You Only Put Your Nanoparticle: A Fully Automated System for Nanoparticle Washing Enabled by Vision and Language AI

You Only Put Your Nanoparticle: A Fully Automated System for Nanoparticle Washing Enabled by Vision and Language AI | 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 You Only Put Your Nanoparticle: A Fully Automated System for Nanoparticle Washing Enabled by Vision and Language AI Sang Soo Han, Heeseung Lee, Daeho Kim, Hyein Lee, Namyoung Gwak, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7060706/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Self-driving laboratories are reshaping materials discovery by combining automated experimentation with AI-driven decision-making. However, the lack of automation in key preprocessing steps such as nanoparticle (NP) washing—remains a major barrier to achieving full experimental autonomy. Effective automation of NP washing requires visual adaptivity, to detect subtle changes in the appearance of dispersions or precipitates, and cognitive adaptivity, to handle failure cases like incomplete sedimentation or phase separation. These demands make NP washing uniquely challenging despite its apparent simplicity. We introduce a fully integrated NP washing platform that combines computer vision with a large language model (LLM) to enable intelligent, end-to-end preprocessing in self-driving labs. The system employs YOLACT for real-time robotic manipulation and latent mask R-CNN for uncertainty-aware image segmentation, achieving 100% task success across 60 trials and accurately processing 45 diverse precipitate images. A retrieval-augmented LLM autonomously generates and continuously refines washing protocols based on cognitive feedback from failure detection. The platform was validated on NiFe layered double hydroxides, IrRu nanoparticles, and CdSe/CdS quantum dots. Electrochemical and photoluminescence analyses demonstrated that the automated washing matches or exceeds the quality of expert-level manual processing. This work represents a critical step toward fully autonomous, closed-loop experimentation by bridging synthesis and characterization through intelligent preprocessing. Physical sciences/Materials science/Techniques and instrumentation/Design, synthesis and processing Physical sciences/Materials science/Techniques and instrumentation/Imaging techniques Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Main Automation in chemical experiments has gained significant attention for its potential to minimize human labor and enhance experimental reproducibility. 1 – 5 With rapid advancements in artificial intelligence (AI), robotics, and machine learning, the concept of fully autonomous laboratories that integrate AI and automated experimental systems is becoming increasingly feasible. 6 – 13 While automation has demonstrated considerable success in predefined repetitive tasks such as chemical reaction screening and material synthesis, real experimentation still requires sophisticated decision-making and adaptability to dynamic conditions, which are challenges that traditional rule-based approaches often fail to address. 4 For example, chemical laboratories operate in highly dynamic physical environments, where simple XYZ-based actuators or preprogrammed robotic arms are often inadequate. Factors such as object positioning, lighting conditions, vessel orientation, and the efficiency of material separation can vary significantly between experimental trials. These uncertainties call for an intelligent system that can perceive, reason, and adapt dynamically to the experimental conditions by integrating physical motion planning, visual recognition, and cognitive decision-making. 14 – 18 In the development of materials, the preprocessing of synthesized material samples is an essential step for accurate characterization and evaluation of their properties. Among these preprocessing steps, centrifuge-based washing of as-synthesized nanoparticles (NPs) is particularly critical, as it directly affects the accuracy and reliability of the subsequent characterization processes. 19 – 25 However, automating this washing process presents significant challenges because of the need for high adaptability. For instance, an automated system must be capable of detecting variable hole positions in the centrifuge after each operation and distinguishing between phases within falcon tubes (e.g., supernatant vs. precipitate). These requirements suggest that, under conditions of high environmental variability, fully automating the washing process necessitates not only advanced visual adaptability but also cognitive capabilities, particularly the ability to reason about and respond intelligently to dynamic and uncertain experimental contexts 18 , 26 – 30 . To address these challenges, we propose an intelligent system that integrates vision and language AI models to fully automate nanoparticle washing workflows. Our platform leverages the combined capabilities of computer vision, a large language model (LLM), and robotic motion planning to execute experimental procedures with minimal human intervention. Together, these components form a visually and cognitively adaptive framework capable of recognizing experimental setups, dynamically adjusting motion strategies on the basis of real-time visual input, and recommending optimal washing parameters—such as solvent selection and centrifugation conditions—by extracting and interpreting contextual chemical knowledge from natural language instructions and reagent information. As illustrated in Fig. 1 , our automated washing platform begins operation in response to a simple user command (e.g., ‘Please wash my nanoparticles’ and ‘Nanoparticles synthesized with the following reagents’ ). Upon receiving such a request, the system autonomously retrieves relevant protocols on the basis of the LLM, plans the necessary sequence of actions, and executes the entire washing procedure—including solution addition, centrifugation, and supernatant removal. An integrated vision system accurately identifies the location of Falcon tubes, including the synthesized NPs, and evaluates the success of phase separation, whereas the LLM extracts contextual knowledge from the scientific literature to generate on-demand experimental recipes. This entire workflow is implemented as a retrieval-augmented generation (RAG) process, with each iteration incorporating experimental feedback. 