Integrating Tagged Neutron Inspection with Explainable AI for Threat Material Identification

Integrating Tagged Neutron Inspection with Explainable AI for Threat Material Identification | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Integrating Tagged Neutron Inspection with Explainable AI for Threat Material Identification Hadi Shahabinejad, Davorin Sudac, Karlo Nad, Isabelle Espagnon, and 12 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4661721/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Here we present an innovative approach for detecting threat materials within a sealed container by integrating tagged fast neutron activation analysis with Explainable Artificial Intelligence (XAI). Two AI models, a Feed-Forward Neural Network (FFNN) and a Convolutional Neural Network (CNN), were developed to analyze the emitted gamma rays to identify materials like explosives and drugs based on depth profiles of carbon, nitrogen, and oxygen concentrations. XAI was applied to make the models' decision-making process transparent. The method is adaptable to various spectrometric analyses. We demonstrate its effectiveness using data obtained by the Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), which is irreplaceable for inspecting sealed cargo containers, despite challenges such as variable material placement, background noise, and shielding effects. Our approach successfully locates and categorizes threat materials, both alone and within surrounding materials, at various locations within sealed cargo containers. Threat material identification Neutron activation Spectrum analysis Deep learning eXplainable Artificial Intelligence (XAI) Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Tagged Neutron Inspection System (TNIS) was initially used to detect explosives hidden in luggages in 2004 [ 1 ] and proposed for cargo container inspection in 2005 [ 2 ]. Using the TNIS, the elemental composition of a suspected item can be determined by coincidence measurements between alpha particles ( α ) and photons ( γ ) resulting from 14 MeV neutron ( n ) interactions in the inspected voxel (or slice) of the container (see Fig. 1 in the next section) [ 3 ]. The target voxel can be identified using neutron time of flight. Alpha particles are usually detected by a position-sensitive Yttrium Aluminum Perovskite (YAP) scintillator in associated alpha particle neutron generators and gamma-rays are detected by commercial scintillation detectors such as NaI(Tl), LaBr 3 (Ce) and BGO. Cargo inspection using fast tagged neutrons was discussed in detail in 2006 and 2007 [4;5] under the EURITRACK (EURopean Illicit TRAfficking Countermeasures Kit) project as a secondary sensor for detailed inspection of suspected container volumes, as determined by an X-ray system. The TNIS configurations for cargo inspection were further elaborated in 2016 [ 6 ], and a final setup named RRTNIS (Rapidly Relocatable Tagged Neutron Inspection System) was proposed for cargo inspection in the same year [ 7 ] under the C-BORD (effective Container inspection at BORDer control points) project [ 8 ]. The details of the RRTNIS detection system in the C-BORD project, including detector types and configurations, data acquisition system, electronics, and its role in anomaly detection, have been presented in [ 9 – 12 ]. The RRTNIS has been further developed by funding received from the European Union’s Horizon 2020 research and innovation program under ENTRANCE [ 13 ] (EfficieNT Risk-bAsed iNspection of freight Crossing bordErs without disrupting business) project. To analyze the gamma-ray spectra obtained from TNIS, the library least square method [ 14 ] was applied to the spectra recorded by a BGO detector in 2016 [ 15 ], aiming to determine the elemental composition of explosives and drugs. Although the library least square method provides good results, the occurrence of negative values for elemental concentrations was reported as its main drawback [ 15 ], as also noted in other studies [16;17]. To address this issue in the full spectrum analysis of gamma-ray spectra, the non-negative least square method [ 18 – 20 ] and artificial intelligence techniques such as hybrid fuzzy-genetic algorithms, particle swarm optimization and artificial neural networks [16; 21–23] have been exploited in similar applications in recent years. Currently, a step-by-step non-negative least square method is being used to analyze gamma-ray spectra obtained from RRTNIS, a topic which is not discussed here. In this work, the RRTNIS and two Artificial Neural Networks (ANNs) supported by eXplainable Artificial Intelligence (XAI)—a Feed Forward Neural Network (FFNN) and a deep Convolutional Neural Network (CNN)—are integrated to identify various threat materials inside a cargo container. The ANNs are designed to facilitate one-dimensional inspection of elements within cargo containers, allowing for the precise identification of threat materials. By leveraging the capabilities of the RRTNIS and the AI, this study aimed to enhance cargo security measures by providing efficient and accurate detection of threat materials, particularly focusing on the identification of carbon, oxygen, and nitrogen, which are the primary components of explosives and drugs. 2. Material and methods In the framework of the project ENTRANCE, the RRTNIS was reassembled at IRESNE Institute of CEA Cadarache, France to improve the experimental setup [ 24 ]. 2.1. Experimental setup The general procedure for the 1D inspection of the container used in this work is presented in Fig. 1 , and the improved version of the RRTNIS under ENTRANCE project is shown in Fig. 2 (a) and (b) [ 24 ]. As seen in Figs. 1 and 2 , the RRTNIS configuration includes an associated alpha particle DT neutron generator with an emission rate of 2×10 7 n/s, encased in a polyethylene shield to ensure radiation protection. The tagged neutron cone, with a total aperture angle of 35°, escapes from the polyethylene shield through a conical aperture (see Fig. 1 ). The distance between polyethylene shield and NaI detectors is 45 cm. Gamma rays resulting from tagged neutron interactions within the inspected container are detected by twenty 5” × 5” × 10” parallelepiped-shaped NaI(Tl) detectors. In addition, the coincidence time between the alpha (position sensitive alpha detector shown in Fig. 1 ) and neutron induced gamma detections allows determining the neutron time of flight (TOF), and subsequently its distance to interaction by assuming that all interactions occur along the beam axis at a constant neutron velocity of 5.13 cm/ns (14-MeV neutrons). As shown in Fig. 1 , the distance utilized in this work was calibrated relative to the wall of the inspected container; specifically, zero distance corresponds to the location of the container wall, indicating zero depth within the container. The 1D inspections of the container with various targets were conducted in slices with a depth resolution of 10 cm and a step size of 5 cm in the horizontal direction (Fig. 1 ). The obtained coincident gamma-ray spectra corresponding to the slices are analyzed using ANN models, resulting in depth profiles for all elements in the database. The XAI is applied to find out which parts of the gamma spectrum contribute to estimating elemental concentrations (C in Fig. 1 ) in output of ANNs for each slice. The suspicious region is determined using elemental depth profiles, and the material type (TNT in Fig. 1 ) is identified through the triangle plot showing the material distributions. The details of all parts will be presented in the following sections. Using the improved version of the RRTNIS setup, an upgraded database of 20 pure element gamma spectra, as well as random and correlated background, was acquired [ 24 ]. Figure 2 (c) and (d) depicts the gamma spectra of C, N, O, Si and Fe as well as the correlated background measured in three different positions normalized to their maximum counts, respectively. To obtain the gamma-spectra of pure elements, the container was removed, and the measured background was subtracted from the obtained spectra. The database of pure elements consists of C, O, N, Si, Al, Ca, Cl, Cr, Cu, F, Fe, K, Mg, Mn, Na, Ni, P, Pb, S, Zn. Graphite was utilized to measure the carbon signature, and water was used for oxygen, as fast neutrons do not produce gamma rays on hydrogen. In instances where a pure element was unavailable, compounds such as Teflon (C 2 F 4 ) were employed for fluorine, after subtracting the carbon signature [ 24 ]. The energy distribution of the random background remains consistent regardless of the position of the slice in the container inspection process. In contrast, the energy distribution of the correlated background depends on the position, specifically the distance from the container wall. The correlated background was measured for different positions by conducting empty measurements, wherein no targets (see Fig. 1 ) were placed in front of the neutron beam. To measure correlated background and different targets, inspection of the container in one dimension was done in slices with a depth of 10 cm and a 5 cm shift. As the correlated background contains its part of random background, only correlated background was used for producing spectra to train the ANNs in this work. These normalized gamma signatures were used to generate data for training ANNs discussed in the following sections. To evaluate the ability of the ANNs in identifying threat materials inside containers using RRTNIS, simulants of dangerous substances with varying C, N, and O weight fractions were created. The components and chemical formulas of the threat materials and their simulants are presented in Table 1 . It can be observed that the chemical formulas of the simulants closely resemble those of the actual dangerous substances. In addition to measuring gamma spectra of the simulants presented in Table 1 and applying ANNs to them, gamma spectra of TNT simulant with varying sizes of iron and wood matrices were also measured, and the ANNs were then evaluated in identifying the threat hidden within these matrices. Table 1 Threat simulants used for the RRTNIS tests in DANAIDES irradiation cell of CEA Cadarache. Threat Description Chemical formula Chemical formula of simulant Simulant mass composition RDX Explosive C 3 H 6 N 6 O 6 Si 3 C 3 H 6 N 6 O 6 Melamine 0.412 Quartz sand 0.588 TNT Explosive C 7 H 5 N 3 O 6 C 7 H 6 N 3 O 6 Graphite 0.158 Oxalic acid dehydrated 0.276 Cyanuric acid 0.566 Cocaine Drug C 17 H 21 NO 4 C 17 H 21 NO 4 Melamine 0.069 – Graphite 0.243 Sucrose 0.41 – Paraffin 0.277 Heroin Drug C 21 H 23 NO 5 C 21 H 23 NO 5 Melamine 0.057 – Graphite 0.294 Sucrose 0.421 – Paraffin 0.227 Yperite Chemical warfare C 4 H 8 Cl 2 S C 4 H 8.2−8.4 Cl 2 SNa 2 Paraffin 0.274 Sodium-chloride 0.57 Sulfur 0.156 Tetranitromethane Liquid explosive C(NO 2 ) 4 Fe 4/9 C 4/3 (NO2) 4 H 32/3 Iron (III) nitrate non-hydrate 0.762 Melamine 0.238 Peroxide methylethylketone Liquid explosive C 8 H 18 O 6 C 8 H 12 O 6 Sucrose 0.914 Graphite 0.086 Nitromethane Liquid explosive CH 3 NO 2 Fe 1/3 CH 5/2 NO 2 Oxalic acid dehydrated 0.398 Melamine 0.266 Iron(III) oxide 0.336 Ethyleneglycol dinitrate Liquid explosive (CH 2 ONO 2 ) 2 FeCH 5/2 NO 3 Oxalic acid dehydrated 0.238 Melamine 0.159 Iron(III) oxide 0.603 2.2. Artificial Neural Network and deep learning Artificial neural networks are computational models inspired by the structure and function of biological neural systems [ 25 ]. These networks consist of interconnected nodes, or artificial neurons, which process and transmit information through weighted connections. ANNs, a subset of which includes deep learning, are designed to mimic the parallel processing and learning capabilities observed in the brain, enabling them to perform tasks such as pattern recognition, classification, and prediction. A crucial aspect of neural networks is that they learn from data by adjusting the weights of connections through algorithms during the training process. Over time, several neural networks architectures have been developed and applied for different applications [ 26 – 33 ], notably including Feed Forward Neural Network (FFNN) and Convolutional Neural Network (CNN), with the term "deep" often attributed to networks with multiple layers. To analyze gamma-ray spectra of RRTNIS using an ANN, after recording gamma-ray spectra from 20 detectors using a data acquisition system and performing energy calibration, the resulting accumulated spectrum is pre-processed by applying energy threshold of 0.6 MeV, compressing the spectrum to have only 246 channels if necessary, and normalizing it to the total gross count of the spectrum. Subsequently, the processed spectrum is fed into the trained ANN model, the outputs of which comprise the count contribution of 20 elements as well as background radiation. To estimate the effect of statistical uncertainties on ANN results, a “small” dataset of size 500 was constructed for each measured spectrum (for instance to build the error bars in Figs. 3 and 5 or cloud around the measured points in Figs. 4 and 6 ). In this process, the counts of each channel of the recorded spectrum were considered to follow a Poisson distribution with the expected counts (λ) equal to the counts in the channel. Subsequently, 500 counts were sampled for each channel using their corresponding Poisson distributions, resulting in 500 gamma spectra for a measured gamma spectrum. The dataset were pre-processed and fed into the network, resulting in 500 sets of outputs. For each output, a Gaussian distribution was fitted to the corresponding 500 data points, with its mean and standard deviation considered as the output of the ANN. 2.2.1. Data preparation for training ANN To train the ANN models in this work, 100,000 gamma spectra were generated by mixing the gamma signatures of 20 pure elements, each with a randomly sampled proportion, and of the correlated background. The generated spectra were divided into 70% training data, 15% testing data, and 15% validation data to train the ANN models described in the following sections. It is worth mentioning that the training data are utilized to optimize the parameters of the ANN models during the training process. Validation data are used to evaluate the generalization ability of the trained models and to prevent overfitting during network training. Additionally, testing data are employed to assess the performance of the trained models on unseen data [ 25 ]. 