Machine learning assisted real-time acoustic monitoring of laser cleaning in Heritage conservation

Machine learning assisted real-time acoustic monitoring of laser cleaning in Heritage conservation | 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 Machine learning assisted real-time acoustic monitoring of laser cleaning in Heritage conservation Georgios D. Barmparis, Anastasios-Nikolaos Raikidis, Kristalia Melessanaki, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6565434/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 04 Dec, 2025 Read the published version in npj Heritage Science → Version 1 posted 9 You are reading this latest preprint version Abstract Laser-assisted cleaning has become an indispensable tool in heritage conservation due to its precision, control, and environmentally friendly nature. However, the complexity of deposition layers and the fragile condition of original surfaces necessitate careful monitoring to avoid irreversible damage. This work explores the integration of machine learning-assisted real-time acoustic monitoring in laser cleaning processes to enhance conservation efforts. By combining acoustic signals generated during laser-material interaction with machine learning, we elevate the precision and reliability of laser cleaning in the delicate context of cultural heritage restoration. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Over the past 30 years, laser-assisted removal of unwanted material has become a common cleaning tool in Heritage conservation, often replacing conventional methods based on chemicals and mechanical action. This is due to its unique advantages, including selective and gradual material removal, high precision and control, and its environmentally friendly nature [ 1 , 2 ]. However, this delicate and irreversible process requires careful selection of irradiation parameters and a thorough understanding of ablation mechanisms, especially given the complex nature of deposition layers and the fragile condition of original heritage surfaces. Self-limiting processes control the near-infrared (NIR) laser cleaning of black pollution crusts from stonework, ensuring that the cleaning intervention halts immediately after the unwanted crust is removed. Consequently, laser cleaning has become increasingly popular in restoration, as it effectively and safely reveals the original surfaces of cultural heritage assets, such as the Acropolis Sculptures in Athens [ 2 , 3 ]. Successful applications of laser cleaning have also been reported for various materials [ 1 – 7 ], including the removal of burial crusts from stone sculptures and leather objects, cleaning soiling from paper, and eliminating corrosion layers from historical metal objects [ 5 ]. Additionally, it has proven effective in removing tarnishing from gilded silver threads [ 6 ], among other uses. Laser cleaning is particularly valuable in addressing a wide range of cleaning challenges. These often involve situations in which self-limiting conditions may not apply, i.e., challenges in which the critical parameters needed to remove unwanted material are very close to, or even exceed, the thresholds that could damage the original substrate [ 7 ]. One of the most challenging areas of laser cleaning research is the removal of aged or polymerized varnish and overpainting layers from artworks, particularly those that are painted. These layers are composed of polymeric materials that are transparent to near-infrared (NIR) laser radiation. In contrast, most degraded polymeric overlayers highly absorb ultraviolet (UV) laser radiation, making them an effective alternative for layer-by-layer ablation and the controlled removal of unwanted materials. However, the fact that varnish films and paint layers exhibit very similar optical properties, absorbing UV radiation highly, poses limitations to their laser cleaning, as self-limiting conditions do not apply. For this reason, a critical assessment of the cleaning process and real-time monitoring of its progress are crucial. Over the years, several imaging and spectroscopic techniques have been explored to assess the results and monitor the cleaning process in heritage conservation. Imaging techniques offer a non-destructive and non-invasive approach and are the most widely used. Colorimetry, multispectral reflectography, and optical coherence tomography (OCT) have been employed to assess the color [ 8 , 9 ] and stratigraphy [ 10 ] of the objects under study and during restoration. Microscopies, including optical microscopy (OM), scanning electron microscopy (SEM), and atomic force microscopy (AFM), have also been used to investigate the surface morphology of the treated objects [ 11 ]. In parallel, the need for minimal chemical alteration to the underlying original surfaces has influenced several studies that employed spectroscopic techniques, such as Raman spectroscopy and Fourier Transform Infrared Spectroscopy (FTIR) [ 12 ], as well as gas chromatography/mass spectrometry (GC-MS), to investigate the chemistry of the cleaned surfaces. Significant research has also been focused on the complementary application of imaging and chemical techniques for a thorough approach to cleaning assessment [ 13 ]. Careful post-treatment appraisal and monitoring during the restoration process are becoming increasingly critical in the context of laser cleaning. While laser cleaning offers precision, it also presents challenges, especially in finding the right balance between effective cleaning and preserving the substrate. In industrial applications, the effectiveness of the cleaning process is evaluated based on two key factors: temporal aspects, such as cleaning time and removal rates, and spatial aspects, which focus on the complete elimination of unwanted materials. However, in heritage conservation, the primary concern is to protect the integrity and chemistry of the original surface. Significant research has focused not only on the ideal in-situ evaluation of the results [ 14 , 15 , 16 , 17 ] but also on real-time monitoring of the process [ 18 ]. In contrast to industrial applications, where monitoring typically depends on post-process assessments of cleaning effectiveness and removal rates, laser cleaning of heritage objects requires a different focus. The most critical factor is to identify the key parameters to ensure irradiation stops at the right time, thereby preserving the original surface. Acoustic monitoring of the laser cleaning process offers unique advantages, particularly because it provides real-time results [ 18 , 19 , 20 , 21 , 22 ]. Early studies [ 23 , 24 , 25 , 26 ] highlighted its potential, and recent research efforts have increasingly focused on this concept [ 19 – 22 ]. This method involves recording and processing the acoustic signals generated when light interacts with matter. Previous studies [ 20 ] have demonstrated a correlation between the amplitude of the acoustic wave and the amount and composition of removed material. This relationship enables monitoring of the cleaning process progress, as the acoustic signal changes gradually when the overlayer is reduced, followed by an abrupt change upon reaching the substrate. One of the key advantages of this method is that the acoustic signals can be easily and non-invasively detected using non-contact air-coupled ultrasonic transducers, which have a frequency response in the MHz range. These transducers can be placed near the laser-treated surfaces, enabling online and real-time measurements without interrupting the cleaning process for analysis and data recording. The effectiveness of this approach has been tested on a series of technical encrusted marble coupons [ 20 ] and varnished painting mock-ups [ 21 ], which simulate various cleaning challenges. The relevant results have been cross-checked using other analytical techniques. The proposed acoustic monitoring strategy enables the identification of the critical laser pulse responsible for complete encrustation removal. This capability can assist conservators and restorers (C-Rs) in monitoring laser treatments and making informed decisions regarding the progress of the crust or overlayer removal process. Artificial Intelligence and Machine Learning (ML) algorithms are increasingly integrated into scientific research and industrial applications, offering powerful tools for data analysis, process optimization, and decision-making. ML models can identify patterns and correlations within experimental data, and they have been employed to various applications in cultural heritage research and conservation [ 27 – 33 ], such as to identify cracks [ 29 ] and other deterioration features [ 30 ], to identify rising damp on monuments promptly [ 31 ] or to study certain type of objects [ 33 ] to name a few. Herein, we discuss the implementation of ML in monitoring the laser cleaning process and ensuring a safe intervention on objects of heritage value. Among the available ML algorithms, we choose the Random Forest algorithm (RF), introduced by Breiman in 2001 [ 34 ], as an extension to the “the random subspace” method by Tim Kam Ho [ 35 ], which does a random selection of a subset of features to grow each decision tree. RF is an explainable tree-based ensemble learning method that has emerged as one of the most potent and versatile machine learning algorithms for classification and regression tasks. Among its advantages, an RF algorithm can handle high-dimensional datasets, model complex nonlinear relationships, perform well with limited amounts of data, and provide insights through feature importance metrics. Previous works have applied ML to optimize laser parameters in industrial cleaning applications [ 36 , 37 , 38 , 39 ] or to monitor and evaluate the cleaning results [ 39 , 40 , 41 ]. However, in Heritage science and conservation practice, the focus shifts away from material removal efficiency, as seen in industrial applications. The primary goal in heritage conservation is to determine the precise moment to stop the cleaning process, thereby preventing damage to the underlying historical surface that must be preserved and revealed. This approach differentiates conservation work from industrial processes and is consistently followed in this context. This study presents the first feasibility study to monitor the restoration process of sensitive cultural heritage objects in real-time. It combines acoustic wave signals produced during laser cleaning and explainable ML algorithms to identify the pulse that completely removes the unwanted paint without damaging the substrate. The methodology, including details on individual irradiation and monitoring parameters, as well as processing procedures and machine learning algorithms used in this study, is outlined in Fig. 1 . Methods Technical mock-ups [subsection] A series of technical mock-ups were created for data-gathering purposes [Fig. 2 (left)]. The base was made from thin slabs of white marble sourced from the Greek island of Thasos in the Northern Aegean Sea. Black acrylic paint (Motip Matt Black Acrylic Varnish) was applied to the marble from approximately 30 cm. The mock-ups were then left to dry naturally for at least one week before conducting irradiation experiments. This black acrylic overlayer, intended to simulate the "unwanted crust," was applied in multiple layers, resulting in varying thicknesses ranging from 8 to 55 µm. The mock-up design facilitated simple laser ablation processes characterized by self-limiting conditions. Cleaning methodology and criteria for quality assessment [subsection] Laser cleaning is affected by several parameters that are directly related to the materials involved [ 2 ]. The most critical parameters are the appropriate wavelength , λ, and pulse duration , τ p , of the laser system, the laser fluence , F, and the number of laser pulses applied , N. The wavelength and the pulse duration ensure the laser ablation mechanism is well-suited to the specific material. F is the energy, E, per unit area, S, of the