31 , 32 If the initial parameters fail to yield effective separation of the liquid and solid phases in Falcon tubes, the system engages a closed-loop refinement cycle: the experimental results are fed back to the LLM, which then updates its recommendations and proposes a revised protocol. This adaptive loop enables continuous system improvement without requiring further human intervention. We also demonstrate the implementation and validation of this platform using three representative nanomaterials: NiFe-layered double hydroxide (LDH), IrRu bimetallic NPs, and CdSe/Cds quantum dots (QDs). Comparative evaluations against manual washing and untreated controls show that our system achieves equivalent or superior reproducibility, enhanced removal of residual organics/inorganics, and improved material performance, highlighting the promise of AI-driven automation in next-generation laboratory workflows. Results and Discussion Development of Automated Hardware Platform for Nanoparticle Washing Various techniques have been suggested for washing synthesized NPs, including centrifugation, filtration, dialysis, and magnetic separation. The efficiency of these methods depends largely on factors such as the NP size, surface properties, and aggregation tendencies. 19 , 21 – 25 , 33 – 36 Among these methods, centrifugation is particularly effective because of its high separation efficiency, which is based on differences in NP size and density. This makes it suitable for applications requiring high-purity NPs, such as catalysts and quantum dots. Based on these advantages, this study employs a centrifuge-based washing method. To enhance both efficiency and consistency, we designed a fully automated NP washing system by systematically integrating multiple laboratory automation components. Figure 2 presents the front and top views of the system, whereas Supplementary Video S1 shows the operation of each hardware component, facilitating a clearer understanding of the system’s functionality. The automated platform comprises a 6-axis robotic arm, a multichannel dispenser, container holder. The robotic arm (Doosan Robotics, M0609; Supplementary Figure S1 ) significantly enhances the system’s degrees of freedom, enabling a broader range of automated tasks. It is equipped with a tool changer (Supplementary Figure S2(a)) that supports autonomous switching between a 2-finger gripper (OnRobot RG6) and a vacuum gripper (VCG10), enabling secure handling and precise placement of chemical containers, even within the confined space of the centrifuge. Additionally, a webcam mounted on the robotic arm (Supplementary Figure S2(b)) provides real-time visual recognition, enabling the arm to identify containers and dynamically adjust its position. This feature further enhances the accuracy and reliability of the system’s operations. The multidispenser is equipped with a liquid pump (NEXT 50FJ) for removing supernatants and syringe pumps (Cavro XCalibur) for the precise dispersal and extraction of solvents (Supplementary Figure S3). The system comprises linear stages, 3D-printed components, and servo motors and accommodates up to five dispensing tips for the simultaneous handling of multiple solutions. Compared with manual techniques, pump-based solvent removal minimizes the risk of precipitate loss. The dispensing tips are precisely aligned within conical tubes to ensure high positional accuracy during operation. After each use, the tips are rinsed with ethanol to prevent contamination and maintain experimental consistency. Additionally, the actuator is used to laterally transfer the chemical vessels to the dispensing unit, which is located outside the direct reach of the robotic arm (Supplementary Figure S4). NP washing was performed using a conventional benchtop centrifuge (Hanil Scientific CombiR515, Supplementary Figure S5). To automate lid operation, a custom 3D-printed component was integrated with a wall-mounted linear actuator, providing operational stability. As the centrifuge lacks native support for external communication, we developed a control system utilizing an Arduino microcontroller and a relay module. This setup enables serial communication between the centrifuge and the central control system, allowing it to be integrated into the automated process. An ultrasonic bath (BANDELIN SONOREX Digitec; Supplementary Figure S6) was used to disperse the precipitates after the addition of solvent, promoting the dissolution and removal of residual impurities. A 3D-printed holder within the bath allows simultaneous processing of up to six samples, improving throughput. Additional hardware specifications are detailed in the supplementary information (SI). Reliability is a key consideration in the development of automated chemical systems. To evaluate the performance of the liquid handling module, we measured its injection precision. The system consistently dispensed 5 mL of various solvents—water, ethanol, acetone, and toluene—with standard deviations below 0.004 mL, as illustrated in Supplementary Figure S9. Vision-based Robot Motion Planning: Centrifugation Task We developed a visual feedback system to enable automated experimentation by adaptively recognizing the positions and states of samples within a dynamic centrifuge environment. After each configuration cycle, the position of the rotor inside the centrifuge varies, whereas the robotic arm operates from a fixed, predefined location. This positional inconsistency limits the accuracy of robotic sample placement. Furthermore, due to the bilaterally symmetric design of the rotor, the postcentrifugation arrangement of samples is also symmetric, making it difficult to distinguish between individual sample locations. To overcome this challenge, three color stickers (red, blue, and yellow) were attached inside the rotor. These colors were selected to contrast effectively with the commonly used cap colors of chemical vessels in laboratories. A color recognition algorithm was implemented to automatically assign pick-and-place sequences for each sample, thereby ensuring precise handling and preventing misidentification. The system operates through three main stages: (1) perception, (2) robot motion planning, and (3) robot execution. The overall system architecture is illustrated in Fig. 3 . During the perception phase, object detection and instance segmentation were performed to classify three key elements within the centrifuge: the empty hole, the filled hole, and the rotor center . For this task, we employed YOLACT, a real-time instance segmentation model known for its balance between speed and accuracy. 