2.2.2. Feed Forward Neural Network The Feed Forward Neural Network (FFNN), also commonly known as the multilayer perceptron (MLP), represents one of the foundational architectures in the realm of artificial neural networks. In FFNNs, information flows strictly in one direction, from input to output layers, without forming any feedback loops. This unidirectional flow enables FFNNs to effectively process and extract features from complex datasets, making them particularly adept at tasks such as regression, classification, and function approximation [22;23]. Moreover, FFNNs typically comprise multiple layers of interconnected neurons, including input, hidden, and output layers, with each layer playing a distinct role in information processing and abstraction. To determine the optimal FFNN structure in this study, various parameters were systematically explored within predefined ranges. These parameters included the number of hidden layers (1 or 2), the number of neurons in each layer (ranging from 50 to 1500 with an increment of 50), the dropout rate (ranging from 0 to 0.2) and the initial learning rate (ranging from 1e-4 to 1e-3 with an increment of 1e-4). The number of epochs was set 200, and from the 100 randomly tested network structures, the network with the lowest mean square error was identified as the optimal structure. The specifications of this optimal network structure are detailed in Table 2 . As neural networks inherently possess a stochastic nature, different outputs can be obtained for a specific network structure through a “multi-start” process for the initialization of the ANN, including randomized initial weights. To mitigate the effects of randomness, the optimal FFNN architecture was trained independently across 10 runs using identical data. By averaging the outputs obtained from these 10 independent runs, we aimed to obtain a more stable and representative estimation of the network's performance. 2.2.3. Convolutional Neural Network The Convolutional Neural Network (CNN) stands out as a groundbreaking architecture specifically designed to excel in tasks involving visual imagery, such as image classification, object detection, and image segmentation. With their ability to automatically learn hierarchical representations of visual data, CNNs have revolutionized various fields, including computer vision, medical image analysis, and autonomous driving. Recently, CNNs have been applied in analyzing gamma-ray spectra specifically for radioisotope identification [29;30]. Unlike traditional neural networks, CNNs leverage convolutional layers to systematically extract and learn spatial hierarchies of features from input spectra. These convolutional layers employ filters, also referred to as kernels, to perform localized operations across the input spectrum, thereby capturing essential patterns and structures. Additionally, CNNs often incorporate pooling layers to down-sample feature maps, reducing computational complexity and enhancing translation invariance. To determine the optimal CNN structure in this study, various parameters were systematically explored within predefined ranges. These parameters included the number of filters in the Conv1D layers, ranging from 4 to 16, with kernel sizes of 3 to 15. Additionally, the MaxPool layers were tested with stride values of 1 and 2, and kernel sizes of 2, 3 and 5. The number of dense layers was varied between 1 and 2, with the number of neurons in each dense layer ranging from 50 to 1500 with increments of 50. Dropout rates were evaluated from 0 to 0.2. The initial learning rate was adjusted from 1e-4 to 1e-3 with an increment of 1e-4. The number of epochs was set to 300, and from the 200 randomly tested network structures, the network with the lowest mean square error was identified as the optimal structure. The specifications of this optimal network structure are detailed in Table 2 . To mitigate the effects of randomness, the optimal CNN architecture was trained independently across 10 runs using identical data and the outputs were averaged over 10 runs. Table 2 Specifications of the optimal FFNN and CNN models used in this work. ANN type FFNN CNN Network parameters Size (Type) Network parameters Size (Type) Number of inputs 246 Number of inputs 246 Number of hidden layers 2 Conv1D # filters: 8 - kernel: (5×1) - dim: (246,8) Number of neurons in the first hidden layer (dropout rate) 1100 (0.03) Conv1D # filters: 8 - kernel: (3×1) - dim: (244,8) Number of neurons in the second hidden layer (dropout rate) 300 (0) MaxPool1D Strides: 2 - kernel: (2×1) - dim: (122,8) Activation function of hidden layers Relu Conv1D # filters: 4 - kernel: (3×1) - dim: (120,4) Activation function of output layer Softmax MaxPool1D Strides: 2 - kernel: (2×1) - dim: (60,4) Loss function Mean square error Flatten dim:240 Training algorithm Adam Number of Dense layers 2 Initial learning rate 0.0005 Number of neurons in the first dense layer (dropout rate) 900 (0.05) Number of outputs 21 Number of neurons in the second dense layer (dropout rate) 200 (0) Activation function of dense layers Relu Activation function of output layer Softmax Loss function Mean Square Error (MSE) Training algorithm Adam Initial learning rate 0.0005 Number of outputs 21 2.2.4. Explainable artificial intelligence As machine learning models like ANNs continue to advance, they often operate as opaque "black boxes," making it hard to understand how they make decisions. To tackle this, the field of explainable artificial intelligence (XAI) has emerged [34;35]. XAI develops ways to clarify how machine learning models work. Applying XAI to the analysis of the gamma-ray spectrum using full information [30;36] helps to find out which parts of the gamma spectrum contribute to estimating a parameter in output of a machine learning method. This not only makes results more trustworthy but also helps people understand and make decisions faster during real-world tasks. By giving clear explanations, users can better understand and handle alerts, making applications more reliable and effective. The performance of different XAI methods have been examined in the full spectrum analysis of gamma-ray spectra using ANNs in the field of radioisotope identification [36;37], with the best performance obtained from the SHAP (SHapley Additive exPlanations) method in terms of providing more precise explanations based on gamma signatures for correctly identified and misidentified isotopes. SHAP [ 38 ] is based on the concept of Shapley values from cooperative game theory [ 38 ]. In the context of machine learning, SHAP values quantify the impact of each feature on the model's output by considering all possible combinations of features and their contributions. This approach ensures that the importance of each feature is evaluated in relation to its interactions with other features, providing a more comprehensive understanding of their influence on the model's predictions. In this work, the SHAP library was installed using the pip package manager and implemented within the Spyder environment. 3. Results and discussion The correlation coefficients between the targets and predicted outputs of FFNN and CNN models were studied for all elements as well as background radiation related to both train and test data. The study demonstrated the efficient training of both FFNN and CNN models. These trained networks were subsequently utilized to analyze gamma spectra obtained during the inspection of containers with various targets in following section. 3.1. Container inspection using trained ANN models The 1D inspection of the container was conducted in slices with a depth of 10 cm and a 5 cm horizontal shift (see Fig. 1 ), with a measuring time of 10 minutes and with various targets as outlined in Table 1 . The position of the middle of the targets was approximately fixed at 75 cm deep inside the container, and a gamma-ray spectrum was obtained for each slice of the container. The positions of slices were determined using neutron time-of-flight, assuming a constant neutron velocity of 5.13 cm/ns (14-MeV neutrons). The distance of each slice from the iron wall of the container, as illustrated in Fig. 1 , was considered, with the center of each slice used as the reference point. Figure 3 (a) and (b) illustrates the depth profiles of C, N, and O count contributions for a 1 kg of TNT simulant as the target in the container (without cargo matrix around), employing FFNN and CNN models for analyzing the obtained gamma-ray spectra. Similar profiles were obtained for 1 kg of each of the other simulants presented in Table 1 . The error bars of count contributions at each depth were derived from a “small” dataset of size 500 constructed for each measured spectrum, as detailed in section 2.2 . In addition to C, O, and N, other elements were also inspected for material and threat identification in a container. As an example, the depth profiles of Cl and S count contributions corresponding to measuring Yperite simulant (mustard gas, a warfare chemical) as the target (still without cargo matrix around) were obtained, and the results demonstrated precise identification of this simulant. Figure 3 (a) and (b) highlights the importance of further investigating the area around 75 cm (65 to 85 cm) within the container to identify material in this region. Initially, we assessed how well the ANN models can reconstruct the gamma-ray spectrum at a distance of 75 cm for TNT explosive as an example case. The reconstructed spectra, obtained using the mean values from both FFNN and CNN models, are compared to the measured spectrum in Fig. 3 (c). According to Fig. 3 (c), both ANN models perform fairly well in reconstructing the measured spectrum. Furthermore, Fig. 3 (d) to (i) overlays the SHAP values corresponding to the C, O, and N elements on the measured spectrum, to visualize the impact of individual energy bins on the predictions made by the FFNN and CNN models. These SHAP values were obtained from both the FFNN and CNN models. The measured spectrum is represented by a white line on the x-axis depicting energy (MeV) and normalized counts on the y-axis. The color map indicates the magnitude of the SHAP values, offering insights into the contribution of each energy bin to the overall prediction. Both models primarily focus on the major gamma lines of the elements C (4.44 MeV) and O (6.13 MeV) to predict their count contributions in the TNT gamma spectrum. However, the CNN model demonstrates a more pronounced emphasis on these gamma lines compared to the FFNN model. As the 5.11 MeV gamma line of N element is not observed in the TNT spectrum in this measurement, the CNN model distinctly emphasizes the 2.31 MeV gamma line to predict the count contribution of N. Finally, the count contributions should be converted to weight fractions for C, N, and O for threat identification purposes. To achieve this, calibration lines were obtained separately for each element and for both FFNN and CNN models. For each element, the maximum mean values in the corresponding depth profiles of three simulants, in addition to the (0,0) point, were used to map the count contribution to the weight fraction. The simulants RDX, Cocaine and Peroxide methylethylketone shown in Table 1 were used for C element; RDX, Nitromethane, and Tetranitromethane for N element; and RDX, Cocaine and Tetranitromethane for O element. The obtained calibration lines were then applied to convert the count contributions of C, N, and O elements obtained using ANN models to weight fractions for all simulants presented in Table 1 . Figure 3 (a) and (b) shows variations in the count contributions in any selected suspicious area. To mitigate these variations, the mean of the three maximum count contributions in the selected region was considered for each element and then converted to weight fractions using the calibration lines, followed by normalization to ensure that the sum of the weight fractions for C, N, and O equals one. The resulting normalized weight fractions were then considered as a point in the CNO barycentric triangle [ 12 , 39 ]. The CNO barycentric triangle illustrates material compositions based on normalized weight fractions of C, N, and O. Each vertex represents pure C, N, or O, while the interior shows combinations of these elements. Points within the triangle indicate mixtures of C, N, and O, offering a visual representation of material compositions in our plots. Figure 4 depicts the theoretical point represented by a black triangle (▲) in the barycentric triangle obtained from weight fractions for each simulant (true theoretical composition), except for Yperite, compared to the measured points obtained using FFNN and CNN models, respectively. The measured points are represented as black points (●) for FFNN model and star symbols (★) for CNN model for each simulant listed in Table 1 . To illustrate the uncertainty in the location of the measured point in barycentric triangle, the measured points were calculated for “small” dataset of size 500 constructed for each measured spectrum (as detailed in section 2.2 ) using both ANN models. The uncertainties obtained are shown in Fig. 4 , with grey points forming a “cloud” around the measured points. Additionally, blue points represent some of the illicit drugs, red points represent some explosives, and green points represent some benign materials, illustrating the distribution of materials within the CNO barycentric triangle. As evident in Fig. 4 , both ANN models predict the measured points close to the theoretical points, with the FFNN model showing better performance. The better performance of FFNN is further supported by the depth profiles (TNT results shown in Fig. 3 (a) and (b)), as the FFNN exhibits smaller standard deviations for each element. To evaluate the performance of the ANN models in the presence of a matrix, 10 kg of TNT was hidden within the container behind 40 and 80 cm wood bales, with a density of 0.2 g.cm − 3 (mimicking an organic cargo). In another series of experiments, the same TNT target was hidden behind 40 and 105 cm iron boxes filled with bundles of iron wire, with an apparent density of 0.2 g.cm − 3 . Figure 5 displays the depth profiles of C, N and O count contributions obtained using both FFNN and CNN models for 10 kg of TNT hidden in wood and iron matrices. According to the Fig. 5 , the distances of 70–90 cm and 115–135 cm were considered as the suspicious areas for the TNT hidden at depths of 40 cm and 80 cm in the wood matrix, respectively. Furthermore, the distances of 85–105 cm and 160–180 cm were considered as the suspicious areas for the TNT hidden 40 cm and 105 cm deep inside the iron matrix, respectively. The resulting