irradiated surface, F = E/S (J/cm2). In addition to the laser fluence, we need to define the laser fluence threshold , F thr , i.e., the minimum fluence required to remove the unwanted material, and the critical laser fluence , F damage , i.e., the fluence above which the laser pulse causes damage to the underlying substrate. Finally, N refers to the total number of sequential laser pulses required to remove the unwanted material, which is closely related to the thickness of the unwanted layer or crust and the selected laser fluence. In this study, a QS Nd:YAG laser system emitting at λ = 1064 nm, with pulses of τ p = 10 ns, was employed. The cleaning efficiency of the given technical mock-ups has been well studied and characterized [ 23 ]. The laser fluence threshold was found to be F thr = 0.1 J/cm 2 , and the critical laser fluence, F damage = 2.1 J/cm 2 . During the data collection process, several values of F were carefully chosen between the F thr and the F damage thresholds to investigate the cleaning process at different paces. Laser pulses with low laser fluence slowly remove the unwanted material; thus, a high number of laser pulses is required to clean the surface, resulting in a slow pace. In contrast, laser pulses with high laser fluence transfer more energy to the irradiated surface, thus requiring fewer laser pulses to clean it, resulting in a high pace. Another parameter that affects the pace of the cleaning process is the thickness, d , of the unwanted material. The cleaning of materials with different thicknesses requires varying numbers of laser pulses at the same laser fluence. In this experiment, mock-ups characterized by a thickness, d i , of black crust are irradiated with a given laser fluence, F i , and a given number of applied laser pulses, N i . With this method, we collect data (acoustic signals) from groups of approximately 150 identical irradiated spots, all with the same overlayer thickness, di , laser fluence, Fi , and number of laser pulses, N i . In Fig. 2 , we present a schematic outline of the developed mock-ups and the various groups considered in this work, including thickness, laser fluence, and the number of laser pulses. Successive laser pulses lead to increasingly removing the unwanted material from the surface of the mock-ups. At this point, it is essential to highlight that an uncontrolled cleaning process can damage the underlying surface in the context of laser cleaning in Heritage conservation. To avoid damaging the substrate, we introduce the definition of the cleaning pulse , which is the pulse that sufficiently cleans the surface without damaging the substrate and is defined as the pulse that removes at least 75% of the irradiated area. We determine the ratio R between the cleaned area , S cleaned , and the irradiated area , S irradiated , by processing the collected images with the ImageJ software. In our experiments, we irradiate a few additional laser pulses after identifying the cleaning pulse to investigate the material's response beyond effective cleaning. In Fig. 2 , we present the digital microscope images taken after each of the laser pulses at a given spot with a paint thickness of d = 33 µm, irradiated with a laser fluence of F = 0.7 J/cm². The white area (exposed marble) is the cleaned area, while the black area is the total irradiated area. Pulse #11 cleans 53.9%, and Pulse #12 cleans 89.5% of the irradiated area. In this case, Pulse #12 is identified as the cleaning pulse. Experimental Setup [subsection] A schematic representation of the experimental setup is shown in Fig. 3 . The setup combines laser ablation and acoustic recording modalities. In all experiments, the relative position of the irradiated spot to the laser focusing lens is 32 cm. The acoustic piezoelectric transducer is located approximately 6 cm above and to the right of the spot at a 45-degree angle. A. Laser Cleaning System Α Q-Switched Nd:YAG laser (LITRON Lasers, TRLi 850 Series, Rugby, Warwickshire, England) was employed for performing the laser irradiation treatments. The laser system operated at the fundamental wavelength of 1064 nm, emitting pulses of 10 ns and with a variable repetition rate ranging from 0.25 to 1 Hz. The energy fluence values on the mock-ups ranged from 0.2 J/cm² to 1.2 J/cm² and were estimated by measuring the spot size (~ 0.26 cm²) of the focused beam on black photographic paper. All the irradiation experiments were performed in dry conditions. Optical imaging of the ablated regions was performed by utilizing a portable digital microscope (Dino-Lite Edge AM4113TFV2W) with a magnification in the range of ×50 to ×200. The depth of the laser-induced craters and the thickness of the over-layers were measured by means of a Portable Surface Roughness Tester profilometer (Mitutoyo America Corporation, Surftest SJ-410 Series, Aurora, IL, USA). B. Acoustic monitoring system The laser-induced acoustic response on the examined mock-ups was detected by an air-coupled transducer (NCT1-D7-P10, The Ultran Group, State College, PA, USA; nominal central frequency: 1 MHz; focal distance: 10 mm; numerical aperture: 0.31) placed at approximately 45 degrees in respect to the horizontal plane and in an out-of-focus position around 4 cm away from the irradiation region to avoid signal saturation effects. The signals were subsequently enhanced by two radio frequency amplifiers (TB-414-8A+, Mini-Circuits, Camberley, England; gain:31 dB) connected in series prior to their digitization by an oscilloscope (DSO7034A, Agilent Technologies, Santa Clara, CA, USA; bandwidth: 350 MHz; sample rate: up to 2 GSamples/s) which, in turn, was connected to a laptop computer equipped with custom-made software controlling the measurement procedures. The recorded waveforms were sampled at 1000 points over a 50 µs temporal window (corresponding to a 20 MSamples/s sampling rate) and bandpassed between 0.5 and 2 MHz to reduce mainly high-frequency noise before being saved to the computer as ASCII files. Recording synchronization was achieved through the trigger output of the laser source, which was connected to the oscilloscope’s second channel. C. Monitoring and Data Gathering During data collection, a total of 1,131 spots were irradiated on the studied mock-ups of black paint overlayers of various thicknesses. Each spot was exposed to varying sets of laser fluence, F i , and number of laser pulses, N i . The recorded waveform corresponding to each laser pulse incidence represents the detected acoustic pressure amplitude as a function of time, where positive and negative values can be interpreted as the compression and rarefaction regions of the propagating ultrasonic wave, respectively. The peak-to-peak amplitude, quantified as the difference between the maximum and minimum pressure values, has been demonstrated [ 18 ] to be directly proportional to the effective optical absorption coefficient at the irradiation wavelength of 1064 nm under typical energy fluence conditions. This relationship arises due to the localized thermoelastic expansion and subsequent generation of broadband ultrasonic waves following transient optical energy deposition. Therefore, the peak-to-peak amplitude parameter, along with the acoustic perturbation’s time-of-flight, carries essential information on the optical absorption and structural characteristics of the irradiated region, enabling the precise real-time monitoring of the laser cleaning process when processed with the proposed machine learning models. For each laser ablation pulse, the corresponding acoustic signal responses were recorded. Simultaneously, the cleaning result was visually assessed and documented with a portable digital microscope. This approach allowed for a detailed analysis of the surface interactions that occurred with each individual ablation pulse. In Fig. 2 (right), the acoustic signals recorded upon irradiation on the same spot for successive pulses are shown. The results are in agreement with the fact that upon ongoing irradiation, the peak-to-peak amplitude is much smaller, and there is a further shift to the right. After Pulse #12, a noticeable drop in amplitude can be observed. D. Data pre-processing We preprocess the recorded signals to guarantee a common starting point in time and a consistent duration. This standardization process ensures that our data representation remains independent of the setup used (such as the distance between the cleaning surface and the sound detector), the thickness of the paint being removed (which also affects the distance from the cleaning point to the sound detector), and the fluence of the cleaning laser pulse. Additionally, events occurring on the surface during the cleaning process are recorded simultaneously across all signals. In other words, among two aligned and trimmed acoustic signals, each time step reflects the same effects on the cleaning surface. Thus, the amplitude of each time step can be used as a descriptor of the cleaning process. We identify the time step corresponding to the global minimum amplitude of each acoustic signal and define this point as the reference point for that signal. The reference point is used to align and trim the duration of each acoustic signal. The duration of the acoustic signal is defined to be 10 µs, consisting of 2 µs before the reference point and 8 µs after it. This duration is sufficient to capture all the characteristics of the acoustic wave generated at a given spot during the cleaning process. By aligning the recorded signals based on their reference point, we observe a loss of information on the time delay of each signal reaching the sound detector. During the cleaning process, each laser pulse induces an acoustic wave at the cleaning spot, removing a portion or layer of the unwanted paint. Consequently, each subsequent laser pulse creates an acoustic wave that propagates deeper within the cleaning spot. This means that the acoustic signal from each later laser pulse takes longer to reach the sound detector, and thus, we observe a time delay in the recorded signals. We define the time shift of each acoustic signal as the time duration between the reference point of each signal and the reference point of the signal of the first laser pulse. This time shift contains valuable information on the amount of paint removed by each laser pulse, allowing for the monitoring of the cleaning process. For this reason, we extract the time shift of each signal and treat it as an additional feature that will be used alongside the 200 time steps (10 µs at 0.05 µs per time step) of each acoustic signal in the machine learning algorithm. In Fig. 2 (right), we present the preprocessing of the acoustic signals of a set of sequential laser pulses at a given cleaning spot. On the left, we present the raw recorded signal of each laser pulse, and on the right, the aligned and trimmed signal with the corresponding extracted time shift. E. Machine Learning We analyze our data using an explainable, tree-based machine learning algorithm, specifically the Random Forest (RF). We chose an RF model due to its ability to handle complex data and perform well even with limited data. Additionally, an RF model can offer insights into how the cleaning process is captured in an acoustic signal. Figure 4 shows a schematic representation of the training process. The aligned and trimmed acoustic signals from the 1,131 cleaned spots were divided into a test set of 120 spots, serving as the hold-out set to evaluate the performance of the trained model, and a training set of 1,011 spots to train it. To eliminate the dependence of the model's performance on a specific training set, we employed a 100-fold cross-validation technique by further splitting the initial training set into 100 random folds, each consisting of 891 training spots and 120 validation spots. Then, we train an RF model on each fold. This method helps the model remain resilient to potential outliers in the data and provides a statistical estimation of its errors. To maintain