37 As shown in Supplementary Figure S10, the model achieved a mean average precision (mAP) of 99.99% at an intersection over union (IoU) threshold of 0.5, indicating highly reliable visual recognition performance. To evaluate the operational reliability of the system, the centrifuge was run at 1000 rpm for 1 minute, and randomized robotic pick-and-place tasks were subsequently performed. The experiment was repeated 10 times with 6 samples per trial, yielding a 100% task success rate across all repetitions (Supplementary Video S2). These results demonstrate the robustness and reliability of the system, even under visually variable centrifugation conditions induced by centrifugation. Vision-based Robot Motion Planning: Supernatant Removal Following centrifugation, the removal of the supernatant is a critical step to ensure experimental reproducibility. To automate this process, we developed a vision-guided robotic system that incorporates motion planning. Specifically, the latent mask R-CNN algorithm was employed as the vision model due to its ability to detect and quantify uncertainty. While YOLACT offers the advantage of fast detection for static objects with simple geometries such as the rotor component of a centrifuge, it is less effective in the supernatant removal process, where precipitates often exhibit complex and irregular morphologies. These challenging features limit the applicability of YOLACT in such scenarios. The overall workflow for the supernatant removal task is illustrated in Fig. 4 (a). Before initiating the removal process, the system first determines whether proper precipitation and phase separation have occurred; this step is referred to as the “detect the precipitate” stage. As shown in Figs. 4 (b) and (c), cases in which the precipitate and liquid layers are visually distinct are classified as successful separations. Conversely, if only the liquid phase is observed without any discernible precipitate, as shown in Figs. 4 (d) and 4(e), the system concludes that the centrifugation has failed. When clear separation between the precipitate and supernatant is observed, as shown in Fig. 4 (c), the robotic system proceeds with the aspiration of the supernatant. However, due to distortions caused by the fixed camera angle and the variable orientation of the chemical vessels after centrifugation (Fig. 4 (b)), the vision model may occasionally misclassify a correctly separated sample as a failure. This misclassification arises from two types of uncertainty: epistemic uncertainty, due to insufficient training data covering diverse vessel orientations, and aleatoric uncertainty, resulting from visual distortions caused by the camera perspective and environmental factors. Such errors may necessitate experimental repetition or parameter adjustment, thereby disrupting downstream procedures and reducing overall operational efficiency. To address this challenge, we implemented a feedback mechanism based on uncertainty estimation from the latent mask R-CNN model. 38 When high classification uncertainty is detected as shown in Fig. 4 (a), the model triggers an active feedback loop: the robotic arm rotates the sample by 60° to capture an alternative visual perspective, which is then reanalyzed. This process enhances the robustness and accuracy of sample assessment by enabling a more comprehensive evaluation of the separation state. The feedback mechanism not only improves classification accuracy but also reduces the need for repeated full rotations. By minimizing unnecessary motion, it significantly enhances experimental throughput and mitigates solvent evaporation risks, particularly for volatile washing agents such as acetone, ethanol, and isopropyl alcohol (IPA). To evaluate the effectiveness of the latent mask R-CNN, we conducted a comparative study using 45 images, as presented Supplementary Figure S13. Among the 25 images where centrifugation was successful, the latent mask R-CNN correctly identified all cases, while the conventional mask R-CNN 39 misclassified two cases. For samples lacking visible precipitates, both models accurately recognized the absence of separation. However, in the 10 visually ambiguous cases, the Latent Mask R-CNN consistently reported high uncertainty, whereas the mask R-CNN erroneously classified five of them as lacking precipitates. These results demonstrate the superior performance of the latent mask R-CNN in handling complex or ambiguous morphologies, particularly when subtle visual features such as precipitate boundaries must be accurately distinguished. Adaptive Washing Parameter Recommendation via LLMs with Cognitive Failure Feedback Figure 5 (a) presents the architecture of the LLM-based washing parameter recommendation system, which includes an integrated cognitive failure feedback loop. The system is designed not only to recommend optimal washing parameters—such as solvent type and volume, and centrifugation RPM and duration—based on basic reagent input, but also to adaptively ensure successful precipitation outcomes, particularly when initial attempts fail. This is achieved through an intelligent feedback mechanism that guides experimental modifications in real time. At the core of the system is a Retrieval-Augmented Generation (RAG) framework. It utilizes reagent information provided by the user, typically associated with NP synthesis, to query a user-specific literature database and identify appropriate washing protocols. In Step 1, the system formulates a search query using the input reagents, framed as " nanoparticles synthesized with the following reagents ". In Step 2, OpenAI’s Assistant is used to retrieve the most relevant document from a literature repository in an OpenAI vector store. Relevance is determined via the cosine similarity between embedding vectors, and the document with the highest similarity score is selected for further analysis. 40 A key advantage of this system is its scalability: user documents (e.g., scientific papers) can be seamlessly incorporated without requiring retraining or regeneration of embeddings—an overhead typically encountered in local vector search systems. In Step 3, a contextual compression retriever 41 is used to refine the search and extract specific washing protocol information from the selected document. In Step 4, the extracted washing parameters are incorporated into a RAG-based prompt (see Supplementary Figure S18), which is then processed by the LLM to generate optimal washing recommendations. When relevant literature is unavailable, or RAG is not applied, the system defaults to generate washing parameters based solely on the LLM’s internal knowledge (Supplementary Figure S19). If the recommended parameters lead to a failed outcome, the system triggers the cognitive adaptive system for failure feedback. This mechanism enables the LLM to suggest an alternative recipe by analyzing the previous attempt and its associated failure mode—without relying on retraining or external datasets. The corresponding prompt structure for this process is detailed in Supplementary Figure S20. Figure 5 (b) presents a case study involving the washing of IrRu NPs. When CTAB was used as a surfactant during synthesis, the LLM’s initial recommendation (water + ethanol) successfully induced NP precipitation. However, under the same washing conditions, replacing CTAB with PVP led to precipitation failure (Fig. 5 (c)). In response, the failure feedback loop proposed an alternative solvent combination—ethanol and acetone—with lower polarity, which led to successful precipitation. This result is consistent with established chemical principles: the lower dielectric constant of acetone diminishes electrostatic repulsion among NPs, thereby facilitating effective sedimentation. 