measured points, along with their uncertainty obtained using ANN models in the CNO barycentric triangle for TNT hidden behind wood and iron matrices have been shown in Fig. 6 . The depth profiles of C and O are obscured by the large background profiles of the wood matrix, particularly for TNT hidden in the wood matrix at a depth of 40 cm. This effect is noticeable in Fig. 6 (a) and (b), where the weight fraction of N is underestimated. The SHAP values were calculated for N element for both FFNN and CNN models at distances of 80 cm and 125 cm, which are at the middle of suspicious regions shown in Fig. 5 , for TNT behind 40 cm and 80 cm of wood matrix, respectively. The overlay of SHAP values on the measured spectra of TNT is shown in Fig. 7 . It is obvious in parts (a) and (c) of Fig. 7 that the measured gamma-ray spectrum is mainly influenced by intense gamma lines of C and O for TNT behind 40 cm of wood. The effect is less pronounced for TNT behind 80 cm of wood, resulting in better visibility of the gamma lines of N and consequently facilitating the estimation of N using its major gamma lines (2.31 and 5.11 MeV) by both models. The depth profiles, particularly for N, are significantly affected by the poor counting statistics when TNT is behind 105 cm of iron, as shown in Fig. 5 (g) and (h). This effect is evident in Fig. 6 (g) and (h), showing large uncertainties and fluctuations of the measured points. Due to this critical situation, SHAP values were calculated for the C element for both FFNN and CNN models at distances of 95 cm and 170 cm, which are at the middle of the suspicious regions shown in Fig. 5 , for TNT behind 40 cm and 105 cm of iron matrix, respectively. The overlay of SHAP values on the measured spectra of TNT for the C element is shown in Fig. 8 (a) to (d). Additionally, the overlay of SHAP values for the Fe element is depicted in Fig. 8 (e) and (f) for TNT behind 105 cm of iron matrix for both models. It is evident in parts (a) and (c) of Fig. 8 that both models still mainly utilize the 4.44 MeV gamma line of the measured gamma-ray spectrum to predict C for TNT behind 40 cm of iron, albeit with a lower contrast compared to Fig. 3 (d) and (g). The effect is much more pronounced for TNT behind 105 cm of iron, where no specific region is used for predicting C, but rather the entire spectrum. In fact, there is no specific peak in the measured spectrum corresponding to C, N and O elements, and the measured spectrum is similar to the gamma signature of pure Fe shown in Fig. 2 (c). Since iron is the most abundant material inside the container in this case, both ANN models effectively predict the count contribution of Fe using its major gamma line, as shown in Fig. 8 (e) and (f). Even in this situation where the gamma-rays originating from different elements are heavily obscured by the iron matrix, both ANN models could still fairly determine the position of the threat (Fig. 5 (g) and (h)) and the category of the unknown material as explosive (Fig. 6 (g) and (h)). As can be seen from the depth profiles presented in Figs. 3 and 5 , both the FFNN and CNN models identify the suspicious areas at the same positions (i.e. depths) within the container. This consistency enhances the reliability of container inspection, as it indicates that two different ANN models pinpoint the same locations. Indeed, the consistency observed in the identification of suspicious areas by both FFNN and CNN models was further supported by the results of XAI techniques. Specifically, the SHAP values overlaid on the measured spectrum provide insights into the contribution of individual energy bins to the models' predictions. Figures 3 , 7 , and 8 illustrate how both models primarily focus on major gamma lines of C, O, and N elements to predict their count contributions accurately. Additionally, the CNN model demonstrates a more pronounced emphasis on these gamma lines compared to the FFNN model, highlighting differences in their predictive strategies. This analysis underscores the reliability and interpretability of the models' predictions, enhancing trust in their performance in identifying suspicious areas within cargo containers. Furthermore, the results from Fig. 4 , which depicts the performance of the models in positioning the measured points in the CNO barycentric triangle for simulants without matrices, and Fig. 6 which present the results for TNT simulant hidden within wood and iron matrices, indicate effective performance in identifying material category within the CNO barycentric triangle. Generally, a slightly better performance in terms of proximity to the theoretical points (except for TNT in presence of 105 cm iron matrix) was observed for the FFNN model. However, both models displayed robust predictive capabilities, allowing for the accurate identification of suspicious regions within cargo containers. 4. Conclusion In this study, we conducted a comprehensive evaluation of eXplainable Artificial Intelligence (XAI) to enhance the identification of suspicious areas within cargo containers, integrating tagged fast neutron activation analysis with Feed Forward Neural Network (FFNN) and Convolutional Neural Network (CNN) models. Employing innovative approach outlined in our investigation, our aim was to enhance container screening effectiveness for global trade security. Through the analysis of depth profiles depicting the count contributions of elements, we observed consistent identification of suspicious regions within the container by both FFNN and CNN models, affirming the reliability of their predictions. Furthermore, utilization of XAI provided valuable insight into the predictive strategies and decision-making processes of the ANN models, enhancing their interpretability and overall reliability. We further extended our investigation by visualizing material compositions using the CNO barycentric triangle. This provided valuable insights into the identification of material categories for analyzed targets, both with and without matrices present in the container. The results demonstrated effective performance in identifying threat materials within the respective categories using both ANN models. However, it is noteworthy that the CNN model exhibits a more pronounced emphasis on the major gamma lines of elements compared to the FFNN model, ensuring highly reliable results and potentially serving as a complementary support for FFNN results. For future work, a promising direction is advancing deep learning methods for spectrum analysis in general, particularly considering the insights gained from XAI results. Improving the performance of deep learning models through exploration of different network architectures could be a fruitful area for research. Additionally, studying the likelihood probability of the measured points relative to the closest reference materials (such as explosives, illicit drugs, and benign substances) in the triangle plots could provide valuable insights. Investigating the step-by-step non-negative least square method currently utilized for gamma-ray spectra analysis in RRTNIS presents another avenue for exploration. Declarations Acknowledgments This research was funded by the Horizon Europe Framework Programme under the ENTRANCE (Grant Agreement No. 883424) and UnderSec (Grant Agreement No. 101121288) projects. The authors gratefully acknowledge the active involvement and support of the Croatian Ministry of Interior, the Croatian Ministry of Finance – Customs Administration, and the Port of Rijeka d.d. Author contributions Hadi Shahabinejad analyzed data, developed and wrote the analysis codes and drafted the manuscript; Davorin Sudac, Karlo Nad, Jasmina Obhođaš, Isabelle Espagnon, Clotilde de Sainte Foy, Bertrand Perot, Cedric Carasco, Alix Sardet, Jessica Delgado, Felix Pino, Sandra Moretto, contributed to hardware and software optimization and acquisition of experimental data; Edwin Friedmann, Jean Philippe Poli, Isabelle Espagnon, Clotilde de Sainte Foy, Bertrand Perot, Cedric Carasco, Alix Sardet, Jessica Delgado, Felix Pino, Sandra Moretto analyzed data; Christine Mer and Guillaume Sannie secured funds, planned experiments, and contributed to the acquisition of the experimental data; Hadi Shahabinejad and Jasmina Obhodas conceived the idea of developing AI for experimental data; All authors discussed the results and commented on the manuscript. Competing interests The authors declare no competing interests. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. References Pesente, S. et al. Detection of hidden explosives by using tagged neutron beams with sub-nanosecond time resolution, Nuclear Instruments and Methods in Physics Research A. 531, 657–667 (2004). Viesti, G. et al. Detection of hidden explosives by using tagged neutron beams: Status and perspectives, Nuclear Instruments and Methods in Physics Research B. 241, 748–752 (2005). Sardet A. et al. Performances of C-BORD’s Tagged Neutron Inspection System for Explosives and Illicit Drugs Detection in Cargo Containers, 68, 346 – 353 (2021). Perret, G., Perot, B., Artaud, J-L., and Mariani, A. EURITRACK tagged neutron inspection system design, Journal of Physics: Conference Series 41, 375–383 (2006). Perot, B. et al. Development of the EURITRACK tagged neutron inspection system, Nuclear Instruments and Methods in Physics Research B. 261, 295–298 (2007). Valkovic, V., Sudac, D., Nad, K., and Obhodas J. Container Inspection Utilizing 14 MeV Neutrons, IEEE Transactions on Nuclear Science, 63 (3), 1536-1543 (2016). Sardet, A. et al. Design of the rapidly relocatable tagged neutron inspection system of the C-BORD project, in Proc. IEEE Nucl. Sci. Symp., Med. Imag. Conf. Room-Temp. Semiconductor Detect. Workshop (NSS/MIC/RTSD), Strasbourg, France, Oct., 1–5 (2016). C-BORD project, Effective Container Inspection at BORDer Control Points, https://www.cbord-h2020.eu. Fontana, C.L. et al. Detection System of the First Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), Developed in the Framework of the European H2020 C-BORD Project, Physics Procedia, 90, 279-284 (2017). Pino, F. et al. Advances on the development of the detection system of C-BORD’s rapidly relocatable tagged neutron inspection, International Journal of Modern Physics: Conference Series, 48, 1860125 (2018). Pin, F. et al. Detection module of the C-BORD Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), Nuclear Inst. and Methods in Physics Research, A. 986, 164743 (2021). Pino, F. et al. Non-intrusive inspection of cargo containers using the C-BORD Rapidly Relocatable Tagged Neutron Inspection System, JINST. 17, T12005 (2022). ENTRANCE project, "Efficient risk-based inspection of freight crossing borders without disrupting business", https://www.entrance-h2020.eu. Gardner, R.P., Sood, A., Wang, Y.Y., Liu, L., Guo, P., Gehrke, R.J. Single peak versus library least-squares analysis methods for the PGNAA analysis of vitrified waste, Applied Radiation and Isotope, 48, 1331-1335 (1997). Obhodas, J., Sudac, D., and Valkovic, V. Development of the Quality Assurance/Quality Control Procedures for a Neutron Interrogation System, IEEE Transactions on Nuclear Science. 63 (3), 1544 – 1550 (2016). Shahabinejad, H., and Vosoughi, N. Analysis of complex gamma-ray spectra using particle swarm optimization, Nucl. Instrum. Methods Phys. Res. A. 911, 123-130 (2018). Meric, I., Johansen, G.A., Holstad, M.B., Wang, J., Gardner, R.P. Produced water characterization by prompt gamma-ray neutron activation analysis, Meas. Sci. Technol., 22, 125701 (2011). Caciolli, A., Baldoncini, M., Bezzon, G.P., Broggini, C., Buso, G.P., Callegari, I., Colonna, T., Fiorentini, G., Guastaldi, E., Mantovani, F., Massa, G., Menegazzo, R., Mou, L., Rossi Alvarez, C., Shyti, M., Zanon, A., Xhixha, G. A new FSA approach for in situ γ ray spectroscopy. Sci. Total Environ. 414, 639-645 (2012). Baldoncini, M., Albéri, M., Bottardi, C., Chiarelli, E., Raptis, K.G.C., Strati, V., Mantovani, F. Investigating the potentialities of Monte Carlo simulation for assessing soil water content via proximal gamma-ray spectroscopy, J. Environ. Radioact. 192, 105-116 (2018). Shahabinejad H., and Vosoughi N. SGSD: A novel Sequential Gamma-ray Spectrum Deconvolution algorithm. Annals of Nuclear Energy. 132, 369-380 (2019). Alamaniotis M., and Jevremovic T. Hybrid fuzzy-genetic approach integrating peak identification and spectrum fitting for complex gamma-ray spectra analysis. IEEE Transactions on Nuclear Science. 62, 1262-1277 (2015). Kamuda M., Stinnett J., Sullivan C. Automated isotope identification algorithm using artificial neural networks. IEEE Transactions on Nuclear Science. 64, 1858-1864 (2017). Shahabinejad H., Vosoughi N., Saheli F., “Processing scintillation gamma-ray spectra by artificial neural network”, Journal of Radioanalytical and Nuclear Chemistry, 325, 471-483 (2020). Sainte Foy C.D. et al. Optimization of the ENTRANCE Rapidly Relocatable Tagged Neutron Inspection System, IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), Italy, 1-7 (2022). Fausett, L.V. Fundamentals of neural networks: architectures, algorithms, and applications, vol 3. prentice-Hall Englewood Cliffs (1994). Haykin, S. Neural networks: a comprehensive foundation. Prentice Hall PTR (1994). Ryu, J., Park, C., Park, J., Cho, N., Park, J., and Cho, G. Development of Neural Network Model With Explainable AI for Measuring Uranium Enrichment, IEEE TRANSACTIONS ON NUCLEAR SCIENCE, 68, 2670-2681 (2021). Hossny, K., Hossny, A.H., Magdi, S. Detecting shielded explosives by coupling prompt gamma neutron activation analysis and deep neural networks. Sci Rep. 10, 13467 (2020). Kamuda, M., Zhao, J., Huff, K. A comparison of machine learning methods for automated gamma-ray spectroscopy. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 954, 161385 (2020). Gomez-Fernandez, M., Wong, W.K., Tokuhiro, A., Welter, K., Alhawsawi, A.M., Yang, H., et al. Isotope identification using deep learning: An explanation. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 988, 164925 (2021). Carasco, C., Pérot, B., Sardet, A. Measuring hydrogen with fast neutrons: Application to organic materials identification in cargo containers, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 951, 163030 (2020). Gong, K., He, D.-D., Chen, S.-Y., Jing, S.-W., Zheng, Y.-L. BP neural network analysis for identification of explosive in package by tagged neutron method, Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 101721, 165810 (2021). Gong, K., Xiao, S.-J., Jing, S.-W., Zheng, Y.-L. Back propagation neural network analysis for the detection of explosives based on tagged neutron, Journal of Radioanalytical and Nuclear Chemistry. 