consistency, we keep all the acoustic signals of an irradiated spot in one set, either the training, validation, or test set. Each RF model is trained to identify whether the next laser pulse is a cleaning pulse by using the aligned and trimmed signal with the corresponding extracted time shift [see Fig. 2 (right)] as input. Notably, the algorithm has no information regarding the fluence of the laser pulse and the thickness of the unwanted paint. Subsequently, we fine-tuned the hyperparameters of the RF, including the number of estimators, i.e., the number of decision trees, the maximum allowed depth of each tree, and the maximum number of samples per leaf of each tree. Based on the model’s performance on the validation set, we found that the optimal RF model consists of 100 estimators and has a maximum depth of 12 with a maximum number of samples per leaf equal to 10. Results Cleaning Pulse Prediction Accuracy [subsection] After the training process, the algorithm was evaluated on the 120 test set spots. The results, including the mean accuracy (the average accuracy of all the trained RF models on the 100 validation sets), the number of errors on both the validation and the test sets, together with the corresponding standard deviation across the 100 folds, are presented in Table 1 . Table 1 Mean accuracy and standard deviation over the RF models for the validation and the test set, and the corresponding number of errors in predicting the cleaning pulse of each spot. Validation Test Prediction accuracy (98.8 ± 1.3)% (97.5 ± 0.8)% Error predictions in 120 spots 2 ± 2 3 ± 1 The mean accuracy and its standard deviation for all RF models on the test set are (97.5 ± 0.8)%, indicating that they perform well on unseen data, regardless of the pulse fluence and the thickness of the unwanted paint. Additionally, the mean accuracy and its standard deviation of (98.8 ± 1.3)% on the validation set further confirm the robustness of the model’s performance. It is important to note that if the algorithm predicts the pulse just before or just after the designated cleaning pulse, it will still be considered a successful prediction. This approach is taken because the goal of this tool is not to replace heritage scientists but to assist them in making better, more informed decisions. In our case, we applied a self-limiting criterion, which eliminates the risk of damaging the marble substrate. However, we successfully classified a single pulse from the group that met our predetermined threshold. Given this achievement, we are optimistic that in a laser cleaning scenario where the self-limiting criteria do not apply, we will still be able to accurately predict the onset of cleaning while safeguarding the substrate. The level of precision required for a clean mock-up, where there is no unwanted paint layer to remove, has not been investigated in terms of data collection from this type of irradiation. It was deemed unnecessary to explore this aspect, as the tool will not be utilized when the surface is already clean. Human supervision will always play a key role in this process. Feature importance [subsection] One of the most significant advantages of the RF algorithm is that it provides insights into which features are most important in the decision-making process. In Fig. 5, we present the feature importance in making decisions as calculated by the Random Forest algorithm. The algorithm demonstrates that the time shift of each pulse with respect to the first pulse at that spot plays the most significant role in identifying the cleaning pulse. In addition, the amplitude of the acoustic signal around the first maximum, the global minimum, and the second maximum is essential for predicting the cleaning pulse. This observation can be understood by considering that the volume fluctuations during the acoustic effect are primarily responsible for the maxima and minima highlighted in Fig. 5. The remainder of the waveform is generally regarded as reflections originating from within (the bulk) or outside the material, arriving later at the detector. Discussion This study aims to enable real-time monitoring of laser cleaning interventions by analyzing laser-induced acoustic signals using machine learning algorithms. The results demonstrate that by processing the generated acoustic waves with random forest algorithms, it is possible to identify the pulse responsible for removing a specific amount of material set at 75% of the irradiated surface. This critical pulse serves as a threshold for terminating the process in a timely manner, preventing any damage to the underlying surface of historical or artistic significance. This research represents the first instance of monitoring laser cleaning interventions in the Heritage field using machine learning algorithms to determine when to stop the irradiation and move the laser beam to an adjacent area. The RF model demonstrated robustness and accuracy in predicting the cleaning pulse. Furthermore, it provides insights into how the cleaning process is captured by the acoustic signal. The time shift emerged as the most significant feature in monitoring the cleaning process. However, this feature cannot be extracted from images. This finding underscores an additional advantage of using acoustic signals for monitoring the cleaning process, as they provide a highly informative feature for this purpose. Declarations Contributions GDB, GJT, and PP conceived the concept and designed the experiments. GDB performed machine learning processing of the recorded data and was a major contributor to writing the manuscript. A-NR performed the repetitive laser irradiation experiments and recorded the acoustic signals. GJT developed the setup and the system for recording and processing the acoustic signals. KM developed the tested mock-ups and contributed to the assessment of the cleaning tests. GPT involved in the analysis. PP coordinated all phases of the study and was a major contributor to the manuscript writing. All authors read and approved the final manuscript. Competing interests All authors declare no financial or non-financial competing interests. Data availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. References Cooper, M. Laser Cleaning in Conservation: An Introduction (Oxford: Butterworth- Heinemann, 1998) Pouli, P. Laser Cleaning on Stonework: Principles, Case Studies, and Future Prospects in: Conserving Stone Heritage (ed. Gherardi, F., Maravelaki, P.N.) 75-100 (Cultural Heritage Science. Springer, Cham., 2022) https://doi.org/10.1007/978-3-030-82942-1_3 Pouli, P. et al. The two-wavelength laser cleaning methodology; theoretical background and examples from its application on CH objects and monuments with emphasis to the Athens Acropolis sculptures. Herit Sci 4, 9 (2016). https://doi.org/10.1186/s40494-016-0077-2. Siano, S. & Salimbeni, R., Advances in Laser Cleaning of Artwork and Objects of Historical Interest: The Optimized Pulse Duration Approach, Accounts of Chemical Research 43 (6), 739-750 (2010) https://doi.org/10.1021/ar900190f Bertasa, M., & Korenberg, K. Successes and challenges in laser cleaning metal artefacts: A review. J. Cultural Heritage 53 , 100-117 (2022), https://doi.org/10.1016/j.culher.2021.10.010 Kono, M., Baldwin, K. G. H., Wain, A. & Rode, A.V., Treating the Untreatable in Art and Heritage Materials: Ultrafast Laser Cleaning of “Cloth-of-Gold, Langmuir , 31 (4), 1596–1604 (2015) https://doi.org/10.1021/la504400h Pouli, P., Selimis, A., Georgiou, S. & Fotakis, C., Recent Studies of Laser Science in Paintings Conservation and Research. Accounts of Chemical Research 43 (6) , 771-781 (2010), DOI: 10.1021/ar900224n Cucci, C, Delaney, J.K., & Picollo, M. Reflectance Hyperspectral Imaging for Investigation of Works of Art: Old Master Paintings and Illuminated Manuscripts, Accounts of Chemical Research 49 (10) , 2070-2079 (2016) DOI: 10.1021/acs.accounts.6b00048 Fontana, R. et al. Application of non-invasive optical monitoring methodologies to follow and record painting cleaning processes. Appl. Phys. A 121 , 957–966 (2015). https://doi.org/10.1007/s00339-015-9505-5 Iwanicka, M., Sylwestrzak, M. & Targowski, P. Optical Coherence Tomography (OCT) for Examination of Artworks in Advanced Characterization Techniques, Diagnostic Tools and Evaluation Methods in Heritage Science (ed. Bastidas, D., Cano, E.) 49–59 (Springer, Cham. 2018) https://doi.org/10.1007/978-3-319-75316-4_4 Crina Anca Sandu, I., de Sá, M. H., & Pereira, M. C. Ancient ‘gilded’ art objects from European cultural heritage: a review on different scales of characterization. Surface and Interface Analysis 43(8), 1134-1151 (2011) DOI 10.1002/sia.3740. Moretti, P. et al. Non-invasive reflection FT-IR spectroscopy for on-site detection of cleaning system residues on polychrome surfaces, Microchem. J. 157, 105033 (2020) https://doi.org/10.1016/j.microc.2020.105033 Iwanicka, M. et al. Complementary use of optical coherence tomography (OCT) and reflection FTIR spectroscopy for in situ non-invasive monitoring of varnish removal from easel paintings. Microchem J . 138, 7–18 (2018), https://doi.org/10.1016/j.microc.2017.12.016 . Melessanaki, K., Stringari, C., Fotakis, C., Anglos, D., Laser Cleaning and Spectroscopy: A Synergistic Approach in the Conservation of a Modern Painting, Laser Chemistry 2006 , 42709 (2006) https://doi.org/10.1155/2006/42709 Striova, J. et al. Optical devices provide unprecedented insights into the laser cleaning of calcium oxalate layers. Microchemical J. 124, 331-337(2015). https://doi.org/10.1016/j.microc.2015.09.005. Moretti, P., et al. Laser cleaning of paintings: in situ optimization of operative parameters through non-invasive assessment by optical coherence tomography (OCT), reflection FT-IR spectroscopy and laser induced fluorescence spectroscopy (LIF). Herit. Sci 7 , 44 (2019). https://doi.org/10.1186/s40494-019-0284-8 Kokkinaki, O., et al. Laser-induced fluorescence as a non-invasive tool to monitor laser-assisted thinning of aged varnish layers on paintings: fundamental issues and critical thresholds. Eur. Phys. J. Plus 136 , 938 (2021). https://doi.org/10.1140/epjp/s13360-021-01929-4 Tserevelakis, G. J. et al. On-line photoacoustic monitoring of laser cleaning on stone: Evaluation of cleaning effectiveness and detection of potential damage to the substrate, J. Cultural Heritage 35, 108-115 (2019), https://doi.org/10.1016/j.culher.2018.05.014 Tserevelakis, G.J., Pouli, P. & Zacharakis, G. Listening to laser light interactions with objects of art: a novel photoacoustic approach for diagnosis and monitoring of laser cleaning interventions. Herit Sci 8, 98 (2020). https://doi.org/10.1186/s40494-020-00440-w Papanikolaou, A. et al. Development of a hybrid photoacoustic and optical monitoring system for the study of laser ablation processes upon the removal of encrustation from stonework, Opto-Electronic Advances 3, 190037(2020), https://doi.org/10.29026/oea.2020.190037 Chen, Y., Deng, GL., Zhou, Q., & Feng, G., Acoustic signal monitoring in laser paint cleaning, Laser Phys. 30 , 066001 (2020) DOI. 10.1088/1555-6611/ab85c7 Dimitroulaki, E. et al. Photoacoustic real-time monitoring of UV laser ablation of aged varnish coatings on Heritage objects, J Cultural Heritage 63, 230-239(2023), https://doi.org/10.1016/j.culher.2023.08.006 Cooper, M.I., Emmony D.C. & Larson, J., Characterization of laser cleaning of limestone, Opt. Laser Technol. 