42 These findings demonstrate the ability of the LLM to incorporate fundamental solvent properties into experimental decision-making, extending its role beyond simple prompt-response applications. Furthermore, the cognitive failure feedback system offers more targeted and efficient optimization of experimental conditions than traditional approaches, such as merely increasing centrifugation speed or time. Validation of the Automated Washing System To evaluate the generality and reliability of the proposed automated washing system, we conducted experiments on three representative classes of nanomaterials, namely, NiFe-based LDH, bimetallic IrRu NPs, and CdSe/CdS quantum dots (QDs), as shown in Fig. 6 . For each material, we compared the effects of three conditions: no washing (baseline), manual washing by domain experts, and an automated system using parameters recommended by an LLM. This comparison allows us to evaluate the impact of washing on the structural, chemical, and functional properties of the nanomaterials. NiFe-Based LDH NiFe-based LDH is a prototypical layered structure widely used as a catalyst. 43 , 44 Following synthesis, residual surfactants such as PVP and precursor-derived byproducts often remain on the surface, potentially impairing the catalytic performance of the material. Thus, we investigated how both the presence and method of washing affect the structural integrity and electrochemical characteristics of the material. As shown in Fig. 6 (a), we compared the overpotentials for the oxygen evolution reaction (OER) among the three conditions. The unwashed sample exhibited a relatively high overpotential, indicative of poor catalytic efficiency. In contrast, both the manually washed and automatically washed samples showed significantly reduced overpotentials, with no discernible difference between them. This finding demonstrates that the automated system can match the washing effectiveness of a human expert. The observed reduction in overpotential highlights the role of washing in removing surface impurities and electrochemically inactive species, ultimately enhancing the active surface area and improving the catalytic activity. Figure 6 (b) presents scanning electron microscopy (SEM) images of the washed sample, revealing the characteristic sheet-like morphology of LDH on the surface. This structural feature is absent in the unwashed sample, which suggests that surface contaminants obscure or inhibit the formation of the typical LDH morphology. The results demonstrate that the washing process not only purifies the surface but also facilitates the manifestation of intrinsic nanostructural features. IrRu NPs IrRu alloy catalysts are well recognized for their excellent electrocatalytic performance in the OER. 45 , 46 As shown in Fig. 6 (c), both the manually washed and system-washed IrRu NPs exhibited markedly improved OER activity compared with that of the unwashed sample. The unwashed catalyst showed a significantly higher overpotential and lower current density, indicative of surface contamination by insulating organic residues. These contaminants, likely residual surfactants and synthesis byproducts, hinder effective charge transfer and catalytic activity. In contrast, the catalysts washed either by a human expert or through the automated system—both using LLM-recommended parameters—demonstrated comparable and enhanced electrocatalytic activity. This result highlights the effectiveness of the LLM-guided automated washing protocol in replicating expert-level outcomes, confirming its reliability in the preparation of clean and active catalyst surfaces. Figure 6 (d) presents transmission electron microscopy (TEM) images of the IrRu NP catalysts before and after washing. In the unwashed sample, densely packed surface residues are clearly observed, indicating the presence of unremoved organic species. After washing, the NP surfaces appeared significantly cleaner, with a more uniform dispersion of individual particles. This morphological improvement suggests the successful elimination of organic impurities, further supporting the observed increase in electrochemical performance. CdSe/CdS QDs For CdSe/CdS QDs, achieving uniform particle size is essential, as their optical and electronic properties are highly sensitive to size distribution. 47 – 49 Accordingly, purification is one of the most critical steps in the synthesis workflow, as it directly influences the homogeneity, purity, and reproducibility of the final product. The effectiveness of this purification process is strongly governed by the washing parameters applied. Figure 6 (e) shows the photoluminescence (PL) spectra of CdSe/CdS QDs subjected to three different washing conditions. The unwashed sample exhibited two distinct emission peaks at 535 nm and 605 nm, indicating the presence of two distinct size populations within the QD ensemble. This result confirms that, without purification, the synthesized QDs consisted of a heterogeneous mixture of particle sizes. Upon applying the initial washing protocol generated by the LLM, the 535 nm peak was substantially reduced, suggesting partial removal of the smaller QDs. However, the persistence of this peak indicated incomplete purification. Through an automated failure feedback loop, the system subsequently refined the washing parameters, ultimately eliminating the 535 nm peak entirely. This indicates the successful removal of size heterogeneity and highlights the efficacy of iterative protocol optimization via feedback. Notably, the PL spectrum of the system-washed sample closely mirrored that of the manually washed QDs, validating the automated system’s ability to match expert-level purification performance. This result illustrates the key capability of the autonomous experimental platform: it can not only induce physical separation processes such as selective precipitation, but also interpret spectral features to refine protocols based on material-specific feedback. As shown in Supplementary Figure S25, the washing recipe generated through the failure feedback loop demonstrates the system’s capacity to detect suboptimal outcomes and autonomously redesign experiments for improved results—exemplifying property-aware optimization in nanomaterial synthesis. Figure 6 (f) shows low-magnification TEM images of the CdSe/CdS QD films before and after washing. In the unwashed sample, numerous dark patches and aggregated domains are evident, indicative of residual surfactants