326, 329-336 (2020). Adadi, A., Berrada, M. Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access. 6, 52138–52160 (2018). Barredo Arrieta, A., Dıaz-Rodrıguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado A. Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI. Information Fusion. 58, 82–115 (2020). Bandstra, M.S., Curtis J.C., Ghawaly, J.M. Jr, Jones, A.C., Joshi, T.H.Y. Explaining machine-learning models for gamma-ray detection and identification. PLoS ONE 18(6), e0286829 (2023). Jeon, B., Kim, J., Moon, M. Explanation of Deep Learning–Based Radioisotope Identifier for Plastic Scintillation Detector, Nuclear Technology, 209, 1, 1-14, (2023). LUNDBERG S.M. and LEE S.-I. A Unified Approach to Interpreting Model Predictions, Proc. 31st Int. Conf. Advances in Neural Information Processing Systems, Long Beach, California, December 4–9, 4768 (2017). Carasco, C. et al. Data Acquisition and Analysis of the UNCOSS Underwater Explosive Neutron Sensor, 2nd International Conference on Advancements in Nuclear Instrumentation, Measurement Methods and their Applications, IEEE, (2011). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4661721","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":329688081,"identity":"9a4d7b13-21d9-4a3a-84e6-5deda679bfe7","order_by":0,"name":"Hadi Shahabinejad","email":"data:image/png;base64,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","orcid":"","institution":"Institute Ruder Boskovic","correspondingAuthor":true,"prefix":"","firstName":"Hadi","middleName":"","lastName":"Shahabinejad","suffix":""},{"id":329688082,"identity":"e2e35bda-c79f-4ff3-acb9-81affd2ae98e","order_by":1,"name":"Davorin Sudac","email":"","orcid":"","institution":"Institute Ruder Boskovic","correspondingAuthor":false,"prefix":"","firstName":"Davorin","middleName":"","lastName":"Sudac","suffix":""},{"id":329688084,"identity":"1458f6aa-349c-4992-9d86-bdb051361f14","order_by":2,"name":"Karlo Nad","email":"","orcid":"","institution":"Institute Ruder Boskovic","correspondingAuthor":false,"prefix":"","firstName":"Karlo","middleName":"","lastName":"Nad","suffix":""},{"id":329688086,"identity":"508235f1-9a39-43fa-b3e5-192c8ad5bb74","order_by":3,"name":"Isabelle Espagnon","email":"","orcid":"","institution":"CEA LIST","correspondingAuthor":false,"prefix":"","firstName":"Isabelle","middleName":"","lastName":"Espagnon","suffix":""},{"id":329688089,"identity":"8e500e08-bae4-4486-b1b5-2c821eab5cb1","order_by":4,"name":"Clotilde de Sainte Foy","email":"","orcid":"","institution":"CEA Cadarache","correspondingAuthor":false,"prefix":"","firstName":"Clotilde","middleName":"de Sainte","lastName":"Foy","suffix":""},{"id":329688090,"identity":"29094fc5-c471-4e80-8c3d-479b88c46533","order_by":5,"name":"Bertrand Perot","email":"","orcid":"","institution":"CEA Cadarache","correspondingAuthor":false,"prefix":"","firstName":"Bertrand","middleName":"","lastName":"Perot","suffix":""},{"id":329688092,"identity":"4e2b454b-12ab-4bff-9946-a75922e4aa3b","order_by":6,"name":"Cedric Carasco","email":"","orcid":"","institution":"CEA Cadarache","correspondingAuthor":false,"prefix":"","firstName":"Cedric","middleName":"","lastName":"Carasco","suffix":""},{"id":329688093,"identity":"5cfe0755-0bf8-4cad-aed2-401893ec7780","order_by":7,"name":"Alix Sardet","email":"","orcid":"","institution":"CEA Cadarache","correspondingAuthor":false,"prefix":"","firstName":"Alix","middleName":"","lastName":"Sardet","suffix":""},{"id":329688094,"identity":"f93fb225-653d-4b4e-963c-28d2621f61a3","order_by":8,"name":"Edwin Friedmann","email":"","orcid":"","institution":"CEA LIST","correspondingAuthor":false,"prefix":"","firstName":"Edwin","middleName":"","lastName":"Friedmann","suffix":""},{"id":329688095,"identity":"df112a59-870a-4601-8563-1d3ef7785b3c","order_by":9,"name":"Jean Philippe Poli","email":"","orcid":"","institution":"CEA LIST","correspondingAuthor":false,"prefix":"","firstName":"Jean","middleName":"Philippe","lastName":"Poli","suffix":""},{"id":329688096,"identity":"9d8157a3-0831-48a0-bba9-20f361b803f3","order_by":10,"name":"Jessica Delgado","email":"","orcid":"","institution":"Università degli Studi di Padova","correspondingAuthor":false,"prefix":"","firstName":"Jessica","middleName":"","lastName":"Delgado","suffix":""},{"id":329688098,"identity":"7d5d6c4d-cb13-4512-8fea-c39fea7d2247","order_by":11,"name":"Felix Pino","email":"","orcid":"","institution":"Università degli Studi di Padova","correspondingAuthor":false,"prefix":"","firstName":"Felix","middleName":"","lastName":"Pino","suffix":""},{"id":329688099,"identity":"6181f562-8c5f-43f0-8586-9d11045308fa","order_by":12,"name":"Sandra Moretto","email":"","orcid":"","institution":"Università degli Studi di Padova","correspondingAuthor":false,"prefix":"","firstName":"Sandra","middleName":"","lastName":"Moretto","suffix":""},{"id":329688100,"identity":"f38576fe-c447-40fe-bb82-7d908bf12ac7","order_by":13,"name":"Christine Mer","email":"","orcid":"","institution":"CEA LIST","correspondingAuthor":false,"prefix":"","firstName":"Christine","middleName":"","lastName":"Mer","suffix":""},{"id":329688101,"identity":"8f75dda3-bc54-43e2-9bb2-07b3aa944266","order_by":14,"name":"Guillaume Sannie","email":"","orcid":"","institution":"CEA LIST","correspondingAuthor":false,"prefix":"","firstName":"Guillaume","middleName":"","lastName":"Sannie","suffix":""},{"id":329688102,"identity":"f2b3d94b-a793-4e6e-994c-a128e65d3626","order_by":15,"name":"Jasmina Obhodas","email":"","orcid":"","institution":"Institute Ruder Boskovic","correspondingAuthor":false,"prefix":"","firstName":"Jasmina","middleName":"","lastName":"Obhodas","suffix":""}],"badges":[],"createdAt":"2024-06-30 07:53:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4661721/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4661721/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":60988194,"identity":"db3a680d-29b3-463c-991a-0e1e7422378d","added_by":"auto","created_at":"2024-07-24 10:36:12","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":358050,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eA brief description of the 1D cargo inspection using ENTRANCE RRTNIS setup and AI.\u003c/strong\u003e The distance and range of slices shown in the container are determined using neutron time-of-flight. The maximum number of slices (m) in this work is 59. The obtained coincident gamma-ray spectra corresponding to the slices are analyzed using ANN models, resulting in depth profiles for all elements in the database. The XAI is applied to find out which parts of the gamma spectrum contribute to estimating elemental concentrations (C in this Figure) in output of the ANNs in each slice. The suspicious region is determined using elemental depth profiles, and the material type (TNT in this Figure) is identified through the triangle plot showing the material distributions.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/4d74a9c68469f52fbe1f44bf.png"},{"id":60988193,"identity":"3e4aa806-3538-4ccd-ad3f-6e26359b933c","added_by":"auto","created_at":"2024-07-24 10:36:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":377940,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eENTRANCE RRTNIS design and gamma signatures.\u003c/strong\u003e (a) Schematic view of the ENTRANCE RRTNIS design, and (b) Experimental setup at CEA Cadarache, France. Gamma-ray energy spectra obtained using RRTNIS for (c) C, O, N, Si and Fe elements, and (d) correlated background in three different distances from container wall shown in Figure 1.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/11c05ee347c8a4915c49b137.png"},{"id":60988195,"identity":"90de0da6-0292-4087-bac1-e6c200ada99a","added_by":"auto","created_at":"2024-07-24 10:36:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":271434,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDepth profiles, reconstructed spectra and XAI results for a 1 kg TNT simulant.\u003c/strong\u003e Depth profiles obtained using (a) FFNN and (b) CNN models. (c) Comparison of reconstructed spectra, generated using mean values from both FFNN and CNN models, with the measured spectrum at a distance of 75 cm. Overlay of SHAP values on the measured spectrum of TNT at a distance of 75 cm in the container for the C, O, and N elements, obtained from FFNN model ((d), (e), and (f)), and CNN model ((g), (h), and (i)). The ANN input spectrum (white line) is plotted against energy (MeV) on the x-axis and normalized counts on the y-axis. SHAP values (color map) represent the impact of each energy bin on the prediction. The color bar indicates the magnitude of the SHAP values.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/4b7a93283dc1d35a18d28322.png"},{"id":60988201,"identity":"aaac8788-12f5-45b2-924e-393828fd5561","added_by":"auto","created_at":"2024-07-24 10:36:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":565306,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTarget points vs measured points for different threats.\u003c/strong\u003e Theoretical point (black triangle) in the barycentric triangle compared to measured points obtained using FFNN (black point) or CNN (black star) models for various simulants measured without matrix. Clouds around measured points depict the uncertainty of the measured point. Blue, red, and green points represent illicit drugs, explosives, and benign materials, respectively. In the plots, abbreviated names such as \"Tetranit\" and \"Peroxide\" are used to represent \"Tetranitromethane\" and \"Peroxide methylethylketone\", respectively, for brevity and clarity.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/382c271e60aa40926e5146ab.png"},{"id":60988198,"identity":"3f57da8b-33d3-4678-bf87-b0a6b4c7f7e4","added_by":"auto","created_at":"2024-07-24 10:36:13","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":241213,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDepth profiles for 10 kg of TNT hidden in wood and Iron matrices.\u003c/strong\u003e Depth profiles of C, O, and N count contributions obtained using FFNN ((a) and (c)) and CNN ((b) and (d)) models for 10 kg of TNT hidden in the wood matrix at a depth of 40 and 80 cm. Depth profiles of C, O, and N count contributions obtained using FFNN ((e) and (g)) and CNN ((f) and (h)) models for 10 kg of TNT hidden in the iron matrix at a depth of 40 and 105 cm.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/bd7b80d6807b10271cb40c79.png"},{"id":60988200,"identity":"da09e545-4514-4b70-a2c2-4c94660b81ca","added_by":"auto","created_at":"2024-07-24 10:36:13","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":258945,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTarget point vs. measured points for 10 kg of TNT hidden in wood and iron matrices. \u003c/strong\u003eMeasured points obtained using ANN models in the barycentric triangle for TNT hidden in the wood matrix at a depth of 40 cm ((a - FFNN) and (b - CNN)), and 80 cm ((c - FFNN) and (d - CNN)). Measured points were also obtained using ANN models in the barycentric triangle for TNT hidden in the iron matrix at a depth of 40 cm ((e - FFNN) and (f - CNN)), and 105 cm ((g - FFNN) and (h - CNN)). The black triangle represents the theoretical point for TNT simulant. The cloud around measured point depict the uncertainty of measured point.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/7f75a02062840de90d37a09c.png"},{"id":60988197,"identity":"d7266249-049a-4e51-b2e2-1c28d1887568","added_by":"auto","created_at":"2024-07-24 10:36:13","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":416192,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eXAI results for TNT hidden within the wood matrix.\u003c/strong\u003e Overlay of SHAP values on the measured spectrum of TNT at a distance of 80 cm within the container for the element N, when the TNT is hidden within the wood matrix at depth of 40 cm, obtained from a FFNN model ((a)), and a CNN model ((c)). Results obtained at a distance of 125 cm, with TNT hidden within a wood matrix at depth of 80 cm, are shown for the FFNN model in (b), and the CNN model in (d). The measured spectrum (white line) is plotted against energy (MeV) on the x-axis and normalized counts on the y-axis. SHAP values (color map) represent the impact of each energy bin on the prediction. The color bar indicates the magnitude of the SHAP values.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/0342c62af8dd4a3db4232ff0.png"},{"id":60988199,"identity":"9b8750cf-03ca-45af-abc1-1f378c4ad02e","added_by":"auto","created_at":"2024-07-24 10:36:13","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":335324,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eXAI results for TNT hidden within the iron matrix.\u003c/strong\u003e Overlay of SHAP values on the measured spectrum of TNT at a distance of 95 cm within the container for the N element, when the TNT is hidden in iron matrix at depth of 40 cm, obtained from a FFNN model ((a)), and a CNN model ((c)). Results obtained at a distance of 170 cm, with TNT hidden within iron matrix at depth of 105 cm, are shown for the FFNN model in (b), and the CNN model in (d). The Overlay of SHAP values for element Fe are also shown for the FFNN model in (e), and the CNN model in (f), with TNT hidden within iron matrix at depths of 105 cm. The measured spectrum (white line) is plotted against energy (MeV) on the x-axis and normalized counts on the y-axis. SHAP values (color map) represent the impact of each energy bin on the prediction. The color bar indicates the magnitude of the SHAP values.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/14eae934b0250d5d469c11b3.png"},{"id":80776896,"identity":"6cf6cde5-123e-406c-a833-35ad809aee40","added_by":"auto","created_at":"2025-04-17 03:31:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3867294,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4661721/v1/82ae66e3-6103-4208-8b53-e113814be57b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating Tagged Neutron Inspection with Explainable AI for Threat Material Identification","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eTagged Neutron Inspection System (TNIS) was initially used to detect explosives hidden in luggages in 2004 [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] and proposed for cargo container inspection in 2005 [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Using the TNIS, the elemental composition of a suspected item can be determined by coincidence measurements between alpha particles (\u003cem\u003eα\u003c/em\u003e) and photons (\u003cem\u003eγ\u003c/em\u003e) resulting from 14 MeV neutron (\u003cem\u003en\u003c/em\u003e) interactions in the inspected voxel (or slice) of the container (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e in the next section) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The target voxel can be identified using neutron time of flight. Alpha particles are usually detected by a position-sensitive Yttrium Aluminum Perovskite (YAP) scintillator in associated alpha particle neutron generators and gamma-rays are detected by commercial scintillation detectors such as NaI(Tl), LaBr\u003csub\u003e3\u003c/sub\u003e(Ce) and BGO.