27 (1) , 69-73 (1995), https://doi.org/10.1016/0030-3992(95)93962-Q Lee, J.M. & Watkins, K.G., In-process monitoring techniques for laser cleaning, Opt. Lasers Eng. 34 (4) , 429-442 (2000), https://doi.org/10.1016/S0143-8166(00)00073-7 Bregar, V.B., & Mozina, J., Optoacoustic analysis of the laser-cleaning process, Appl. Surf. Sci. 185 (3-4), 277-288 (2002) , https://doi.org/10.1016/S0169-4332(01)00981-3 Jankowska, M. & Śliwiński, G., Acoustic monitoring for the laser cleaning of sandstone, J. Cultural Heritage 4, 65-71 (2003), https://doi.org/10.1016/S1296-2074(02)01230-X Fiorucci, M. et al., Machine Learning for Cultural Heritage: A Survey. Pattern Recognition Letters 133 , 102 – 108 (2020), DOI. 10.1016/j.patrec.2020.02.017 Argyrou, A. & Agapiou, A. A Review of Artificial Intelligence and Remote Sensing for Archaeological Research. Remote Sens. 14 , 6000 (2022). https://doi.org/10.3390/rs14236000 Zhang, Ch.et al. Application of deep learning algorithms for identifying deterioration in the ushnisha (Head Bun) of the Leshan Giant Buddha. Herit Sci 12 , 399 (2024) https://doi.org/10.1186/s40494-024-01514-9 Kwon, D. & Yu J., Automatic damage detection of stone cultural property based on deep learning algorithm, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences , XLII-2/W15 , 639-643, (2019) https://doi.org/10.5194/isprs-archives-XLII-2-W15-639-2019 Alexakis, E.et al. A novel application of deep learning approach over IRT images for the automated detection of rising damp on historical masonries, Case Studies in Construction Materials 20 , e02889 (2024), https://doi.org/10.1016/j.cscm.2024.e02889 Towarek, A., Halicz, L., Matwin, S., & Wagner, B., Machine learning in analytical chemistry for cultural heritage: A comprehensive review. J. Cul. Her. 70, 64-70 (2024) DOI. 10.1016/j.culher.2024.08.014 Festa, G., et al. Studying ancient Egyptian copper-alloy objects via X-ray diffraction and Machine Learning. J. Cul. Her. 72 , 48-58 (2025) https://doi.org/10.1016/j.culher.2025.01.002 Breiman, L., Random Forests, Machine Learning 45 , 5-32 (2001) https://doi.org/10.1023/A:1010933404324 Ho, T.K. Random decision forests in Proceedings of Third International Conference on Document Analysis and Recognition, vol. 1(ICDAR '95), IEEE Computer Society, USA, 278-282 (1995) Wang, G. et al. Multi-Objective Optimization of Laser Cleaning Quality of Q390 Steel Rust Layer Based on Response Surface Methodology and NSGA-II Algorithm. Materials 17 , 3109 (2024). https://doi.org/10.3390/ma17133109 Fang, C. et al. Effect of Laser Cleaning Parameters on Surface Filth Removal of Porcelain Insulator. Photonics 10(3) , 269 (2023). https://doi.org/10.3390/photonics10030269 Yu, J. & Chen Y., Prediction of simulation parameters of fiber laser cleaning Range hood based on BP neural network. J. Phys.: Conf. Ser. 1820 012118 (2021) https://doi.org/10.1088/1742-6596/1820/1/012118 Chu, D. et al. Real-time acoustic monitoring of laser paint removal based on deep learning, Opt. Express 33 , 1421-1436 (2025) https://doi.org/10.1364/OE.545906 Sun, B. et al. Cleanliness prediction of rusty iron in laser cleaning using convolutional neural networks. Appl. Phys. A 126 , 179 (2020). https://doi.org/10.1007/s00339-020-3363-5 Lin, D., et al., Real-time LIBS monitoring of laser-based layered controlled paint removal from aircraft skin based on random forest, Appl. Opt. 62 , 2569-2576 (2023), https://doi.org/10.1364/AO.484404 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 04 Dec, 2025 Read the published version in npj Heritage Science → Version 1 posted Editorial decision: Revision requested 21 Jul, 2025 Reviews received at journal 21 Jul, 2025 Reviewers agreed at journal 09 Jul, 2025 Reviews received at journal 29 Jun, 2025 Reviewers agreed at journal 27 Jun, 2025 Reviewers invited by journal 27 May, 2025 Editor assigned by journal 25 May, 2025 Submission checks completed at journal 15 May, 2025 First submitted to journal 15 May, 2025 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-6565434","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":462633488,"identity":"9885330d-64a2-4b35-b642-e0a3080c224b","order_by":0,"name":"Georgios D. Barmparis","email":"","orcid":"","institution":"FORTH Institute of Electronic Structure and Laser","correspondingAuthor":false,"prefix":"","firstName":"Georgios","middleName":"D.","lastName":"Barmparis","suffix":""},{"id":462633489,"identity":"ee66797d-170b-4c75-a6a9-cc7a90cdf680","order_by":1,"name":"Anastasios-Nikolaos Raikidis","email":"","orcid":"","institution":"Department of Physics, University of Crete","correspondingAuthor":false,"prefix":"","firstName":"Anastasios-Nikolaos","middleName":"","lastName":"Raikidis","suffix":""},{"id":462633490,"identity":"6cf02155-954b-401c-b08b-ac55416850fd","order_by":2,"name":"Kristalia Melessanaki","email":"","orcid":"","institution":"FORTH Institute of Electronic Structure and Laser","correspondingAuthor":false,"prefix":"","firstName":"Kristalia","middleName":"","lastName":"Melessanaki","suffix":""},{"id":462633491,"identity":"e47921e5-ceda-47db-ac4d-a6f30ab4c03b","order_by":3,"name":"Giorgos P. Tsironis","email":"","orcid":"","institution":"Department of Physics, University of Crete","correspondingAuthor":false,"prefix":"","firstName":"Giorgos","middleName":"P.","lastName":"Tsironis","suffix":""},{"id":462633492,"identity":"467d053a-5080-4d43-b5fc-cbb66e6737dc","order_by":4,"name":"George J. Tserevelakis","email":"","orcid":"","institution":"FORTH Institute of Electronic Structure and Laser","correspondingAuthor":false,"prefix":"","firstName":"George","middleName":"J.","lastName":"Tserevelakis","suffix":""},{"id":462633493,"identity":"ed2f0bbc-76f8-4cb7-839e-1d839889748a","order_by":5,"name":"Paraskevi Pouli","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIie3RsQrCMBCA4Ssduhx0TdGHuEktiH2VlEInB8GlkwQEHV07+BA+QkqhLsVuUhCk4Avo7mCQCrqEugnmHwI3fHBJAEymX62hMQLYAhI1UCfCKVbEElACsI4EcuhOhr1lwfis6rtuJkAm58UQnONVR/xtETNOJ/TSUJFyznyB81RHqJ4OnoRqS9i3FWckMdIu1pIDBpVaLPuCSCQIX8TJtUTdJRpxipDVoZCy5N4uR1tL1Itl9fU+CdxNnjUy4S7t1xf9YuxtkM/TRv3XfJA2p9ESk8lk+rse0AJKOaeb1akAAAAASUVORK5CYII=","orcid":"","institution":"FORTH Institute of Electronic Structure and Laser","correspondingAuthor":true,"prefix":"","firstName":"Paraskevi","middleName":"","lastName":"Pouli","suffix":""}],"badges":[],"createdAt":"2025-04-30 13:38:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6565434/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6565434/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s40494-025-02146-3","type":"published","date":"2025-12-04T15:57:20+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":83619596,"identity":"54f986bf-4f16-4041-96a5-7d4462257ac2","added_by":"auto","created_at":"2025-05-29 14:50:00","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":473985,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic outline of the methodology employed in this study\u003c/p\u003e","description":"","filename":"FIGURE01.png","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/a0b2c03831bfc1bd5ed0a84f.png"},{"id":83619599,"identity":"1aaf9bc0-245d-4cfc-aa3f-70d6fc313165","added_by":"auto","created_at":"2025-05-29 14:50:00","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":639215,"visible":true,"origin":"","legend":"\u003cp\u003eLeft: Schematic outline of the developed mock-ups and the groups of overlayer thickness, laser fluence, and number of laser pulses. Right: The cleaning process of successive laser pulses on a given spot. The recorded and the preprocessed acoustic signals of successive laser pulses at a cleaning spot. From left to right: Each recorded signal is aligned based on the timestep of its global minimum amplitude (the reference point), and its duration is trimmed between 2 μs before and 8 μs after the reference point. The time difference between the reference point of each pulse with respect to the reference point of the first pulse is extracted as the time shift of the acoustic signal. From top to bottom: The signal before cleaning (cyan), the signal of the cleaning laser pulse (red), and the signal after cleaning (green). In blue is the extracted time shift of each acoustic signal.\u003c/p\u003e","description":"","filename":"FIGURE02.png","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/f2800ec16fe3b1f876e15a88.png"},{"id":83619602,"identity":"15b0ecdf-1640-4335-a7ac-1da1b04fba37","added_by":"auto","created_at":"2025-05-29 14:50:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":166390,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of the experimental setup.\u003c/p\u003e","description":"","filename":"FIGURE03.png","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/108d61a69f8b6e59bc706041.png"},{"id":83620568,"identity":"6a1e228f-3bea-4cc5-b792-1b3207ccb169","added_by":"auto","created_at":"2025-05-29 15:06:00","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":269969,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of the 100-fold cross-validation training process of the Random Forest algorithm. After optimizing the hyperparameters of the model, we conclude with a model that features 100 estimators (decision trees), a maximum depth of 12 for each tree, and a minimum number of samples per leaf of 10.\u003c/p\u003e","description":"","filename":"FIGURE04.png","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/13dc4fb089632ceadedb2b18.png"},{"id":83620401,"identity":"238cee36-021c-4b99-b15d-c52e8541173b","added_by":"auto","created_at":"2025-05-29 14:58:00","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":835571,"visible":true,"origin":"","legend":"\u003cp\u003eThe feature importance of the Random Forest algorithm. Left: The processed acoustic signals of the successive pulses on a spot. In cyan, the signal of the pulses before the cleaning pulse, in red, the cleaning pulse, and in green, the signal of the pulses after the cleaning pulse. The intensity of the background color demonstrates the importance of that time step. Right: The time shift is the most important feature, followed by the amplitude of the acoustic signal around its first maximum, its global minimum, and its second maximum.\u003c/p\u003e","description":"","filename":"FIGURE05.png","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/361593a74a11ed6ec10c7961.png"},{"id":97724105,"identity":"47e8c94c-a0ed-41a2-9422-06003c7edab9","added_by":"auto","created_at":"2025-12-08 16:11:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3073061,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6565434/v1/8c81d3bd-6999-4685-b112-1958840adb6c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine learning assisted real-time acoustic monitoring of laser cleaning in Heritage conservation","fulltext":[{"header":"Introduction","content":"\u003cp\u003eOver the past 30 years, laser-assisted removal of unwanted material has become a common cleaning tool in Heritage conservation, often replacing conventional methods based on chemicals and mechanical action. This is due to its unique advantages, including selective and gradual material removal, high precision and control, and its environmentally friendly nature [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. However, this delicate and irreversible process requires careful selection of irradiation parameters and a thorough understanding of ablation mechanisms, especially given the complex nature of deposition layers and the fragile condition of original heritage surfaces.\u003c/p\u003e \u003cp\u003eSelf-limiting processes control the near-infrared (NIR) laser cleaning of black pollution crusts from stonework, ensuring that the cleaning intervention halts immediately after the unwanted crust is removed. Consequently, laser cleaning has become increasingly popular in restoration, as it effectively and safely reveals the original surfaces of cultural heritage assets, such as the Acropolis Sculptures in Athens [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Successful applications of laser cleaning have also been reported for various materials [\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5 CR6\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], including the removal of burial crusts from stone sculptures and leather objects, cleaning soiling from paper, and eliminating corrosion layers from historical metal objects [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Additionally, it has proven effective in removing tarnishing from gilded silver threads [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], among other uses.