and poor particle dispersion. After washing, the film displays a notably cleaner surface with reduced contamination and improved uniformity. High-resolution inset images further confirm the presence of well-dispersed, size-uniform QDs, underscoring the effectiveness of the purification protocol employed by the LLM-guided method. Conclusions In this study, we developed an adaptive automation system for centrifuge-based NP washing by integrating computer vision and an LLM. The vision module of this system employs YOLACT for real-time detection and localization of centrifuge tubes. After centrifugation, the latent mask R-CNN is used to classify the supernatant and precipitate regions and to quantify the segmentation uncertainty. When the segmentation uncertainty is high, the system initiates a visual feedback loop by reorienting the sample to reduce ambiguity and improve the classification accuracy. To support decision-making, the system incorporates the GPT-4o-mini LLM in combination with a RAG pipeline to suggest initial centrifugation parameters. These parameters are iteratively refined through experimental feedback, particularly in response to failure cases identified during operation. This system was validated across three representative nanomaterial classes: NiFe-LDHs, IrRu NPs, and CdSe/CdS QDs. The results demonstrate that the integration of perception (via computer vision) and cognition (via LLM reasoning) enables a transition from rigid, rule-based procedures to adaptive, intelligent experimental workflows. By facilitating robust operation under uncertainty and enabling data-driven refinement of protocols, our system paves the way for intelligent, reproducible, and scalable experimental design. The success of this system highlights the broader potential of autonomous workflows in both material synthesis and characterization. Methods Training datasets for computer vision models YOLACT for the Centrifugation Task A YOLACT instance segmentation model was developed to detect key components within the centrifugation environment, including the rotors, Falcon tubes, and empty holes. The training dataset comprised 295 original images, which were augmented using color transformations including red and purple overlays, resulting in a total of 885 images. Model training was based on a customized yolact_base_config 37 derived from the original COCO 50 configuration, featuring a ResNet-101 backbone, an input image size of 550×550 pixels, and a maximum of 1,000 iterations. The model was trained on an NVIDIA RTX A4000 GPU, with the dataset split into 80% for training and 20% for validation. Additional training parameters, including learning rate steps and IOU thresholds, followed the default base configuration. Latent Mask R-CNN for removing the supernatant To classify and detect the supernatant removed after centrifugation, a latent mask R-CNN model was trained using 1,926 manually annotated experimental images. The model utilized the mask_rcnn_R_50_FPN_1x architecture from the Detectron2 51 framework. Training was conducted with standard hyperparameters, including an initial learning rate of 1x10 -4 and a batch size of 4, on an NVIDIA RTX 3060 Ti GPU. The dataset was divided to achieve a training-to-validation ratio of 80:20. The incorporation of latent representations enhanced the model's generalization capability across diverse experimental variations, including differences in supernatant appearance, precipitate shape, sample holder, and Falcon tube positioning. Robot motion planning for the centrifugation task To enable precise robotic manipulation within the centrifuge, key objects, including the chemical vessels, rotor holes, and rotor center, were first detected from the input image and used to generate processed data suitable for motion planning. The image coordinates of these objects were then computed. Because the robot arm observes the centrifuge interior from a fixed position, the rotor center serves as a consistent reference point. This reference was mapped to the robot’s coordinate system, enabling the calculation of relative positions for the chemical vessels and holes. To determine the correct processing sequence, color-based object recognition was employed using distinct stickers (red, yellow, and blue) affixed inside the rotor. These stickers were carefully selected to avoid color overlap with Falcon tube caps. For example, if a Falcon tube with a red cap is used, blue or yellow stickers are selected to prevent recognition errors. Once the center coordinates of the color-coded stickers were identified, the objects were sorted in a clockwise order based on their angular positions relative to the reference sticker. The angle 𝜃 between the reference sticker and each object was computed using the following equations: $$\:\theta\:=\text{atan}2({y}_{color}-{y}_{object},{x}_{color}-{x}_{object})$$ 1 $$\:{\theta\:}_{deg}=\theta\:\:\times\:\left(\frac{180}{{\pi\:}}\right)$$ 2 If the calculated angle was negative, 360° was added to convert it to a positive value. The angular difference, Δ𝜃, between the reference sticker and each object was then calculated as follows: Δ𝜃 = \(\:{\theta\:}_{color}-{\theta\:}_{object}\) (3) This normalized angular difference allows for consistent sorting of objects in a clockwise sequence. Once the object sequence is determined, the positional data of each object are transformed into the robot’s coordinate system. By using the rotor center previously mapped in the robot’s coordinate space as a fixed anchor, the relative image-based coordinates of each object were added to the robot center’s robot-frame coordinates ( \(\:{x}_{center\:of\:rotor},\:\:{y}_{center\:of\:rotor})\) . This transformation enables the robot to accurately position and manipulate each sample. Uncertainty quantification using the latent mask R-CNN We utilized latent mask R-CNN to mitigate image distortions encountered during the supernatant removal process and to improve the segmentation accuracy between the liquid (supernatant) and precipitate regions. Unlike traditional deterministic segmentation methods, latent mask R-CNN incorporates a latent variable \(\:\mathcal{z}\) into the mask prediction process, enabling a probabilistic framework. This allows the model to generate multiple plausible segmentation masks, thereby capturing the inherent distributional uncertainty present in the data. To quantify segmentation uncertainty, we sampled \(\:\mathcal{n}\) masks from the latent variable \(\:\mathcal{z}\) and computed the pixelwise mean \(\:{\mu\:}\) and variance \(\:{{\sigma\:}}^{2}\) . The coefficient of variation (CV) for each pixel is defined as follows: $$\:\text{C}\text{V}\:=\frac{{\sigma\:}}{{\mu\:}}$$ 4 It was then computed separately for the supernatant and precipitate classes. Higher CV values correspond to