\u003c/p\u003e \u003cp\u003eCargo inspection using fast tagged neutrons was discussed in detail in 2006 and 2007 [4;5] under the EURITRACK (EURopean Illicit TRAfficking Countermeasures Kit) project as a secondary sensor for detailed inspection of suspected container volumes, as determined by an X-ray system. The TNIS configurations for cargo inspection were further elaborated in 2016 [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], and a final setup named RRTNIS (Rapidly Relocatable Tagged Neutron Inspection System) was proposed for cargo inspection in the same year [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] under the C-BORD (effective Container inspection at BORDer control points) project [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. The details of the RRTNIS detection system in the C-BORD project, including detector types and configurations, data acquisition system, electronics, and its role in anomaly detection, have been presented in [\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. The RRTNIS has been further developed by funding received from the European Union\u0026rsquo;s Horizon 2020 research and innovation program under ENTRANCE [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] (EfficieNT Risk-bAsed iNspection of freight Crossing bordErs without disrupting business) project.\u003c/p\u003e \u003cp\u003eTo analyze the gamma-ray spectra obtained from TNIS, the library least square method [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] was applied to the spectra recorded by a BGO detector in 2016 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], aiming to determine the elemental composition of explosives and drugs. Although the library least square method provides good results, the occurrence of negative values for elemental concentrations was reported as its main drawback [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], as also noted in other studies [16;17]. To address this issue in the full spectrum analysis of gamma-ray spectra, the non-negative least square method [\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and artificial intelligence techniques such as hybrid fuzzy-genetic algorithms, particle swarm optimization and artificial neural networks [16; 21\u0026ndash;23] have been exploited in similar applications in recent years. Currently, a step-by-step non-negative least square method is being used to analyze gamma-ray spectra obtained from RRTNIS, a topic which is not discussed here.\u003c/p\u003e \u003cp\u003eIn this work, the RRTNIS and two Artificial Neural Networks (ANNs) supported by eXplainable Artificial Intelligence (XAI)\u0026mdash;a Feed Forward Neural Network (FFNN) and a deep Convolutional Neural Network (CNN)\u0026mdash;are integrated to identify various threat materials inside a cargo container. The ANNs are designed to facilitate one-dimensional inspection of elements within cargo containers, allowing for the precise identification of threat materials. By leveraging the capabilities of the RRTNIS and the AI, this study aimed to enhance cargo security measures by providing efficient and accurate detection of threat materials, particularly focusing on the identification of carbon, oxygen, and nitrogen, which are the primary components of explosives and drugs.\u003c/p\u003e"},{"header":"2. Material and methods","content":"\u003cp\u003eIn the framework of the project ENTRANCE, the RRTNIS was reassembled at IRESNE Institute of CEA Cadarache, France to improve the experimental setup [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Experimental setup\u003c/h2\u003e \u003cp\u003eThe general procedure for the 1D inspection of the container used in this work is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and the improved version of the RRTNIS under ENTRANCE project is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e(a) and (b) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. As seen in Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the RRTNIS configuration includes an associated alpha particle DT neutron generator with an emission rate of 2\u0026times;10\u003csup\u003e7\u003c/sup\u003e n/s, encased in a polyethylene shield to ensure radiation protection. The tagged neutron cone, with a total aperture angle of 35\u0026deg;, escapes from the polyethylene shield through a conical aperture (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe distance between polyethylene shield and NaI detectors is 45 cm. Gamma rays resulting from tagged neutron interactions within the inspected container are detected by twenty 5\u0026rdquo; \u0026times; 5\u0026rdquo; \u0026times; 10\u0026rdquo; parallelepiped-shaped NaI(Tl) detectors. In addition, the coincidence time between the alpha (position sensitive alpha detector shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) and neutron induced gamma detections allows determining the neutron time of flight (TOF), and subsequently its distance to interaction by assuming that all interactions occur along the beam axis at a constant neutron velocity of 5.13 cm/ns (14-MeV neutrons). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the distance utilized in this work was calibrated relative to the wall of the inspected container; specifically, zero distance corresponds to the location of the container wall, indicating zero depth within the container. The 1D inspections of the container with various targets were conducted in slices with a depth resolution of 10 cm and a step size of 5 cm in the horizontal direction (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The obtained coincident gamma-ray spectra corresponding to the slices are analyzed using ANN models, resulting in depth profiles for all elements in the database. The XAI is applied to find out which parts of the gamma spectrum contribute to estimating elemental concentrations (C in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) in output of ANNs for each slice. The suspicious region is determined using elemental depth profiles, and the material type (TNT in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) is identified through the triangle plot showing the material distributions. The details of all parts will be presented in the following sections.\u003c/p\u003e \u003cp\u003eUsing the improved version of the RRTNIS setup, an upgraded database of 20 pure element gamma spectra, as well as random and correlated background, was acquired [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e(c) and (d) depicts the gamma spectra of C, N, O, Si and Fe as well as the correlated background measured in three different positions normalized to their maximum counts, respectively. To obtain the gamma-spectra of pure elements, the container was removed, and the measured background was subtracted from the obtained spectra. The database of pure elements consists of C, O, N, Si, Al, Ca, Cl, Cr, Cu, F, Fe, K, Mg, Mn, Na, Ni, P, Pb, S, Zn. Graphite was utilized to measure the carbon signature, and water was used for oxygen, as fast neutrons do not produce gamma rays on hydrogen. In instances where a pure element was unavailable, compounds such as Teflon (C\u003csub\u003e2\u003c/sub\u003eF\u003csub\u003e4\u003c/sub\u003e) were employed for fluorine, after subtracting the carbon signature [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The energy distribution of the random background remains consistent regardless of the position of the slice in the container inspection process. In contrast, the energy distribution of the correlated background depends on the position, specifically the distance from the container wall. The correlated background was measured for different positions by conducting empty measurements, wherein no targets (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) were placed in front of the neutron beam. To measure correlated background and different targets, inspection of the container in one dimension was done in slices with a depth of 10 cm and a 5 cm shift. As the correlated background contains its part of random background, only correlated background was used for producing spectra to train the ANNs in this work. These normalized gamma signatures were used to generate data for training ANNs discussed in the following sections.\u003c/p\u003e \u003cp\u003eTo evaluate the ability of the ANNs in identifying threat materials inside containers using RRTNIS, simulants of dangerous substances with varying C, N, and O weight fractions were created. The components and chemical formulas of the threat materials and their simulants are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. It can be observed that the chemical formulas of the simulants closely resemble those of the actual dangerous substances. In addition to measuring gamma spectra of the simulants presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and applying ANNs to them, gamma spectra of TNT simulant with varying sizes of iron and wood matrices were also measured, and the ANNs were then evaluated in identifying the threat hidden within these matrices.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThreat simulants used for the RRTNIS tests in DANAIDES irradiation cell of CEA Cadarache.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThreat\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eChemical formula\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eChemical formula\u003c/p\u003e \u003cp\u003eof simulant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSimulant mass\u003c/p\u003e \u003cp\u003ecomposition\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRDX\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExplosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e3\u003c/sub\u003eH\u003csub\u003e6\u003c/sub\u003eN\u003csub\u003e6\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSi\u003csub\u003e3\u003c/sub\u003eC\u003csub\u003e3\u003c/sub\u003eH\u003csub\u003e6\u003c/sub\u003eN\u003csub\u003e6\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMelamine 0.412\u003c/p\u003e \u003cp\u003eQuartz sand 0.588\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTNT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExplosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e7\u003c/sub\u003eH\u003csub\u003e5\u003c/sub\u003eN\u003csub\u003e3\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u003csub\u003e7\u003c/sub\u003eH\u003csub\u003e6\u003c/sub\u003eN\u003csub\u003e3\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGraphite 0.158\u003c/p\u003e \u003cp\u003eOxalic acid dehydrated 0.276\u003c/p\u003e \u003cp\u003eCyanuric acid 0.566\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCocaine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDrug\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e17\u003c/sub\u003eH\u003csub\u003e21\u003c/sub\u003eNO\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u003csub\u003e17\u003c/sub\u003eH\u003csub\u003e21\u003c/sub\u003eNO\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMelamine 0.069 \u0026ndash; Graphite\u003c/p\u003e \u003cp\u003e0.243 Sucrose 0.41 \u0026ndash; Paraffin\u003c/p\u003e \u003cp\u003e0.277\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeroin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDrug\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e21\u003c/sub\u003eH\u003csub\u003e23\u003c/sub\u003eNO\u003csub\u003e5\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u003csub\u003e21\u003c/sub\u003eH\u003csub\u003e23\u003c/sub\u003eNO\u003csub\u003e5\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMelamine 0.057 \u0026ndash; Graphite\u003c/p\u003e \u003cp\u003e0.294 Sucrose 0.421 \u0026ndash; Paraffin 0.227\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYperite\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChemical warfare\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e4\u003c/sub\u003eH\u003csub\u003e8\u003c/sub\u003eCl\u003csub\u003e2\u003c/sub\u003eS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u003csub\u003e4\u003c/sub\u003eH\u003csub\u003e8.2\u0026minus;8.4\u003c/sub\u003eCl\u003csub\u003e2\u003c/sub\u003eSNa\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eParaffin 0.274\u003c/p\u003e \u003cp\u003eSodium-chloride 0.57\u003c/p\u003e \u003cp\u003eSulfur 0.156\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTetranitromethane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiquid explosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC(NO\u003csub\u003e2\u003c/sub\u003e)\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFe\u003csub\u003e4/9\u003c/sub\u003eC\u003csub\u003e4/3\u003c/sub\u003e(NO2)\u003csub\u003e4\u003c/sub\u003eH\u003csub\u003e32/3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eIron (III) nitrate non-hydrate 0.762\u003c/p\u003e \u003cp\u003eMelamine 0.238\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeroxide methylethylketone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiquid explosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC\u003csub\u003e8\u003c/sub\u003eH\u003csub\u003e18\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u003csub\u003e8\u003c/sub\u003eH\u003csub\u003e12\u003c/sub\u003eO\u003csub\u003e6\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSucrose 0.914\u003c/p\u003e \u003cp\u003eGraphite 0.086\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNitromethane\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiquid explosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCH\u003csub\u003e3\u003c/sub\u003eNO\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFe\u003csub\u003e1/3\u003c/sub\u003eCH\u003csub\u003e5/2\u003c/sub\u003eNO\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOxalic acid dehydrated 0.398 Melamine 0.266\u003c/p\u003e \u003cp\u003eIron(III) oxide 0.336\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEthyleneglycol dinitrate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiquid explosive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(CH\u003csub\u003e2\u003c/sub\u003eONO\u003csub\u003e2\u003c/sub\u003e)\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFeCH\u003csub\u003e5/2\u003c/sub\u003eNO\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOxalic acid dehydrated 0.238 Melamine 0.159\u003c/p\u003e \u003cp\u003eIron(III) oxide 0.603\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Artificial Neural Network and deep learning\u003c/h2\u003e \u003cp\u003eArtificial neural networks are computational models inspired by the structure and function of biological neural systems [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. These networks consist of interconnected nodes, or artificial neurons, which process and transmit information through weighted connections. ANNs, a subset of which includes deep learning, are designed to mimic the parallel processing and learning capabilities observed in the brain, enabling them to perform tasks such as pattern recognition, classification, and prediction. A crucial aspect of neural networks is that they learn from data by adjusting the weights of connections through algorithms during the training process. Over time, several neural networks architectures have been developed and applied for different applications [\u003cspan additionalcitationids=\"CR27 CR28 CR29 CR30 CR31 CR32\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], notably including Feed Forward Neural Network (FFNN) and Convolutional Neural Network (CNN), with the term \"deep\" often attributed to networks with multiple layers.