\u003c/p\u003e \u003cp\u003eLaser cleaning is particularly valuable in addressing a wide range of cleaning challenges. These often involve situations in which self-limiting conditions may not apply, i.e., challenges in which the critical parameters needed to remove unwanted material are very close to, or even exceed, the thresholds that could damage the original substrate [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. One of the most challenging areas of laser cleaning research is the removal of aged or polymerized varnish and overpainting layers from artworks, particularly those that are painted. These layers are composed of polymeric materials that are transparent to near-infrared (NIR) laser radiation. In contrast, most degraded polymeric overlayers highly absorb ultraviolet (UV) laser radiation, making them an effective alternative for layer-by-layer ablation and the controlled removal of unwanted materials. However, the fact that varnish films and paint layers exhibit very similar optical properties, absorbing UV radiation highly, poses limitations to their laser cleaning, as self-limiting conditions do not apply. For this reason, a critical assessment of the cleaning process and real-time monitoring of its progress are crucial.\u003c/p\u003e \u003cp\u003eOver the years, several imaging and spectroscopic techniques have been explored to assess the results and monitor the cleaning process in heritage conservation. Imaging techniques offer a non-destructive and non-invasive approach and are the most widely used. Colorimetry, multispectral reflectography, and optical coherence tomography (OCT) have been employed to assess the color [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and stratigraphy [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] of the objects under study and during restoration. Microscopies, including optical microscopy (OM), scanning electron microscopy (SEM), and atomic force microscopy (AFM), have also been used to investigate the surface morphology of the treated objects [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In parallel, the need for minimal chemical alteration to the underlying original surfaces has influenced several studies that employed spectroscopic techniques, such as Raman spectroscopy and Fourier Transform Infrared Spectroscopy (FTIR) [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], as well as gas chromatography/mass spectrometry (GC-MS), to investigate the chemistry of the cleaned surfaces. Significant research has also been focused on the complementary application of imaging and chemical techniques for a thorough approach to cleaning assessment [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCareful post-treatment appraisal and monitoring during the restoration process are becoming increasingly critical in the context of laser cleaning. While laser cleaning offers precision, it also presents challenges, especially in finding the right balance between effective cleaning and preserving the substrate. In industrial applications, the effectiveness of the cleaning process is evaluated based on two key factors: temporal aspects, such as cleaning time and removal rates, and spatial aspects, which focus on the complete elimination of unwanted materials. However, in heritage conservation, the primary concern is to protect the integrity and chemistry of the original surface.\u003c/p\u003e \u003cp\u003eSignificant research has focused not only on the ideal in-situ evaluation of the results [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] but also on real-time monitoring of the process [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. In contrast to industrial applications, where monitoring typically depends on post-process assessments of cleaning effectiveness and removal rates, laser cleaning of heritage objects requires a different focus. The most critical factor is to identify the key parameters to ensure irradiation stops at the right time, thereby preserving the original surface.\u003c/p\u003e \u003cp\u003eAcoustic monitoring of the laser cleaning process offers unique advantages, particularly because it provides real-time results [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Early studies [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] highlighted its potential, and recent research efforts have increasingly focused on this concept [\u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis method involves recording and processing the acoustic signals generated when light interacts with matter. Previous studies [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] have demonstrated a correlation between the amplitude of the acoustic wave and the amount and composition of removed material. This relationship enables monitoring of the cleaning process progress, as the acoustic signal changes gradually when the overlayer is reduced, followed by an abrupt change upon reaching the substrate. One of the key advantages of this method is that the acoustic signals can be easily and non-invasively detected using non-contact air-coupled ultrasonic transducers, which have a frequency response in the MHz range. These transducers can be placed near the laser-treated surfaces, enabling online and real-time measurements without interrupting the cleaning process for analysis and data recording. The effectiveness of this approach has been tested on a series of technical encrusted marble coupons [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and varnished painting mock-ups [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], which simulate various cleaning challenges. The relevant results have been cross-checked using other analytical techniques. The proposed acoustic monitoring strategy enables the identification of the critical laser pulse responsible for complete encrustation removal. This capability can assist conservators and restorers (C-Rs) in monitoring laser treatments and making informed decisions regarding the progress of the crust or overlayer removal process.\u003c/p\u003e \u003cp\u003eArtificial Intelligence and Machine Learning (ML) algorithms are increasingly integrated into scientific research and industrial applications, offering powerful tools for data analysis, process optimization, and decision-making. ML models can identify patterns and correlations within experimental data, and they have been employed to various applications in cultural heritage research and conservation [\u003cspan additionalcitationids=\"CR28 CR29 CR30 CR31 CR32\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], such as to identify cracks [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] and other deterioration features [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], to identify rising damp on monuments promptly [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] or to study certain type of objects [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] to name a few.\u003c/p\u003e \u003cp\u003eHerein, we discuss the implementation of ML in monitoring the laser cleaning process and ensuring a safe intervention on objects of heritage value. Among the available ML algorithms, we choose the Random Forest algorithm (RF), introduced by Breiman in 2001 [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e], as an extension to the \u0026ldquo;the random subspace\u0026rdquo; method by Tim Kam Ho [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], which does a random selection of a subset of features to grow each decision tree. RF is an explainable tree-based ensemble learning method that has emerged as one of the most potent and versatile machine learning algorithms for classification and regression tasks. Among its advantages, an RF algorithm can handle high-dimensional datasets, model complex nonlinear relationships, perform well with limited amounts of data, and provide insights through feature importance metrics.\u003c/p\u003e \u003cp\u003ePrevious works have applied ML to optimize laser parameters in industrial cleaning applications [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e] or to monitor and evaluate the cleaning results [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. However, in Heritage science and conservation practice, the focus shifts away from material removal efficiency, as seen in industrial applications. The primary goal in heritage conservation is to determine the precise moment to stop the cleaning process, thereby preventing damage to the underlying historical surface that must be preserved and revealed. This approach differentiates conservation work from industrial processes and is consistently followed in this context.\u003c/p\u003e \u003cp\u003eThis study presents the first feasibility study to monitor the restoration process of sensitive cultural heritage objects in real-time. It combines acoustic wave signals produced during laser cleaning and explainable ML algorithms to identify the pulse that completely removes the unwanted paint without damaging the substrate.\u003c/p\u003e \u003cp\u003eThe methodology, including details on individual irradiation and monitoring parameters, as well as processing procedures and machine learning algorithms used in this study, is outlined in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e \u003cb\u003eTechnical mock-ups\u003c/b\u003e \u003cb\u003e[subsection]\u003c/b\u003e\u003c/p\u003e \u003cp\u003eA series of technical mock-ups were created for data-gathering purposes [Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (left)]. The base was made from thin slabs of white marble sourced from the Greek island of Thasos in the Northern Aegean Sea. Black acrylic paint (Motip Matt Black Acrylic Varnish) was applied to the marble from approximately 30 cm. The mock-ups were then left to dry naturally for at least one week before conducting irradiation experiments. This black acrylic overlayer, intended to simulate the \"unwanted crust,\" was applied in multiple layers, resulting in varying thicknesses ranging from 8 to 55 \u0026micro;m. The mock-up design facilitated simple laser ablation processes characterized by self-limiting conditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCleaning methodology and criteria for quality assessment\u003c/b\u003e \u003cb\u003e[subsection]\u003c/b\u003e\u003c/p\u003e \u003cp\u003eLaser cleaning is affected by several parameters that are directly related to the materials involved [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The most critical parameters are the appropriate \u003cem\u003ewavelength\u003c/em\u003e, λ, and \u003cem\u003epulse duration\u003c/em\u003e, τ\u003csub\u003ep\u003c/sub\u003e, of the laser system, the \u003cem\u003elaser fluence\u003c/em\u003e, F, and the \u003cem\u003enumber of laser pulses applied\u003c/em\u003e, N. The wavelength and the pulse duration ensure the laser ablation mechanism is well-suited to the specific material. F is the energy, E, per unit area, S, of the irradiated surface, F\u0026thinsp;=\u0026thinsp;E/S (J/cm2). In addition to the laser fluence, we need to define the \u003cem\u003elaser fluence threshold\u003c/em\u003e, F\u003csub\u003ethr\u003c/sub\u003e, i.e., the minimum fluence required to remove the unwanted material, and the \u003cem\u003ecritical laser fluence\u003c/em\u003e, F\u003csub\u003edamage\u003c/sub\u003e, i.e., the fluence above which the laser pulse causes damage to the underlying substrate. Finally, N refers to the total number of sequential laser pulses required to remove the unwanted material, which is closely related to the thickness of the unwanted layer or crust and the selected laser fluence.