regions of greater segmentation uncertainty, indicating lower confidence in the predicted mask boundaries. This uncertainty-aware approach facilitates more reliable downstream decision-making in automated experimental workflows. Recipe generator via LLM To generate optimal washing conditions, we used the GPT-4o-mini model from OpenAI within the LangChain framework. The model was configured with a temperature setting of 0 to ensure deterministic and consistent outputs. Input prompts were constructed from structured experimental metadata, including reagent names, volumes, concentrations, relevant literature on washing protocols (retrieved text through the retrieval process), and records of prior experimental failures when available. To tailor prompt generation based on the available context, we used different prompt augmentation strategies: RAG-based augmentation was applied when relevant literature could be retrieved; failure-based augmentation was used when a previous washing attempt had failed; and metadata-only prompts were used in the absence of both literature and failure history. The specific prompt templates used for each case are provided in Supplementary Figures S18 ~ S20. Model outputs were returned in JSON format and processed programmatically following a preprocessing step to ensure correct formatting and interpretability. In cases of failure, a feedback loop was implemented: the failure conditions were incorporated into the next prompt, allowing the model to refine its recommendations iteratively. This iterative framework enables continuous improvement of washing protocols without requiring additional fine-tuning of the model or repeated manual intervention. Implementation of the retriever system To build the retrieval system, we utilized OpenAI’s vector store to create the embedding database. PDF documents were uploaded and automatically embedded via the OpenAI API (Supplementary Figure S16). For efficient and context-preserving retrieval, the documents were segmented into overlapping chunks—each consisting of 800 tokens with a 400-token overlap. The retrieval process consisted of two stages: Initial Retrieval: Using a list of synthesized reagents as input, the GPT Assistant API queried the embedded document database in the OpenAI vector store to identify literature relevant to the synthesis and washing process. Contextual Compression: In the second stage, we applied ContextualCompressionRetriever to filter the initial results and extract only the segments specifically related to the washing protocols. This component integrated a base retriever with an LLM-based compressor, LLMChainExtractor, powered by the GPT-4o-mini model. The model was executed with a temperature setting of 0 to maintain consistency and determinism. This two-tiered retrieval system enabled accurate identification of procedural content, which was subsequently used to inform and enhance the generation of washing condition recommendations. By isolating contextually relevant information from broader scientific literature, the retriever significantly improved the precision of downstream prompt engineering for the recipe generation module. Declarations Data and code availability Several examples of our result data, the codes, and related explanations are provided in the following GitHub repository (https://github.com/KIST-CSRC/washingModule). All codes are written in Python 3.9 and all environments could be created via requirements.txt file. Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Sang Soo Han ( [email protected] ). Acknowledgements This work was supported by the National Research Foundation of Korea funded by the Ministry of Science and ICT [NRF-2022M3H4A7046278 and RS-2024-00450102]. Author Contributions S.S.H. and H.S.L.(Heeseung Lee) conceived the idea. S.S.H. and S.S.S. supervised the project. H.S.L., D.K, N.K, and H.J.Y developed the hardware, and H.S.L. integrated vision and language AI into the system. H.L.(Heyin Lee) and T.Y. synthesized the LDH and IrRu samples and conducted their electrochemical measurements. N.Y.K. and N.O. synthesized the CdSe/Cds quantum dot samples and evaluated their optical properties. All authors contributed to the analysis of the results and the writing of the manuscript. Competing interests We have a patent pending titled “System and method for automated nanoparticle washing” (Korean Patent Application No. 10-2024-0170078). References Shi Y, Prieto PL, Zepel T, Grunert S, Hein JE (2021) Automated Experimentation Powers Data Science in Chemistry. Acc Chem Res 54:546–555. https://doi.org:10.1021/acs.accounts.0c00736 Du H, Corkan LA, Yang K, Kuo PY, Lindsey JS (1999) An automated microscale chemistry workstation capable of parallel, adaptive experimentation. Chemometr Intell Lab Syst 48:181–203 King RD et al (2009) The Automation of Science. Science 324:85–89. https://doi.org:10.1126/science.1165620 Selekman JA et al (2017) High-throughput automation in chemical process development. Annual Rev Chem Biomol Eng 8:525–547 Gromski PS, Henson AB, Granda JM, Cronin L (2019) How to explore chemical space using algorithms and automation. Nat Reviews Chem 3:119–128 Jiang Y et al (2022) An artificial intelligence enabled chemical synthesis robot for exploration and optimization of nanomaterials. Sci Adv 8:eabo2626. https://doi.org:10.1126/sciadv.abo2626 Yoo HJ et al (2024) Bespoke Metal Nanoparticle Synthesis at Room Temperature and Discovery of Chemical Knowledge on Nanoparticle Growth via Autonomous Experimentations. Adv Funct Mater 34. https://doi.org:10.1002/adfm.202312561 Burger B et al (2020) A mobile robotic chemist. Nature 583:237–. https://doi.org:10.1038/s41586-020-2442-2 Granda JM, Donina L, Dragone V, Long DL, Cronin L (2018) Controlling an organic synthesis robot with machine learning to search for new reactivity. Nature 559:377–381. https://doi.org:10.1038/s41586-018-0307-8 Rapp JT, Bremer BJ, Romero PA (2024) Self-driving laboratories to autonomously navigate the protein fitness landscape. Nat Chem Eng 1:97–107. https://doi.org:10.1038/s44286-023-00002-4 Tao HC et al (2021) Self-Driving Platform for Metal Nanoparticle Synthesis: Combining Microfluidics and Machine Learning. Adv Funct Mater 31 https://doi.org:ARTN 2106725. 10.1002/adfm.202106725 Yoo HJ, Lee KY, Kim D, Han SS (2024) OCTOPUS: operation control system for task optimization and job parallelization via a user-optimal scheduler. Nat Commun 15 https://doi.org:ARTN 9669.10.1038/s41467-024-54067-7 Roch LM et al (2020) ChemOS: An orchestration software to democratize autonomous discovery. PLoS ONE 15. https://doi.org:ARTN e0229862 10.1371/journal.pone.0229862 Du Toit NE, Burdick JW (2011) Robot motion planning in dynamic, uncertain environments. IEEE Trans Robot 28:101–115 Taherimakhsousi N et al (2021) A machine vision tool for facilitating the optimization of large-area perovskite photovoltaics. Npj Computational Materials 7 https://doi.org:ARTN 190.10.1038/s41524-021-00657-8 Darvish K et al (2025) ORGANA: A robotic assistant for automated chemistry experimentation and characterization. Matter-Us 8 https://doi.org:ARTN 101897. 