\u003c/p\u003e \u003cp\u003eTo analyze gamma-ray spectra of RRTNIS using an ANN, after recording gamma-ray spectra from 20 detectors using a data acquisition system and performing energy calibration, the resulting accumulated spectrum is pre-processed by applying energy threshold of 0.6 MeV, compressing the spectrum to have only 246 channels if necessary, and normalizing it to the total gross count of the spectrum. Subsequently, the processed spectrum is fed into the trained ANN model, the outputs of which comprise the count contribution of 20 elements as well as background radiation.\u003c/p\u003e \u003cp\u003eTo estimate the effect of statistical uncertainties on ANN results, a \u0026ldquo;small\u0026rdquo; dataset of size 500 was constructed for each measured spectrum (for instance to build the error bars in Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e or cloud around the measured points in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e and \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). In this process, the counts of each channel of the recorded spectrum were considered to follow a Poisson distribution with the expected counts (λ) equal to the counts in the channel. Subsequently, 500 counts were sampled for each channel using their corresponding Poisson distributions, resulting in 500 gamma spectra for a measured gamma spectrum. The dataset were pre-processed and fed into the network, resulting in 500 sets of outputs. For each output, a Gaussian distribution was fitted to the corresponding 500 data points, with its mean and standard deviation considered as the output of the ANN.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1. Data preparation for training ANN\u003c/h2\u003e \u003cp\u003eTo train the ANN models in this work, 100,000 gamma spectra were generated by mixing the gamma signatures of 20 pure elements, each with a randomly sampled proportion, and of the correlated background. The generated spectra were divided into 70% training data, 15% testing data, and 15% validation data to train the ANN models described in the following sections. It is worth mentioning that the training data are utilized to optimize the parameters of the ANN models during the training process. Validation data are used to evaluate the generalization ability of the trained models and to prevent overfitting during network training. Additionally, testing data are employed to assess the performance of the trained models on unseen data [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2. Feed Forward Neural Network\u003c/h2\u003e \u003cp\u003eThe Feed Forward Neural Network (FFNN), also commonly known as the multilayer perceptron (MLP), represents one of the foundational architectures in the realm of artificial neural networks. In FFNNs, information flows strictly in one direction, from input to output layers, without forming any feedback loops. This unidirectional flow enables FFNNs to effectively process and extract features from complex datasets, making them particularly adept at tasks such as regression, classification, and function approximation [22;23]. Moreover, FFNNs typically comprise multiple layers of interconnected neurons, including input, hidden, and output layers, with each layer playing a distinct role in information processing and abstraction. To determine the optimal FFNN structure in this study, various parameters were systematically explored within predefined ranges. These parameters included the number of hidden layers (1 or 2), the number of neurons in each layer (ranging from 50 to 1500 with an increment of 50), the dropout rate (ranging from 0 to 0.2) and the initial learning rate (ranging from 1e-4 to 1e-3 with an increment of 1e-4). The number of epochs was set 200, and from the 100 randomly tested network structures, the network with the lowest mean square error was identified as the optimal structure. The specifications of this optimal network structure are detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. As neural networks inherently possess a stochastic nature, different outputs can be obtained for a specific network structure through a \u0026ldquo;multi-start\u0026rdquo; process for the initialization of the ANN, including randomized initial weights. To mitigate the effects of randomness, the optimal FFNN architecture was trained independently across 10 runs using identical data. By averaging the outputs obtained from these 10 independent runs, we aimed to obtain a more stable and representative estimation of the network's performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3. Convolutional Neural Network\u003c/h2\u003e \u003cp\u003eThe Convolutional Neural Network (CNN) stands out as a groundbreaking architecture specifically designed to excel in tasks involving visual imagery, such as image classification, object detection, and image segmentation. With their ability to automatically learn hierarchical representations of visual data, CNNs have revolutionized various fields, including computer vision, medical image analysis, and autonomous driving. Recently, CNNs have been applied in analyzing gamma-ray spectra specifically for radioisotope identification [29;30]. Unlike traditional neural networks, CNNs leverage convolutional layers to systematically extract and learn spatial hierarchies of features from input spectra. These convolutional layers employ filters, also referred to as kernels, to perform localized operations across the input spectrum, thereby capturing essential patterns and structures. Additionally, CNNs often incorporate pooling layers to down-sample feature maps, reducing computational complexity and enhancing translation invariance. To determine the optimal CNN structure in this study, various parameters were systematically explored within predefined ranges. These parameters included the number of filters in the Conv1D layers, ranging from 4 to 16, with kernel sizes of 3 to 15. Additionally, the MaxPool layers were tested with stride values of 1 and 2, and kernel sizes of 2, 3 and 5. The number of dense layers was varied between 1 and 2, with the number of neurons in each dense layer ranging from 50 to 1500 with increments of 50. Dropout rates were evaluated from 0 to 0.2. The initial learning rate was adjusted from 1e-4 to 1e-3 with an increment of 1e-4. The number of epochs was set to 300, and from the 200 randomly tested network structures, the network with the lowest mean square error was identified as the optimal structure. The specifications of this optimal network structure are detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. To mitigate the effects of randomness, the optimal CNN architecture was trained independently across 10 runs using identical data and the outputs were averaged over 10 runs.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSpecifications of the optimal FFNN and CNN models used in this work.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eANN type\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eFFNN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNetwork parameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSize (Type)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNetwork parameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSize (Type)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of inputs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e246\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of inputs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e246\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of hidden layers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eConv1D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e# filters: 8 - kernel: (5\u0026times;1) - dim: (246,8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of neurons in the\u003c/p\u003e \u003cp\u003efirst hidden layer (dropout rate)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1100 (0.03)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eConv1D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e# filters: 8 - kernel: (3\u0026times;1) - dim: (244,8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of neurons in the\u003c/p\u003e \u003cp\u003esecond hidden layer (dropout rate)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e300 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaxPool1D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStrides: 2 - kernel: (2\u0026times;1) - dim: (122,8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActivation function of hidden layers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelu\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eConv1D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e# filters: 4 - kernel: (3\u0026times;1) - dim: (120,4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActivation function of output layer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSoftmax\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaxPool1D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStrides: 2 - kernel: (2\u0026times;1) - dim: (60,4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLoss function\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean square error\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFlatten\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003edim:240\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining algorithm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAdam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of Dense layers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInitial learning rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.0005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of neurons in the\u003c/p\u003e \u003cp\u003efirst dense layer (dropout rate)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e900 (0.05)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of outputs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of neurons in the\u003c/p\u003e \u003cp\u003esecond dense layer (dropout rate)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e200 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eActivation function of dense layers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRelu\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eActivation function of output layer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSoftmax\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLoss function\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean Square Error (MSE)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTraining algorithm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAdam\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInitial learning rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.0005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of outputs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.4. Explainable artificial intelligence\u003c/h2\u003e \u003cp\u003eAs machine learning models like ANNs continue to advance, they often operate as opaque \"black boxes,\" making it hard to understand how they make decisions. To tackle this, the field of explainable artificial intelligence (XAI) has emerged [34;35]. XAI develops ways to clarify how machine learning models work. Applying XAI to the analysis of the gamma-ray spectrum using full information [30;36] helps to find out which parts of the gamma spectrum contribute to estimating a parameter in output of a machine learning method. This not only makes results more trustworthy but also helps people understand and make decisions faster during real-world tasks. By giving clear explanations, users can better understand and handle alerts, making applications more reliable and effective. The performance of different XAI methods have been examined in the full spectrum analysis of gamma-ray spectra using ANNs in the field of radioisotope identification [36;37], with the best performance obtained from the SHAP (SHapley Additive exPlanations) method in terms of providing more precise explanations based on gamma signatures for correctly identified and misidentified isotopes. SHAP [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] is based on the concept of Shapley values from cooperative game theory [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. In the context of machine learning, SHAP values quantify the impact of each feature on the model's output by considering all possible combinations of features and their contributions. This approach ensures that the importance of each feature is evaluated in relation to its interactions with other features, providing a more comprehensive understanding of their influence on the model's predictions. In this work, the SHAP library was installed using the pip package manager and implemented within the Spyder environment.