\u003c/p\u003e \u003cp\u003eIn this study, a QS Nd:YAG laser system emitting at λ\u0026thinsp;=\u0026thinsp;1064 nm, with pulses of τ\u003csub\u003ep\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;10 ns, was employed. The cleaning efficiency of the given technical mock-ups has been well studied and characterized [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The laser fluence threshold was found to be F\u003csub\u003ethr\u003c/sub\u003e = 0.1 J/cm\u003csup\u003e2\u003c/sup\u003e, and the critical laser fluence, F\u003csub\u003edamage\u003c/sub\u003e = 2.1 J/cm\u003csup\u003e2\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDuring the data collection process, several values of F were carefully chosen between the F\u003csub\u003ethr\u003c/sub\u003e and the F\u003csub\u003edamage\u003c/sub\u003e thresholds to investigate the cleaning process at different paces. Laser pulses with low laser fluence slowly remove the unwanted material; thus, a high number of laser pulses is required to clean the surface, resulting in a slow pace. In contrast, laser pulses with high laser fluence transfer more energy to the irradiated surface, thus requiring fewer laser pulses to clean it, resulting in a high pace. Another parameter that affects the pace of the cleaning process is the thickness, \u003cem\u003ed\u003c/em\u003e, of the unwanted material. The cleaning of materials with different thicknesses requires varying numbers of laser pulses at the same laser fluence.\u003c/p\u003e \u003cp\u003eIn this experiment, mock-ups characterized by a thickness, \u003cem\u003ed\u003c/em\u003e\u003csub\u003e\u003cem\u003ei\u003c/em\u003e\u003c/sub\u003e, of black crust are irradiated with a given laser fluence, \u003cem\u003eF\u003c/em\u003e\u003csub\u003e\u003cem\u003ei\u003c/em\u003e\u003c/sub\u003e, and a given number of applied laser pulses, \u003cem\u003eN\u003c/em\u003e\u003csub\u003e\u003cem\u003ei\u003c/em\u003e\u003c/sub\u003e. With this method, we collect data (acoustic signals) from groups of approximately 150 identical irradiated spots, all with the same overlayer thickness, \u003cem\u003edi\u003c/em\u003e, laser fluence, \u003cem\u003eFi\u003c/em\u003e, and number of laser pulses, \u003cem\u003eN\u003c/em\u003e\u003csub\u003e\u003cem\u003ei\u003c/em\u003e\u003c/sub\u003e. In Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, we present a schematic outline of the developed mock-ups and the various groups considered in this work, including thickness, laser fluence, and the number of laser pulses.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSuccessive laser pulses lead to increasingly removing the unwanted material from the surface of the mock-ups. At this point, it is essential to highlight that an uncontrolled cleaning process can damage the underlying surface in the context of laser cleaning in Heritage conservation. To avoid damaging the substrate, we introduce the definition of the \u003cem\u003ecleaning pulse\u003c/em\u003e, which is the pulse that sufficiently cleans the surface without damaging the substrate and is defined as the pulse that removes at least 75% of the irradiated area. We determine the ratio R between the \u003cem\u003ecleaned area\u003c/em\u003e, S\u003csub\u003e\u003cem\u003ecleaned\u003c/em\u003e\u003c/sub\u003e, and the \u003cem\u003eirradiated area\u003c/em\u003e, S\u003csub\u003e\u003cem\u003eirradiated\u003c/em\u003e\u003c/sub\u003e, by processing the collected images with the ImageJ software. In our experiments, we irradiate a few additional laser pulses after identifying the cleaning pulse to investigate the material's response beyond effective cleaning. In Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, we present the digital microscope images taken after each of the laser pulses at a given spot with a paint thickness of d\u0026thinsp;=\u0026thinsp;33 \u0026micro;m, irradiated with a laser fluence of F\u0026thinsp;=\u0026thinsp;0.7 J/cm\u0026sup2;. The white area (exposed marble) is the cleaned area, while the black area is the total irradiated area. Pulse #11 cleans 53.9%, and Pulse #12 cleans 89.5% of the irradiated area. In this case, Pulse #12 is identified as the cleaning pulse.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eExperimental Setup\u003c/b\u003e \u003cb\u003e[subsection]\u003c/b\u003e\u003c/p\u003e \u003cp\u003eA schematic representation of the experimental setup is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The setup combines laser ablation and acoustic recording modalities. In all experiments, the relative position of the irradiated spot to the laser focusing lens is 32 cm. The acoustic piezoelectric transducer is located approximately 6 cm above and to the right of the spot at a 45-degree angle.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eA. Laser Cleaning System\u003c/h2\u003e \u003cp\u003eΑ Q-Switched Nd:YAG laser (LITRON Lasers, TRLi 850 Series, Rugby, Warwickshire, England) was employed for performing the laser irradiation treatments. The laser system operated at the fundamental wavelength of 1064 nm, emitting pulses of 10 ns and with a variable repetition rate ranging from 0.25 to 1 Hz. The energy fluence values on the mock-ups ranged from 0.2 J/cm\u0026sup2; to 1.2 J/cm\u0026sup2; and were estimated by measuring the spot size (~\u0026thinsp;0.26 cm\u0026sup2;) of the focused beam on black photographic paper. All the irradiation experiments were performed in dry conditions.\u003c/p\u003e \u003cp\u003eOptical imaging of the ablated regions was performed by utilizing a portable digital microscope (Dino-Lite Edge AM4113TFV2W) with a magnification in the range of \u0026times;50 to \u0026times;200. The depth of the laser-induced craters and the thickness of the over-layers were measured by means of a Portable Surface Roughness Tester profilometer (Mitutoyo America Corporation, Surftest SJ-410 Series, Aurora, IL, USA).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eB. Acoustic monitoring system\u003c/h3\u003e\n\u003cp\u003eThe laser-induced acoustic response on the examined mock-ups was detected by an air-coupled transducer (NCT1-D7-P10, The Ultran Group, State College, PA, USA; nominal central frequency: 1 MHz; focal distance: 10 mm; numerical aperture: 0.31) placed at approximately 45 degrees in respect to the horizontal plane and in an out-of-focus position around 4 cm away from the irradiation region to avoid signal saturation effects. The signals were subsequently enhanced by two radio frequency amplifiers (TB-414-8A+, Mini-Circuits, Camberley, England; gain:31 dB) connected in series prior to their digitization by an oscilloscope (DSO7034A, Agilent Technologies, Santa Clara, CA, USA; bandwidth: 350 MHz; sample rate: up to 2 GSamples/s) which, in turn, was connected to a laptop computer equipped with custom-made software controlling the measurement procedures. The recorded waveforms were sampled at 1000 points over a 50 \u0026micro;s temporal window (corresponding to a 20 MSamples/s sampling rate) and bandpassed between 0.5 and 2 MHz to reduce mainly high-frequency noise before being saved to the computer as ASCII files. Recording synchronization was achieved through the trigger output of the laser source, which was connected to the oscilloscope\u0026rsquo;s second channel.\u003c/p\u003e\n\u003ch3\u003eC. Monitoring and Data Gathering\u003c/h3\u003e\n\u003cp\u003eDuring data collection, a total of 1,131 spots were irradiated on the studied mock-ups of black paint overlayers of various thicknesses. Each spot was exposed to varying sets of laser fluence, F\u003csub\u003ei\u003c/sub\u003e, and number of laser pulses, N\u003csub\u003ei\u003c/sub\u003e. The recorded waveform corresponding to each laser pulse incidence represents the detected acoustic pressure amplitude as a function of time, where positive and negative values can be interpreted as the compression and rarefaction regions of the propagating ultrasonic wave, respectively. The peak-to-peak amplitude, quantified as the difference between the maximum and minimum pressure values, has been demonstrated [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] to be directly proportional to the effective optical absorption coefficient at the irradiation wavelength of 1064 nm under typical energy fluence conditions. This relationship arises due to the localized thermoelastic expansion and subsequent generation of broadband ultrasonic waves following transient optical energy deposition. Therefore, the peak-to-peak amplitude parameter, along with the acoustic perturbation\u0026rsquo;s time-of-flight, carries essential information on the optical absorption and structural characteristics of the irradiated region, enabling the precise real-time monitoring of the laser cleaning process when processed with the proposed machine learning models. For each laser ablation pulse, the corresponding acoustic signal responses were recorded. Simultaneously, the cleaning result was visually assessed and documented with a portable digital microscope. This approach allowed for a detailed analysis of the surface interactions that occurred with each individual ablation pulse.\u003c/p\u003e \u003cp\u003eIn Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e (right), the acoustic signals recorded upon irradiation on the same spot for successive pulses are shown. The results are in agreement with the fact that upon ongoing irradiation, the peak-to-peak amplitude is much smaller, and there is a further shift to the right. After Pulse #12, a noticeable drop in amplitude can be observed.\u003c/p\u003e\n\u003ch3\u003eD. Data pre-processing\u003c/h3\u003e\n\u003cp\u003eWe preprocess the recorded signals to guarantee a common starting point in time and a consistent duration. This standardization process ensures that our data representation remains independent of the setup used (such as the distance between the cleaning surface and the sound detector), the thickness of the paint being removed (which also affects the distance from the cleaning point to the sound detector), and the fluence of the cleaning laser pulse. Additionally, events occurring on the surface during the cleaning process are recorded simultaneously across all signals. In other words, among two aligned and trimmed acoustic signals, each time step reflects the same effects on the cleaning surface. Thus, the amplitude of each time step can be used as a descriptor of the cleaning process.\u003c/p\u003e \u003cp\u003eWe identify the time step corresponding to the global minimum amplitude of each acoustic signal and define this point as the reference point for that signal. The reference point is used to align and trim the duration of each acoustic signal. The duration of the acoustic signal is defined to be 10 \u0026micro;s, consisting of 2 \u0026micro;s before the reference point and 8 \u0026micro;s after it. This duration is sufficient to capture all the characteristics of the acoustic wave generated at a given spot during the cleaning process. By aligning the recorded signals based on their reference point, we observe a loss of information on the time delay of each signal reaching the sound detector. During the cleaning process, each laser pulse induces an acoustic wave at the cleaning spot, removing a portion or layer of the unwanted paint. Consequently, each subsequent laser pulse creates an acoustic wave that propagates deeper within the cleaning spot. This means that the acoustic signal from each later laser pulse takes longer to reach the sound detector, and thus, we observe a time delay in the recorded signals. We define the time shift of each acoustic signal as the time duration between the reference point of each signal and the reference point of the signal of the first laser pulse. This time shift contains valuable information on the amount of paint removed by each laser pulse, allowing for the monitoring of the cleaning process. For this reason, we extract the time shift of each signal and treat it as an additional feature that will be used alongside the 200 time steps (10 \u0026micro;s at 0.05 \u0026micro;s per time step) of each acoustic signal in the machine learning algorithm.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (right), we present the preprocessing of the acoustic signals of a set of sequential laser pulses at a given cleaning spot. On the left, we present the raw recorded signal of each laser pulse, and on the right, the aligned and trimmed signal with the corresponding extracted time shift.