10.1016/j.matt.2024.10.015 Lisca G, Nyga D, Bálint-Benczédi F, Langer H, Beetz M in 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 5202–5208 (IEEE) Eppel S, Xu H, Bismuth M, Aspuru-Guzik A (2020) Computer Vision for Recognition of Materials and Vessels in Chemistry Lab Settings and the Vector-LabPics Data Set. ACS Cent Sci 6:1743–1752. https://doi.org:10.1021/acscentsci.0c00460 Chiba T et al (2017) High-Efficiency Perovskite Quantum-Dot Light-Emitting Devices by Effective Washing Process and Interfacial Energy Level Alignment. ACS Appl Mater Interfaces 9:18054–18060. https://doi.org:10.1021/acsami.7b03382 Nyongombe G, Kabongo GL, Noto LL, Dhlamini MS (2022) Investigating the Impact of the Washing Steps of Layered Double Hydroxides (LDH) on the Electrochemical Performance. Nanomaterials 12, 578 Adkins H, Billica HR (1948) The preparation of Raney nickel catalysts and their use under conditions comparable with those for platinum and palladium catalysts. J Am Chem Soc 70:695–698 Hoshi K et al (2018) Purification of perovskite quantum dots using low-dielectric-constant washing solvent diglyme for highly efficient light-emitting devices. ACS Appl Mater Interfaces 10:24607–24612 Zhou S (2021) Rapid separation and purification of lead halide perovskite quantum dots through differential centrifugation in nonpolar solvent. RSC Adv 11:28410–28419 Zhang J et al (2022) Surface enrichment of Ir on the IrRu alloy for efficient and stable water oxidation catalysis in acid. Chem Sci 13:12114–12121 Li D et al (2012) Surfactant removal for colloidal nanoparticles from solution synthesis: the effect on catalytic performance. ACS Catal 2:1358–1362 El-Khawaldeh R et al (2024) Keeping an eye on the experiment: computer vision for real-time monitoring and control. Chem Sci 15:1271–1282. https://doi.org:10.1039/d3sc05491h Tiong LCO et al (2024) Machine vision-based detections of transparent chemical vessels toward the safe automation of material synthesis. npj Comput Mater 10:42 Kendall A, Gal Y (2017) What uncertainties do we need in bayesian deep learning for computer vision? Adv Neural Inf Process Syst 30 Xia Y, Zhang J, Jazdi N, Weyrich M (2024) Incorporating Large Language Models into Production Systems for Enhanced Task Automation and Flexibility. arXiv preprint arXiv:2407.08550 Xia Y, Jazdi N, Zhang J, Shah C, Weyrich M (2024) Control Industrial Automation System with Large Language Models. arXiv preprint arXiv:2409.18009 Tang Y, Yang Y (2024) Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries. arXiv preprint arXiv:2401.15391 Lewis P et al (2020) Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Adv Neur In 33 Wei L et al (2008) Effect of centrifugation on the purity of single-walled carbon nanotubes from MCM-41 containing cobalt. J Phys Chem C 112:17567–17575 Zhang Y, Li D, Tian XA, Highly Efficient (2024) Fluorescent Turn-Off Nanosensor for Quantitative Detection of Teicoplanin Antibiotic from Humans, Food, and Water Based on the Electron Transfer between Imprinted Quantum Dots and the Five-Membered Cyclic Boronate Esters. Molecules 29:4115 Weeranoppanant N, Adamo A (2020) In-Line Purification: A Key Component to Facilitate Drug Synthesis and Process Development in Medicinal Chemistry. ACS Med Chem Lett 11:9–15. https://doi.org:10.1021/acsmedchemlett.9b00491 Lordi V, Yao N, Wei J (2001) Method for supporting platinum on single-walled carbon nanotubes for a selective hydrogenation catalyst. Chem Mater 13:733–737 Bolya D, Zhou C, Xiao FY, Lee YJ (2019) YOLACT Real-time Instance Segmentation. Ieee I Conf Comp Vis 9156–9165. https://doi.org:10.1109/Iccv.2019.00925 Liu YX, Mishra N, Abbeel P, Chen X (2023) Distributional Instance Segmentation: Modeling Uncertainty and High Confidence Predictions with Latent-MaskRCNN. Ieee Int Conf Robot 7069–7075. https://doi.org:10.1109/Icra48891.2023.10160812 He K, Gkioxari G, Dollar P, Girshick R, Mask R-CNN (2020) IEEE Trans Pattern Anal Mach Intell 42:386–397. https://doi.org:10.1109/TPAMI.2018.2844175 OpenAI (2025) OpenAI assistant , https://platform.openai.com/docs/assistants/tools Verma S (2024) Contextual Compression in Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv preprint arXiv:2409.13385 Van der Hoeven PC, Lyklema J (1992) Electrostatic stabilization in non-aqueous media. Adv Colloid Interface Sci 42:205–277 Gu Y et al (2024) NiFe layered double hydroxides synthesized based on solvent properties as anode catalysts for enhanced oxygen evolution reaction. Chem Eng J 480 https://doi.org:ARTN 147789. 10.1016/j.cej.2023.147789 Hou CM et al (2021) Rapid large-scale synthesis of ultrathin NiFe-layered double hydroxide nanosheets with tunable structures as robust oxygen evolution electrocatalysts. RSC Adv 11:37624–37630. https://doi.org:10.1039/d1ra05045a Park Y et al (2025) Atomic-level Ru-Ir mixing in rutile-type (RuIr)O for efficient and durable oxygen evolution catalysis. Nat Commun 16. https://doi.org:ARTN 10.1038/s41467-025-55910-1 Bornet A et al (2024) Beyond RDE characterisation - Unveiling IrRu/ATO OER catalyst stability with a GDE setup. Electrochim Acta 501. 10.1016/j.electacta.2024.144773 . https://doi.org:ARTN 144773 Liu YC, Bose S, Fan WJ (2018) Effect of size and shape on electronic and optical properties of CdSe quantum dots. Optik 155:242–250. https://doi.org:10.1016/j.ijleo.2017.10.165 Samadi-Maybodi A, Tirbandpay R (2021) Synthesis, optical properties and tuning size of CdSe quantum dots by variation capping agent. Spectrochim Acta A 250 https://doi.org:ARTN 119369.10.1016/j.saa.2020.119369 Gwak N et al (2025) Comparative Study of Two Different Quasi-Type II Heterostructured Quantum Dots for Enhanced Electroluminescence. Acs Photonics 12:944–951. https://doi.org:10.1021/acsphotonics.4c01995 Lin TY et al (2014) Microsoft COCO: Common Objects in Context. Lect Notes Comput Sc 8693:740–755. https://doi.org:Doi 10.1007/978-3-319-10602-1_48 Girshick W-Y (2019) L. a. R. Detectron2 , https://github.com/facebookresearch/detectron2 Additional Declarations Yes there is potential Competing Interest. We have a patent pending titled “System and method for automated nanoparticle washing” (Korean Patent Application No. 10-2024-0170078). Supplementary Files SupplementaryVideo1.mp4 Supplementary Video_1 SupplementaryVideo2.mp4 Supplementary Video_2 SupplementaryInformation.docx Supplementary infromation Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7060706","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":490841893,"identity":"9bb001fa-d59b-4aaf-a37c-d73810091bd5","order_by":0,"name":"Sang