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Results and discussion","content":"\u003cp\u003eThe correlation coefficients between the targets and predicted outputs of FFNN and CNN models were studied for all elements as well as background radiation related to both train and test data. The study demonstrated the efficient training of both FFNN and CNN models. These trained networks were subsequently utilized to analyze gamma spectra obtained during the inspection of containers with various targets in following section.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Container inspection using trained ANN models\u003c/h2\u003e \u003cp\u003eThe 1D inspection of the container was conducted in slices with a depth of 10 cm and a 5 cm horizontal shift (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), with a measuring time of 10 minutes and with various targets as outlined in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The position of the middle of the targets was approximately fixed at 75 cm deep inside the container, and a gamma-ray spectrum was obtained for each slice of the container. The positions of slices were determined using neutron time-of-flight, assuming a constant neutron velocity of 5.13 cm/ns (14-MeV neutrons). The distance of each slice from the iron wall of the container, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, was considered, with the center of each slice used as the reference point. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and (b) illustrates the depth profiles of C, N, and O count contributions for a 1 kg of TNT simulant as the target in the container (without cargo matrix around), employing FFNN and CNN models for analyzing the obtained gamma-ray spectra. Similar profiles were obtained for 1 kg of each of the other simulants presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The error bars of count contributions at each depth were derived from a \u0026ldquo;small\u0026rdquo; dataset of size 500 constructed for each measured spectrum, as detailed in section \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e. In addition to C, O, and N, other elements were also inspected for material and threat identification in a container. As an example, the depth profiles of Cl and S count contributions corresponding to measuring Yperite simulant (mustard gas, a warfare chemical) as the target (still without cargo matrix around) were obtained, and the results demonstrated precise identification of this simulant. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and (b) highlights the importance of further investigating the area around 75 cm (65 to 85 cm) within the container to identify material in this region. Initially, we assessed how well the ANN models can reconstruct the gamma-ray spectrum at a distance of 75 cm for TNT explosive as an example case. The reconstructed spectra, obtained using the mean values from both FFNN and CNN models, are compared to the measured spectrum in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(c). According to Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(c), both ANN models perform fairly well in reconstructing the measured spectrum. Furthermore, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(d) to (i) overlays the SHAP values corresponding to the C, O, and N elements on the measured spectrum, to visualize the impact of individual energy bins on the predictions made by the FFNN and CNN models. These SHAP values were obtained from both the FFNN and CNN models. The measured spectrum is represented by a white line on the x-axis depicting energy (MeV) and normalized counts on the y-axis. The color map indicates the magnitude of the SHAP values, offering insights into the contribution of each energy bin to the overall prediction. Both models primarily focus on the major gamma lines of the elements C (4.44 MeV) and O (6.13 MeV) to predict their count contributions in the TNT gamma spectrum. However, the CNN model demonstrates a more pronounced emphasis on these gamma lines compared to the FFNN model. As the 5.11 MeV gamma line of N element is not observed in the TNT spectrum in this measurement, the CNN model distinctly emphasizes the 2.31 MeV gamma line to predict the count contribution of N.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFinally, the count contributions should be converted to weight fractions for C, N, and O for threat identification purposes. To achieve this, calibration lines were obtained separately for each element and for both FFNN and CNN models. For each element, the maximum mean values in the corresponding depth profiles of three simulants, in addition to the (0,0) point, were used to map the count contribution to the weight fraction. The simulants RDX, Cocaine and Peroxide methylethylketone shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e were used for C element; RDX, Nitromethane, and Tetranitromethane for N element; and RDX, Cocaine and Tetranitromethane for O element.\u003c/p\u003e \u003cp\u003eThe obtained calibration lines were then applied to convert the count contributions of C, N, and O elements obtained using ANN models to weight fractions for all simulants presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and (b) shows variations in the count contributions in any selected suspicious area. To mitigate these variations, the mean of the three maximum count contributions in the selected region was considered for each element and then converted to weight fractions using the calibration lines, followed by normalization to ensure that the sum of the weight fractions for C, N, and O equals one. The resulting normalized weight fractions were then considered as a point in the CNO barycentric triangle [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. The CNO barycentric triangle illustrates material compositions based on normalized weight fractions of C, N, and O. Each vertex represents pure C, N, or O, while the interior shows combinations of these elements. Points within the triangle indicate mixtures of C, N, and O, offering a visual representation of material compositions in our plots. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e depicts the theoretical point represented by a black triangle (▲) in the barycentric triangle obtained from weight fractions for each simulant (true theoretical composition), except for Yperite, compared to the measured points obtained using FFNN and CNN models, respectively. The measured points are represented as black points (●) for FFNN model and star symbols (★) for CNN model for each simulant listed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. To illustrate the uncertainty in the location of the measured point in barycentric triangle, the measured points were calculated for \u0026ldquo;small\u0026rdquo; dataset of size 500 constructed for each measured spectrum (as detailed in section \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e) using both ANN models. The uncertainties obtained are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, with grey points forming a \u0026ldquo;cloud\u0026rdquo; around the measured points. Additionally, blue points represent some of the illicit drugs, red points represent some explosives, and green points represent some benign materials, illustrating the distribution of materials within the CNO barycentric triangle. As evident in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, both ANN models predict the measured points close to the theoretical points, with the FFNN model showing better performance. The better performance of FFNN is further supported by the depth profiles (TNT results shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and (b)), as the FFNN exhibits smaller standard deviations for each element.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo evaluate the performance of the ANN models in the presence of a matrix, 10 kg of TNT was hidden within the container behind 40 and 80 cm wood bales, with a density of 0.2 g.cm\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e (mimicking an organic cargo). In another series of experiments, the same TNT target was hidden behind 40 and 105 cm iron boxes filled with bundles of iron wire, with an apparent density of 0.2 g.cm\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e displays the depth profiles of C, N and O count contributions obtained using both FFNN and CNN models for 10 kg of TNT hidden in wood and iron matrices. According to the Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the distances of 70\u0026ndash;90 cm and 115\u0026ndash;135 cm were considered as the suspicious areas for the TNT hidden at depths of 40 cm and 80 cm in the wood matrix, respectively. Furthermore, the distances of 85\u0026ndash;105 cm and 160\u0026ndash;180 cm were considered as the suspicious areas for the TNT hidden 40 cm and 105 cm deep inside the iron matrix, respectively. The resulting measured points, along with their uncertainty obtained using ANN models in the CNO barycentric triangle for TNT hidden behind wood and iron matrices have been shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe depth profiles of C and O are obscured by the large background profiles of the wood matrix, particularly for TNT hidden in the wood matrix at a depth of 40 cm. This effect is noticeable in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e (a) and (b), where the weight fraction of N is underestimated.\u003c/p\u003e \u003cp\u003eThe SHAP values were calculated for N element for both FFNN and CNN models at distances of 80 cm and 125 cm, which are at the middle of suspicious regions shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, for TNT behind 40 cm and 80 cm of wood matrix, respectively. The overlay of SHAP values on the measured spectra of TNT is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIt is obvious in parts (a) and (c) of Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e that the measured gamma-ray spectrum is mainly influenced by intense gamma lines of C and O for TNT behind 40 cm of wood. The effect is less pronounced for TNT behind 80 cm of wood, resulting in better visibility of the gamma lines of N and consequently facilitating the estimation of N using its major gamma lines (2.31 and 5.11 MeV) by both models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe depth profiles, particularly for N, are significantly affected by the poor counting statistics when TNT is behind 105 cm of iron, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e (g) and (h). This effect is evident in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e (g) and (h), showing large uncertainties and fluctuations of the measured points.\u003c/p\u003e \u003cp\u003eDue to this critical situation, SHAP values were calculated for the C element for both FFNN and CNN models at distances of 95 cm and 170 cm, which are at the middle of the suspicious regions shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, for TNT behind 40 cm and 105 cm of iron matrix, respectively. The overlay of SHAP values on the measured spectra of TNT for the C element is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e (a) to (d). Additionally, the overlay of SHAP values for the Fe element is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e (e) and (f) for TNT behind 105 cm of iron matrix for both models.\u003c/p\u003e \u003cp\u003eIt is evident in parts (a) and (c) of Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e that both models still mainly utilize the 4.44 MeV gamma line of the measured gamma-ray spectrum to predict C for TNT behind 40 cm of iron, albeit with a lower contrast compared to Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(d) and (g). The effect is much more pronounced for TNT behind 105 cm of iron, where no specific region is used for predicting C, but rather the entire spectrum. In fact, there is no specific peak in the measured spectrum corresponding to C, N and O elements, and the measured spectrum is similar to the gamma signature of pure Fe shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e(c). Since iron is the most abundant material inside the container in this case, both ANN models effectively predict the count contribution of Fe using its major gamma line, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e(e) and (f). Even in this situation where the gamma-rays originating from different elements are heavily obscured by the iron matrix, both ANN models could still fairly determine the position of the threat (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(g) and (h)) and the category of the unknown material as explosive (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(g) and (h)).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs can be seen from the depth profiles presented in Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, both the FFNN and CNN models identify the suspicious areas at the same positions (i.e. depths) within the container. This consistency enhances the reliability of container inspection, as it indicates that two different ANN models pinpoint the same locations.\u003c/p\u003e \u003cp\u003eIndeed, the consistency observed in the identification of suspicious areas by both FFNN and CNN models was further supported by the results of XAI techniques. Specifically, the SHAP values overlaid on the measured spectrum provide insights into the contribution of individual energy bins to the models' predictions. Figures\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e, and \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e illustrate how both models primarily focus on major gamma lines of C, O, and N elements to predict their count contributions accurately. Additionally, the CNN model demonstrates a more pronounced emphasis on these gamma lines compared to the FFNN model, highlighting differences in their predictive strategies. This analysis underscores the reliability and interpretability of the models' predictions, enhancing trust in their performance in identifying suspicious areas within cargo containers.\u003c/p\u003e \u003cp\u003eFurthermore, the results from Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, which depicts the performance of the models in positioning the measured points in the CNO barycentric triangle for simulants without matrices, and Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e which present the results for TNT simulant hidden within wood and iron matrices, indicate effective performance in identifying material category within the CNO barycentric triangle. Generally, a slightly better performance in terms of proximity to the theoretical points (except for TNT in presence of 105 cm iron matrix) was observed for the FFNN model. However, both models displayed robust predictive capabilities, allowing for the accurate identification of suspicious regions within cargo containers.