\u003c/p\u003e\n\u003ch3\u003eE. Machine Learning\u003c/h3\u003e\n\u003cp\u003eWe analyze our data using an explainable, tree-based machine learning algorithm, specifically the Random Forest (RF). We chose an RF model due to its ability to handle complex data and perform well even with limited data. Additionally, an RF model can offer insights into how the cleaning process is captured in an acoustic signal. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows a schematic representation of the training process.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe aligned and trimmed acoustic signals from the 1,131 cleaned spots were divided into a test set of 120 spots, serving as the hold-out set to evaluate the performance of the trained model, and a training set of 1,011 spots to train it. To eliminate the dependence of the model's performance on a specific training set, we employed a 100-fold cross-validation technique by further splitting the initial training set into 100 random folds, each consisting of 891 training spots and 120 validation spots. Then, we train an RF model on each fold. This method helps the model remain resilient to potential outliers in the data and provides a statistical estimation of its errors. To maintain consistency, we keep all the acoustic signals of an irradiated spot in one set, either the training, validation, or test set. Each RF model is trained to identify whether the next laser pulse is a cleaning pulse by using the aligned and trimmed signal with the corresponding extracted time shift [see Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (right)] as input. Notably, the algorithm has no information regarding the fluence of the laser pulse and the thickness of the unwanted paint.\u003c/p\u003e \u003cp\u003eSubsequently, we fine-tuned the hyperparameters of the RF, including the number of estimators, i.e., the number of decision trees, the maximum allowed depth of each tree, and the maximum number of samples per leaf of each tree. Based on the model\u0026rsquo;s performance on the validation set, we found that the optimal RF model consists of 100 estimators and has a maximum depth of 12 with a maximum number of samples per leaf equal to 10.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e \u003cb\u003eCleaning Pulse Prediction Accuracy\u003c/b\u003e \u003cb\u003e[subsection]\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAfter the training process, the algorithm was evaluated on the 120 test set spots. The results, including the mean accuracy (the average accuracy of all the trained RF models on the 100 validation sets), the number of errors on both the validation and the test sets, together with the corresponding standard deviation across the 100 folds, are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\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\u003eMean accuracy and standard deviation over the RF models for the validation and the test set, and the corresponding number of errors in predicting the cleaning pulse of each spot.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValidation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrediction accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e(98.8\u0026thinsp;\u0026plusmn;\u0026thinsp;1.3)%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e(97.5\u0026thinsp;\u0026plusmn;\u0026thinsp;0.8)%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eError predictions in 120 spots\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e2\u0026thinsp;\u0026plusmn;\u0026thinsp;2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e3\u0026thinsp;\u0026plusmn;\u0026thinsp;1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe mean accuracy and its standard deviation for all RF models on the test set are (97.5\u0026thinsp;\u0026plusmn;\u0026thinsp;0.8)%, indicating that they perform well on unseen data, regardless of the pulse fluence and the thickness of the unwanted paint. Additionally, the mean accuracy and its standard deviation of (98.8\u0026thinsp;\u0026plusmn;\u0026thinsp;1.3)% on the validation set further confirm the robustness of the model\u0026rsquo;s performance.\u003c/p\u003e \u003cp\u003eIt is important to note that if the algorithm predicts the pulse just before or just after the designated cleaning pulse, it will still be considered a successful prediction. This approach is taken because the goal of this tool is not to replace heritage scientists but to assist them in making better, more informed decisions.\u003c/p\u003e \u003cp\u003eIn our case, we applied a self-limiting criterion, which eliminates the risk of damaging the marble substrate. However, we successfully classified a single pulse from the group that met our predetermined threshold. Given this achievement, we are optimistic that in a laser cleaning scenario where the self-limiting criteria do not apply, we will still be able to accurately predict the onset of cleaning while safeguarding the substrate.\u003c/p\u003e \u003cp\u003eThe level of precision required for a clean mock-up, where there is no unwanted paint layer to remove, has not been investigated in terms of data collection from this type of irradiation. It was deemed unnecessary to explore this aspect, as the tool will not be utilized when the surface is already clean. Human supervision will always play a key role in this process.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature importance\u003c/b\u003e \u003cb\u003e[subsection]\u003c/b\u003e\u003c/p\u003e \u003cp\u003eOne of the most significant advantages of the RF algorithm is that it provides insights into which features are most important in the decision-making process.\u003c/p\u003e\u003cp\u003eIn Fig. 5, we present the feature importance in making decisions as calculated by the Random Forest algorithm. The algorithm demonstrates that the time shift of each pulse with respect to the first pulse at that spot plays the most significant role in identifying the cleaning pulse. In addition, the amplitude of the acoustic signal around the first maximum, the global minimum, and the second maximum is essential for predicting the cleaning pulse. This observation can be understood by considering that the volume fluctuations during the acoustic effect are primarily responsible for the maxima and minima highlighted in Fig. 5. The remainder of the waveform is generally regarded as reflections originating from within (the bulk) or outside the material, arriving later at the detector.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study aims to enable real-time monitoring of laser cleaning interventions by analyzing laser-induced acoustic signals using machine learning algorithms. The results demonstrate that by processing the generated acoustic waves with random forest algorithms, it is possible to identify the pulse responsible for removing a specific amount of material set at 75% of the irradiated surface. This critical pulse serves as a threshold for terminating the process in a timely manner, preventing any damage to the underlying surface of historical or artistic significance. This research represents the first instance of monitoring laser cleaning interventions in the Heritage field using machine learning algorithms to determine when to stop the irradiation and move the laser beam to an adjacent area.\u003c/p\u003e \u003cp\u003eThe RF model demonstrated robustness and accuracy in predicting the cleaning pulse. Furthermore, it provides insights into how the cleaning process is captured by the acoustic signal. The time shift emerged as the most significant feature in monitoring the cleaning process. However, this feature cannot be extracted from images. This finding underscores an additional advantage of using acoustic signals for monitoring the cleaning process, as they provide a highly informative feature for this purpose.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eContributions\u003c/h2\u003e\n\u003cp\u003eGDB, GJT, and PP conceived the concept and designed the experiments. GDB performed machine learning processing of the recorded data and was a major contributor to writing the manuscript. A-NR performed the repetitive laser irradiation experiments and recorded the acoustic signals. GJT developed the setup and the system for recording and processing the acoustic signals. KM developed the tested mock-ups and contributed to the assessment of the cleaning tests. GPT involved in the analysis. PP coordinated all phases of the study and was a major contributor to the manuscript writing. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eAll authors declare no financial or non-financial competing interests.\u003c/p\u003e\n\u003ch2\u003eData availability\u003c/h2\u003e\n\u003cp\u003eThe datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCooper, M. \u003cem\u003eLaser Cleaning in Conservation: An Introduction\u003c/em\u003e (Oxford: Butterworth- Heinemann, 1998)\u003c/li\u003e\n\u003cli\u003ePouli, P. Laser Cleaning on Stonework: Principles, Case Studies, and Future Prospects in: \u003cem\u003eConserving Stone Heritage\u003c/em\u003e (ed. Gherardi, F., Maravelaki, P.N.) 75-100 (Cultural Heritage Science. Springer, Cham., 2022) https://doi.org/10.1007/978-3-030-82942-1_3 \u003c/li\u003e\n\u003cli\u003ePouli, P. et al. The two-wavelength laser cleaning methodology; theoretical background and examples from its application on CH objects and monuments with emphasis to the Athens Acropolis sculptures. \u003cem\u003eHerit Sci\u003c/em\u003e\u003cstrong\u003e4,\u003c/strong\u003e 9 (2016). https://doi.org/10.1186/s40494-016-0077-2. \u003c/li\u003e\n\u003cli\u003eSiano, S. \u0026amp; Salimbeni, R., Advances in Laser Cleaning of Artwork and Objects of Historical Interest: The Optimized Pulse Duration Approach, \u003cem\u003eAccounts of Chemical Research\u003c/em\u003e\u003cstrong\u003e43 (6),\u003c/strong\u003e 739-750 (2010) https://doi.org/10.1021/ar900190f\u003c/li\u003e\n\u003cli\u003eBertasa, M., \u0026amp; Korenberg, K. Successes and challenges in laser cleaning metal artefacts: A review. \u003cem\u003eJ. Cultural Heritage\u003c/em\u003e\u003cstrong\u003e53\u003c/strong\u003e, 100-117 (2022), https://doi.org/10.1016/j.culher.2021.10.010 \u003c/li\u003e\n\u003cli\u003eKono, M., Baldwin, K. G. H., Wain, A. \u0026amp; Rode, A.V., Treating the Untreatable in Art and Heritage Materials: Ultrafast Laser Cleaning of \u0026ldquo;Cloth-of-Gold, \u003cem\u003eLangmuir\u003c/em\u003e, \u003cstrong\u003e31 \u003c/strong\u003e(4), 1596\u0026ndash;1604 (2015) https://doi.org/10.1021/la504400h\u003c/li\u003e\n\u003cli\u003ePouli, P., Selimis, A., Georgiou, S. \u0026amp; Fotakis, C., Recent Studies of Laser Science in Paintings Conservation and Research. \u003cem\u003eAccounts of Chemical Research\u003c/em\u003e\u003cstrong\u003e43 (6)\u003c/strong\u003e, 771-781 (2010), DOI: 10.1021/ar900224n \u003c/li\u003e\n\u003cli\u003eCucci, C, Delaney, J.K., \u0026amp; Picollo, M. Reflectance Hyperspectral Imaging for Investigation of Works of Art: Old Master Paintings and Illuminated Manuscripts, \u003cem\u003eAccounts of Chemical Research\u003c/em\u003e\u003cstrong\u003e49 (10)\u003c/strong\u003e, 2070-2079 (2016) DOI: 10.1021/acs.accounts.6b00048\u003c/li\u003e\n\u003cli\u003eFontana, R. et al. Application of non-invasive optical monitoring methodologies to follow and record painting cleaning processes. \u003cem\u003eAppl. Phys. A\u003c/em\u003e\u003cstrong\u003e121\u003c/strong\u003e, 957\u0026ndash;966 (2015). https://doi.org/10.1007/s00339-015-9505-5\u003c/li\u003e\n\u003cli\u003eIwanicka, M., Sylwestrzak, M. \u0026amp; Targowski, P. Optical Coherence Tomography (OCT) for Examination of Artworks in \u003cem\u003eAdvanced Characterization Techniques, Diagnostic Tools and Evaluation Methods in Heritage Science\u003c/em\u003e (ed. Bastidas, D., Cano, E.) 49\u0026ndash;59 (Springer, Cham. 2018) https://doi.org/10.1007/978-3-319-75316-4_4\u003c/li\u003e\n\u003cli\u003eCrina Anca Sandu, I., de S\u0026aacute;, M. H., \u0026amp; Pereira, M. C. Ancient \u0026lsquo;gilded\u0026rsquo; art objects from European cultural heritage: a review on different scales of characterization. \u003cem\u003eSurface and Interface Analysis\u003c/em\u003e\u003cstrong\u003e43(8),\u003c/strong\u003e 1134-1151 (2011) DOI 10.1002/sia.3740.