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eIn this study, we conducted a comprehensive evaluation of eXplainable Artificial Intelligence (XAI) to enhance the identification of suspicious areas within cargo containers, integrating tagged fast neutron activation analysis with Feed Forward Neural Network (FFNN) and Convolutional Neural Network (CNN) models. Employing innovative approach outlined in our investigation, our aim was to enhance container screening effectiveness for global trade security. Through the analysis of depth profiles depicting the count contributions of elements, we observed consistent identification of suspicious regions within the container by both FFNN and CNN models, affirming the reliability of their predictions. Furthermore, utilization of XAI provided valuable insight into the predictive strategies and decision-making processes of the ANN models, enhancing their interpretability and overall reliability.\u003c/p\u003e \u003cp\u003eWe further extended our investigation by visualizing material compositions using the CNO barycentric triangle. This provided valuable insights into the identification of material categories for analyzed targets, both with and without matrices present in the container. The results demonstrated effective performance in identifying threat materials within the respective categories using both ANN models. However, it is noteworthy that the CNN model exhibits a more pronounced emphasis on the major gamma lines of elements compared to the FFNN model, ensuring highly reliable results and potentially serving as a complementary support for FFNN results.\u003c/p\u003e \u003cp\u003eFor future work, a promising direction is advancing deep learning methods for spectrum analysis in general, particularly considering the insights gained from XAI results. Improving the performance of deep learning models through exploration of different network architectures could be a fruitful area for research. Additionally, studying the likelihood probability of the measured points relative to the closest reference materials (such as explosives, illicit drugs, and benign substances) in the triangle plots could provide valuable insights. Investigating the step-by-step non-negative least square method currently utilized for gamma-ray spectra analysis in RRTNIS presents another avenue for exploration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was funded by the Horizon Europe Framework Programme under the ENTRANCE (Grant Agreement No. 883424) and UnderSec (Grant Agreement No. 101121288) projects. The authors gratefully acknowledge the active involvement and support of the Croatian Ministry of Interior, the Croatian Ministry of Finance \u0026ndash; Customs Administration, and the Port of Rijeka d.d.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHadi Shahabinejad analyzed data, developed and wrote the analysis codes and drafted the manuscript; Davorin Sudac, Karlo Nad, Jasmina Obhođa\u0026scaron;, Isabelle Espagnon, Clotilde de Sainte Foy, Bertrand Perot, Cedric Carasco, Alix Sardet, Jessica Delgado, Felix Pino, Sandra Moretto, contributed to hardware and software optimization and acquisition of experimental data; Edwin Friedmann, Jean Philippe Poli, Isabelle Espagnon, Clotilde de Sainte Foy, Bertrand Perot, Cedric Carasco, Alix Sardet, Jessica Delgado, Felix Pino, Sandra Moretto analyzed data; Christine Mer and Guillaume Sannie secured funds, planned experiments, and contributed to the acquisition of the experimental data; Hadi Shahabinejad and Jasmina Obhodas conceived the idea of developing AI for experimental data; All authors discussed the results and commented on the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePesente, S. et al. Detection of hidden explosives by using tagged neutron beams with sub-nanosecond time resolution, Nuclear Instruments and Methods in Physics Research A. 531, 657\u0026ndash;667 (2004).\u003c/li\u003e\n\u003cli\u003eViesti, G. et al. Detection of hidden explosives by using tagged neutron beams: Status and perspectives, Nuclear Instruments and Methods in Physics Research B. 241, 748\u0026ndash;752 (2005).\u003c/li\u003e\n\u003cli\u003eSardet A. et al. Performances of C-BORD\u0026rsquo;s Tagged Neutron Inspection System for Explosives and Illicit Drugs Detection in Cargo Containers, 68, 346 \u0026ndash; 353 (2021). \u003c/li\u003e\n\u003cli\u003ePerret, G., Perot, B., Artaud, J-L., and Mariani, A. EURITRACK tagged neutron inspection system design, Journal of Physics: Conference Series 41, 375\u0026ndash;383 (2006). \u003c/li\u003e\n\u003cli\u003ePerot, B. et al. Development of the EURITRACK tagged neutron inspection system, Nuclear Instruments and Methods in Physics Research B. 261, 295\u0026ndash;298 (2007). \u003c/li\u003e\n\u003cli\u003eValkovic, V., Sudac, D., Nad, K., and Obhodas J. Container Inspection Utilizing 14 MeV Neutrons, IEEE Transactions on Nuclear Science, 63 (3), 1536-1543 (2016). \u003c/li\u003e\n\u003cli\u003eSardet, A. et al. Design of the rapidly relocatable tagged neutron inspection system of the C-BORD project, in Proc. IEEE Nucl. Sci. Symp., Med. Imag. Conf. Room-Temp. Semiconductor Detect. Workshop (NSS/MIC/RTSD), Strasbourg, France, Oct., 1\u0026ndash;5 (2016).\u003c/li\u003e\n\u003cli\u003eC-BORD project, Effective Container Inspection at BORDer Control Points, https://www.cbord-h2020.eu. \u003c/li\u003e\n\u003cli\u003eFontana, C.L. et al. Detection System of the First Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), Developed in the Framework of the European H2020 C-BORD Project, Physics Procedia, 90, 279-284 (2017). \u003c/li\u003e\n\u003cli\u003ePino, F. et al. Advances on the development of the detection system of C-BORD\u0026rsquo;s rapidly relocatable tagged neutron inspection, International Journal of Modern Physics: Conference Series, 48, 1860125 (2018).\u003c/li\u003e\n\u003cli\u003ePin, F. et al. Detection module of the C-BORD Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), Nuclear Inst. and Methods in Physics Research, A. 986, 164743 (2021). \u003c/li\u003e\n\u003cli\u003ePino, F. et al. Non-intrusive inspection of cargo containers using the C-BORD Rapidly Relocatable Tagged Neutron Inspection System, JINST. 17, T12005 (2022). \u003c/li\u003e\n\u003cli\u003eENTRANCE project, \u0026quot;Efficient risk-based inspection of freight crossing borders without disrupting business\u0026quot;, https://www.entrance-h2020.eu. \u003c/li\u003e\n\u003cli\u003eGardner, R.P., Sood, A., Wang, Y.Y., Liu, L., Guo, P., Gehrke, R.J. Single peak versus library least-squares analysis methods for the PGNAA analysis of vitrified waste, Applied Radiation and Isotope, 48, 1331-1335 (1997).\u003c/li\u003e\n\u003cli\u003eObhodas, J., Sudac, D., and Valkovic, V. Development of the Quality Assurance/Quality Control Procedures for a Neutron Interrogation System, IEEE Transactions on Nuclear Science. 63 (3), 1544 \u0026ndash; 1550 (2016).\u003c/li\u003e\n\u003cli\u003eShahabinejad, H., and Vosoughi, N. Analysis of complex gamma-ray spectra using particle swarm optimization, Nucl. Instrum. Methods Phys. Res. A. 911, 123-130 (2018).\u003c/li\u003e\n\u003cli\u003eMeric, I., Johansen, G.A., Holstad, M.B., Wang, J., Gardner, R.P. Produced water characterization by prompt gamma-ray neutron activation analysis, Meas. Sci. Technol., 22, 125701 (2011).\u003c/li\u003e\n\u003cli\u003eCaciolli, A., Baldoncini, M., Bezzon, G.P., Broggini, C., Buso, G.P., Callegari, I., Colonna, T., Fiorentini, G., Guastaldi, E., Mantovani, F., Massa, G., Menegazzo, R., Mou, L., Rossi Alvarez, C., Shyti, M., Zanon, A., Xhixha, G. A new FSA approach for in situ \u0026gamma; ray spectroscopy. Sci. Total Environ. 414, 639-645 (2012).\u003c/li\u003e\n\u003cli\u003eBaldoncini, M., Alb\u0026eacute;ri, M., Bottardi, C., Chiarelli, E., Raptis, K.G.C., Strati, V., Mantovani, F. Investigating the potentialities of Monte Carlo simulation for assessing soil water content via proximal gamma-ray spectroscopy, J. Environ. Radioact. 192, 105-116 (2018).\u003c/li\u003e\n\u003cli\u003eShahabinejad H., and Vosoughi N. SGSD: A novel Sequential Gamma-ray Spectrum Deconvolution algorithm. Annals of Nuclear Energy. 132, 369-380 (2019). \u003c/li\u003e\n\u003cli\u003eAlamaniotis M., and Jevremovic T. Hybrid fuzzy-genetic approach integrating peak identification and spectrum fitting for complex gamma-ray spectra analysis. IEEE Transactions on Nuclear Science. 62, 1262-1277 (2015).\u003c/li\u003e\n\u003cli\u003eKamuda M., Stinnett J., Sullivan C. Automated isotope identification algorithm using artificial neural networks. IEEE Transactions on Nuclear Science. 64, 1858-1864 (2017).\u003c/li\u003e\n\u003cli\u003eShahabinejad H., Vosoughi N., Saheli F., \u0026ldquo;Processing scintillation gamma-ray spectra by artificial neural network\u0026rdquo;, Journal of Radioanalytical and Nuclear Chemistry, 325, 471-483 (2020). \u003c/li\u003e\n\u003cli\u003eSainte Foy C.D. et al. Optimization of the ENTRANCE Rapidly Relocatable Tagged Neutron Inspection System, IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), Italy, 1-7 (2022). \u003c/li\u003e\n\u003cli\u003eFausett, L.V. Fundamentals of neural networks: architectures, algorithms, and applications, vol 3. prentice-Hall Englewood Cliffs (1994).\u003c/li\u003e\n\u003cli\u003eHaykin, S. Neural networks: a comprehensive foundation. Prentice Hall PTR (1994). \u003c/li\u003e\n\u003cli\u003eRyu, J., Park, C., Park, J., Cho, N., Park, J., and Cho, G. Development of Neural Network Model With Explainable AI for Measuring Uranium Enrichment, IEEE TRANSACTIONS ON NUCLEAR SCIENCE, 68, 2670-2681 (2021).\u003c/li\u003e\n\u003cli\u003eHossny, K., Hossny, A.H., Magdi, S. Detecting shielded explosives by coupling prompt gamma neutron activation analysis and deep neural networks. Sci Rep. 10, 13467 (2020).\u003c/li\u003e\n\u003cli\u003eKamuda, M., Zhao, J., Huff, K. A comparison of machine learning methods for automated gamma-ray spectroscopy. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 954, 161385 (2020).\u003c/li\u003e\n\u003cli\u003eGomez-Fernandez, M., Wong, W.K., Tokuhiro, A., Welter, K., Alhawsawi, A.M., Yang, H., et al. Isotope identification using deep learning: An explanation. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 988, 164925 (2021). \u003c/li\u003e\n\u003cli\u003eCarasco, C., P\u0026eacute;rot, B., Sardet, A. Measuring hydrogen with fast neutrons: Application to organic materials identification in cargo containers, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 951, 163030 (2020).\u003c/li\u003e\n\u003cli\u003eGong, K., He, D.-D., Chen, S.-Y., Jing, S.-W., Zheng, Y.-L. BP neural network analysis for identification of explosive in package by tagged neutron method, Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers, Detectors and Associated Equipment. 101721, 165810 (2021).\u003c/li\u003e\n\u003cli\u003eGong, K., Xiao, S.-J., Jing, S.-W., Zheng, Y.-L. Back propagation neural network analysis for the detection of explosives based on tagged neutron, Journal of Radioanalytical and Nuclear Chemistry. 326, 329-336 (2020).\u003c/li\u003e\n\u003cli\u003eAdadi, A., Berrada, M. Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access. 6, 52138\u0026ndash;52160 (2018).\u003c/li\u003e\n\u003cli\u003eBarredo Arrieta, A., Dıaz-Rodrıguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado A. Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI. Information Fusion. 58, 82\u0026ndash;115 (2020). \u003c/li\u003e\n\u003cli\u003eBandstra, M.S., Curtis J.C., Ghawaly, J.M. Jr, Jones, A.C., Joshi, T.H.Y. Explaining machine-learning models for gamma-ray detection and identification. PLoS ONE 18(6), e0286829 (2023). \u003c/li\u003e\n\u003cli\u003eJeon, B., Kim, J., Moon, M. Explanation of Deep Learning\u0026ndash;Based Radioisotope Identifier for Plastic Scintillation Detector, Nuclear Technology, 209, 1, 1-14, (2023).\u003c/li\u003e\n\u003cli\u003eLUNDBERG S.M. and LEE S.-I. A Unified Approach to Interpreting Model Predictions, Proc. 31st Int. Conf. Advances in Neural Information Processing Systems, Long Beach, California, December 4\u0026ndash;9, 4768 (2017). \u003c/li\u003e\n\u003cli\u003eCarasco, C. et al. Data Acquisition and Analysis of the UNCOSS Underwater Explosive Neutron Sensor, 2nd International Conference on Advancements in Nuclear Instrumentation, Measurement Methods and their Applications, IEEE, (2011). \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Threat material identification, Neutron activation, Spectrum analysis, Deep learning, eXplainable Artificial Intelligence (XAI), Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS)","lastPublishedDoi":"10.21203/rs.3.rs-4661721/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4661721/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHere we present an innovative approach for detecting threat materials within a sealed container by integrating tagged fast neutron activation analysis with Explainable Artificial Intelligence (XAI). Two AI models, a Feed-Forward Neural Network (FFNN) and a Convolutional Neural Network (CNN), were developed to analyze the emitted gamma rays to identify materials like explosives and drugs based on depth profiles of carbon, nitrogen, and oxygen concentrations. XAI was applied to make the models' decision-making process transparent. The method is adaptable to various spectrometric analyses. We demonstrate its effectiveness using data obtained by the Rapidly Relocatable Tagged Neutron Inspection System (RRTNIS), which is irreplaceable for inspecting sealed cargo containers, despite challenges such as variable material placement, background noise, and shielding effects. Our approach successfully locates and categorizes threat materials, both alone and within surrounding materials, at various locations within sealed cargo containers.\u003c/p\u003e","manuscriptTitle":"Integrating Tagged Neutron Inspection with Explainable AI for Threat Material Identification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-07-24 10:36:07","doi":"10.21203/rs.3.rs-4661721/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"903d1391-3851-4cdd-a9bd-cc2a30d2ab87","owner":[],"postedDate":"July 24th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-04-17T03:23:40+00:00","versionOfRecord":[],"versionCreatedAt":"2024-07-24 10:36:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4661721","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4661721","identity":"rs-4661721","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}