\u003c/li\u003e\n\u003cli\u003eMoretti, P. et al. Non-invasive reflection FT-IR spectroscopy for on-site detection of cleaning system residues on polychrome surfaces, \u003cem\u003eMicrochem. J.\u003c/em\u003e\u003cstrong\u003e157,\u003c/strong\u003e 105033 (2020) https://doi.org/10.1016/j.microc.2020.105033 \u003c/li\u003e\n\u003cli\u003eIwanicka, M. et al. Complementary use of optical coherence tomography (OCT) and reflection FTIR spectroscopy for in situ non-invasive monitoring of varnish removal from easel paintings. \u003cem\u003eMicrochem J\u003c/em\u003e. \u003cstrong\u003e138,\u003c/strong\u003e 7\u0026ndash;18 (2018), https://doi.org/10.1016/j.microc.2017.12.016 .\u003c/li\u003e\n\u003cli\u003eMelessanaki, K., Stringari, C., Fotakis, C., Anglos, D., Laser Cleaning and Spectroscopy: A Synergistic Approach in the Conservation of a Modern Painting, \u003cem\u003eLaser Chemistry\u003c/em\u003e\u003cstrong\u003e2006\u003c/strong\u003e, 42709 (2006) \u003cu\u003ehttps://doi.org/10.1155/2006/42709\u003c/u\u003e\u003c/li\u003e\n\u003cli\u003eStriova, J. et al. Optical devices provide unprecedented insights into the laser cleaning of calcium oxalate layers. \u003cem\u003eMicrochemical J.\u003c/em\u003e\u003cstrong\u003e124,\u003c/strong\u003e 331-337(2015). https://doi.org/10.1016/j.microc.2015.09.005.\u003c/li\u003e\n\u003cli\u003eMoretti, P., et al. Laser cleaning of paintings: in situ optimization of operative parameters through non-invasive assessment by optical coherence tomography (OCT), reflection FT-IR spectroscopy and laser induced fluorescence spectroscopy (LIF). \u003cem\u003eHerit. Sci\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 44 (2019). https://doi.org/10.1186/s40494-019-0284-8\u003c/li\u003e\n\u003cli\u003eKokkinaki, O., et al. Laser-induced fluorescence as a non-invasive tool to monitor laser-assisted thinning of aged varnish layers on paintings: fundamental issues and critical thresholds. \u003cem\u003eEur. Phys. J. Plus\u003c/em\u003e\u003cstrong\u003e136\u003c/strong\u003e, 938 (2021). https://doi.org/10.1140/epjp/s13360-021-01929-4 \u003c/li\u003e\n\u003cli\u003eTserevelakis, G. J. et al. On-line photoacoustic monitoring of laser cleaning on stone: Evaluation of cleaning effectiveness and detection of potential damage to the substrate, \u003cem\u003eJ. Cultural Heritage\u003c/em\u003e\u003cstrong\u003e35,\u003c/strong\u003e 108-115 (2019), https://doi.org/10.1016/j.culher.2018.05.014\u003c/li\u003e\n\u003cli\u003eTserevelakis, G.J., Pouli, P. \u0026amp; Zacharakis, G. Listening to laser light interactions with objects of art: a novel photoacoustic approach for diagnosis and monitoring of laser cleaning interventions. \u003cem\u003eHerit Sci\u003c/em\u003e\u003cstrong\u003e8,\u003c/strong\u003e 98 (2020). https://doi.org/10.1186/s40494-020-00440-w \u003c/li\u003e\n\u003cli\u003ePapanikolaou, A. et al. Development of a hybrid photoacoustic and optical monitoring system for the study of laser ablation processes upon the removal of encrustation from stonework, \u003cem\u003eOpto-Electronic Advances\u003c/em\u003e\u003cstrong\u003e3,\u003c/strong\u003e 190037(2020), https://doi.org/10.29026/oea.2020.190037\u003c/li\u003e\n\u003cli\u003eChen, Y., Deng, GL., Zhou, Q., \u0026amp; Feng, G., Acoustic signal monitoring in laser paint cleaning, \u003cem\u003eLaser Phys. \u003c/em\u003e\u003cstrong\u003e30\u003c/strong\u003e, 066001 (2020) DOI. 10.1088/1555-6611/ab85c7\u003c/li\u003e\n\u003cli\u003eDimitroulaki, E. et al. Photoacoustic real-time monitoring of UV laser ablation of aged varnish coatings on Heritage objects, \u003cem\u003eJ Cultural Heritage \u003c/em\u003e\u003cstrong\u003e63,\u003c/strong\u003e 230-239(2023), https://doi.org/10.1016/j.culher.2023.08.006\u003c/li\u003e\n\u003cli\u003eCooper, M.I., Emmony D.C. \u0026amp; Larson, J., Characterization of laser cleaning of limestone, \u003cem\u003eOpt. Laser Technol.\u003c/em\u003e\u003cstrong\u003e 27 (1)\u003c/strong\u003e, 69-73 (1995), https://doi.org/10.1016/0030-3992(95)93962-Q\u003c/li\u003e\n\u003cli\u003eLee, J.M. \u0026amp; Watkins, K.G., In-process monitoring techniques for laser cleaning, \u003cem\u003eOpt. Lasers Eng.\u003c/em\u003e\u003cstrong\u003e34 (4)\u003c/strong\u003e, 429-442 (2000), https://doi.org/10.1016/S0143-8166(00)00073-7\u003c/li\u003e\n\u003cli\u003eBregar, V.B., \u0026amp; Mozina, J., Optoacoustic analysis of the laser-cleaning process, \u003cem\u003eAppl. Surf. Sci. \u003c/em\u003e\u003cstrong\u003e185 (3-4),\u003c/strong\u003e 277-288 (2002)\u003cstrong\u003e, \u003c/strong\u003ehttps://doi.org/10.1016/S0169-4332(01)00981-3\u003c/li\u003e\n\u003cli\u003eJankowska, M. \u0026amp; Śliwiński, G., Acoustic monitoring for the laser cleaning of sandstone, \u003cem\u003eJ. Cultural Heritage \u003c/em\u003e\u003cstrong\u003e4,\u003c/strong\u003e 65-71 (2003), https://doi.org/10.1016/S1296-2074(02)01230-X\u003c/li\u003e\n\u003cli\u003eFiorucci, M. et al., Machine Learning for Cultural Heritage: A Survey. \u003cem\u003ePattern Recognition Letters\u003c/em\u003e\u003cstrong\u003e133\u003c/strong\u003e, 102 \u0026ndash; 108 (2020), DOI. 10.1016/j.patrec.2020.02.017\u003c/li\u003e\n\u003cli\u003eArgyrou, A. \u0026amp; Agapiou, A. A Review of Artificial Intelligence and Remote Sensing for Archaeological Research. \u003cem\u003eRemote Sens.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 6000 (2022). https://doi.org/10.3390/rs14236000\u003c/li\u003e\n\u003cli\u003eZhang, Ch.et al. Application of deep learning algorithms for identifying deterioration in the ushnisha (Head Bun) of the Leshan Giant Buddha. \u003cem\u003eHerit Sci\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 399 (2024) https://doi.org/10.1186/s40494-024-01514-9\u003c/li\u003e\n\u003cli\u003eKwon, D. \u0026amp; Yu J., Automatic damage detection of stone cultural property based on deep learning algorithm, \u003cem\u003eThe International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences\u003c/em\u003e, \u003cstrong\u003eXLII-2/W15\u003c/strong\u003e, 639-643, (2019) https://doi.org/10.5194/isprs-archives-XLII-2-W15-639-2019\u003c/li\u003e\n\u003cli\u003eAlexakis, E.et al. A novel application of deep learning approach over IRT images for the automated detection of rising damp on historical masonries, \u003cem\u003eCase Studies in Construction Materials\u003c/em\u003e\u003cstrong\u003e20\u003c/strong\u003e, e02889 (2024), https://doi.org/10.1016/j.cscm.2024.e02889\u003c/li\u003e\n\u003cli\u003eTowarek, A., Halicz, L., Matwin, S., \u0026amp; Wagner, B., Machine learning in analytical chemistry for cultural heritage: A comprehensive review. \u003cem\u003eJ. Cul. Her. \u003c/em\u003e\u003cstrong\u003e70,\u003c/strong\u003e 64-70 (2024) DOI. 10.1016/j.culher.2024.08.014\u003c/li\u003e\n\u003cli\u003eFesta, G., et al. Studying ancient Egyptian copper-alloy objects via X-ray diffraction and Machine Learning. \u003cem\u003eJ. Cul. Her. \u003c/em\u003e\u003cstrong\u003e72\u003c/strong\u003e, 48-58 (2025) https://doi.org/10.1016/j.culher.2025.01.002\u003c/li\u003e\n\u003cli\u003eBreiman, L., Random Forests, \u003cem\u003eMachine Learning\u003c/em\u003e\u003cstrong\u003e45\u003c/strong\u003e, 5-32 (2001) https://doi.org/10.1023/A:1010933404324\u003c/li\u003e\n\u003cli\u003eHo, T.K. Random decision forests in Proceedings of Third International Conference on Document Analysis and Recognition, vol. 1(ICDAR \u0026apos;95), IEEE Computer Society, USA, 278-282 (1995)\u003c/li\u003e\n\u003cli\u003eWang, G. et al. Multi-Objective Optimization of Laser Cleaning Quality of Q390 Steel Rust Layer Based on Response Surface Methodology and NSGA-II Algorithm. \u003cem\u003eMaterials\u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 3109 (2024). https://doi.org/10.3390/ma17133109 \u003c/li\u003e\n\u003cli\u003eFang, C. et al. Effect of Laser Cleaning Parameters on Surface Filth Removal of Porcelain Insulator. \u003cem\u003ePhotonics\u003c/em\u003e\u003cstrong\u003e10(3)\u003c/strong\u003e, 269 (2023). https://doi.org/10.3390/photonics10030269 \u003c/li\u003e\n\u003cli\u003eYu, J. \u0026amp; Chen Y., Prediction of simulation parameters of fiber laser cleaning Range hood based on BP neural network. \u003cem\u003eJ. Phys.: Conf. Ser.\u003c/em\u003e\u003cstrong\u003e1820\u003c/strong\u003e 012118 (2021) https://doi.org/10.1088/1742-6596/1820/1/012118\u003c/li\u003e\n\u003cli\u003eChu, D. et al. Real-time acoustic monitoring of laser paint removal based on deep learning, \u003cem\u003eOpt. Express\u003c/em\u003e\u003cstrong\u003e33\u003c/strong\u003e, 1421-1436 (2025) https://doi.org/10.1364/OE.545906\u003c/li\u003e\n\u003cli\u003eSun, B. et al. Cleanliness prediction of rusty iron in laser cleaning using convolutional neural networks. \u003cem\u003eAppl. Phys. A\u003c/em\u003e\u003cstrong\u003e126\u003c/strong\u003e, 179 (2020). https://doi.org/10.1007/s00339-020-3363-5\u003c/li\u003e\n\u003cli\u003eLin, D., et al., Real-time LIBS monitoring of laser-based layered controlled paint removal from aircraft skin based on random forest, \u003cem\u003eAppl. Opt.\u003c/em\u003e\u003cstrong\u003e62\u003c/strong\u003e, 2569-2576 (2023), https://doi.org/10.1364/AO.484404\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6565434/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6565434/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLaser-assisted cleaning has become an indispensable tool in heritage conservation due to its precision, control, and environmentally friendly nature. However, the complexity of deposition layers and the fragile condition of original surfaces necessitate careful monitoring to avoid irreversible damage. This work explores the integration of machine learning-assisted real-time acoustic monitoring in laser cleaning processes to enhance conservation efforts. By combining acoustic signals generated during laser-material interaction with machine learning, we elevate the precision and reliability of laser cleaning in the delicate context of cultural heritage restoration.\u003c/p\u003e","manuscriptTitle":"Machine learning assisted real-time acoustic monitoring of laser cleaning in Heritage conservation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-29 14:49:55","doi":"10.21203/rs.3.rs-6565434/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-07-21T16:07:29+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-21T14:01:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"144951795094081727525252932353256062261","date":"2025-07-09T13:03:59+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-29T11:38:06+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"259789176974225520052170200331196438803","date":"2025-06-27T06:09:52+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-27T15:53:42+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-25T19:58:56+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-15T09:48:24+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj heritage science","date":"2025-05-15T09:47:16+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7b500dd8-bdad-45c2-9a15-3e29819e66cb","owner":[],"postedDate":"May 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-12-08T16:07:15+00:00","versionOfRecord":{"articleIdentity":"rs-6565434","link":"https://doi.org/10.1038/s40494-025-02146-3","journal":{"identity":"npj-heritage-science","isVorOnly":false,"title":"npj Heritage Science"},"publishedOn":"2025-12-04 15:57:20","publishedOnDateReadable":"December 4th, 2025"},"versionCreatedAt":"2025-05-29 14:49:55","video":"","vorDoi":"10.1038/s40494-025-02146-3","vorDoiUrl":"https://doi.org/10.1038/s40494-025-02146-3","workflowStages":[]},"version":"v1","identity":"rs-6565434","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6565434","identity":"rs-6565434","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}