Soo Han","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1klEQVRIiWNgGAWjYBADOQlmKEuCWC3GQC2MDSRpSZzBQKwW+fazh1/+qLBJn9nO/vwB4x4bBsnZB/BrMTiTl2bNcyYtdzYzj2EDw7M0Bmm+BAJaGHLMjBnbDufOY+YBOuzAYQY5HkIO639jZviz7XC6HDP7Q6CW/4S1MNzIMX7A23Y4QZqZAeiwAwcYpAlpMbjxxowZ6BfDmc08hjMSDiTzSPYQdFiO8UdgiMlLnD/+4MOHA3ZyEmcIOYyBgQ0REwkMDAR9AgLMH4hRNQpGwSgYBSMYAAC1vjwpaK0nugAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-7925-8105","institution":"Korea Institute of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Sang","middleName":"Soo","lastName":"Han","suffix":""},{"id":490841894,"identity":"3948f690-b077-4f4c-a81f-7f3204feb6c6","order_by":1,"name":"Heeseung Lee","email":"","orcid":"","institution":"Korea Institute of Science and Technolog","correspondingAuthor":false,"prefix":"","firstName":"Heeseung","middleName":"","lastName":"Lee","suffix":""},{"id":490841895,"identity":"45fc6e80-92f6-4d44-8e4e-4f9921adfee7","order_by":2,"name":"Daeho Kim","email":"","orcid":"","institution":"Korea Institute of Science and Technolog","correspondingAuthor":false,"prefix":"","firstName":"Daeho","middleName":"","lastName":"Kim","suffix":""},{"id":490841896,"identity":"ba07d415-6421-4b65-b11a-77feb757bef3","order_by":3,"name":"Hyein Lee","email":"","orcid":"https://orcid.org/0009-0001-8617-3927","institution":"Kyung Hee University, Yongin","correspondingAuthor":false,"prefix":"","firstName":"Hyein","middleName":"","lastName":"Lee","suffix":""},{"id":490841897,"identity":"aef5e9e9-b69e-42cf-a802-76bb778a4a91","order_by":4,"name":"Namyoung Gwak","email":"","orcid":"","institution":"Hanyang University","correspondingAuthor":false,"prefix":"","firstName":"Namyoung","middleName":"","lastName":"Gwak","suffix":""},{"id":490841898,"identity":"1bbeccfc-3178-408e-ae8d-ade84535ac55","order_by":5,"name":"Nayeon Kim","email":"","orcid":"","institution":"Korea Institute of Science and Technolog","correspondingAuthor":false,"prefix":"","firstName":"Nayeon","middleName":"","lastName":"Kim","suffix":""},{"id":490841899,"identity":"e685cbfd-d1b2-40b1-8391-985dcc912895","order_by":6,"name":"Hyuk Jun Yoo","email":"","orcid":"","institution":"Korea Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Hyuk","middleName":"Jun","lastName":"Yoo","suffix":""},{"id":490841900,"identity":"18ec33ac-b279-4488-ba3d-2650c9b9b3bb","order_by":7,"name":"Taekyung Yu","email":"","orcid":"https://orcid.org/0000-0003-4703-3523","institution":"Kyung Hee University","correspondingAuthor":false,"prefix":"","firstName":"Taekyung","middleName":"","lastName":"Yu","suffix":""},{"id":490841901,"identity":"924e3fc3-d3bb-456d-bde5-9772343c78a1","order_by":8,"name":"Nuri Oh","email":"","orcid":"https://orcid.org/0000-0001-9145-8911","institution":"Hanyang University","correspondingAuthor":false,"prefix":"","firstName":"Nuri","middleName":"","lastName":"Oh","suffix":""},{"id":490841902,"identity":"8a0f3652-cfa8-4e33-a7db-33c0cfc082a2","order_by":9,"name":"Seok Su Sohn","email":"","orcid":"https://orcid.org/0000-0002-4504-0028","institution":"Korea University","correspondingAuthor":false,"prefix":"","firstName":"Seok","middleName":"Su","lastName":"Sohn","suffix":""}],"badges":[],"createdAt":"2025-07-07 03:05:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7060706/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7060706/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87813883,"identity":"ccab212d-c5b5-492d-87eb-65f7661e8580","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":303525,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic of our fully automated nanoparticle washing system. The platform operates in an adaptive experimental environment by integrating computer vision and a large language model to respond intelligently to user commands.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/30bc932dafef8dcb9ccc8bad.png"},{"id":87813880,"identity":"7be21322-0f6d-41f2-9030-b5bdfea28898","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":443208,"visible":true,"origin":"","legend":"\u003cp\u003eActual implementation (left) and workflow (right) of the automated system for nanoparticle washing. The system executes a three-step washing action task consisting of solution addition, centrifugation, and supernatant removal. Computer vision is used to verify the correct placement of chemical vessels in the centrifuge and to evaluate the success or failure of centrifugation before proceeding to the next step.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/13eef3a7c4126ecd9374265c.png"},{"id":87813881,"identity":"4d5dc9de-657e-4d14-b95d-d9801e717c6a","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":803224,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Workflow for placing and retrieving Falcon tubes in and out of the centrifuge. The system uses input images for perception and motion planning to execute precise robotic actions. (b–d) Actual images captured from the centrifuge. (e–g) Corresponding inference results generated by the vision system.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/1c0db75c1c92487ae35dc99f.png"},{"id":87814227,"identity":"e326f302-11ff-4c17-81b1-9b049328ebc2","added_by":"auto","created_at":"2025-07-29 09:52:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":317130,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Workflow for the removal of the supernatant using Vision AI. The system detects the precipitate through visual perception and then performs robotic motion planning to remove the supernatant. If the boundary between the liquid and precipitate is unclear, the tube is rotated 60° to reduce uncertainty. (b, c) Successful supernatant removal: (b) a centrifuged sample with an unclear boundary; (c) the same sample after 60° rotation. (d, e) Failure cases: (d) a sample with failed separation; (e) the same sample after 60° rotation. (f–j) Inference results corresponding to (b–e), showing segmentation of the liquid, precipitate, Falcon tube, and holder by the perception system.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/2166d5adbe2f74d23cd5bc0f.png"},{"id":87813898,"identity":"1e060076-91cd-4c78-80ad-764751115d17","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":842327,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Workflow of an LLM-based system for recommending NP washing parameters and managing failure cases through an adaptive feedback loop. Upon receiving a list of reagents from the user, the system retrieves relevant literature and suggests optimal washing conditions, including solvent type, volume, and centrifugation parameters. In cases of experimental failure, the system iteratively refines the recommendations using feedback to improve the outcome. (b, c) Case study demonstrating the system’s application to IrRu NP washing. In (b), when CTAB was used as the surfactant, the LLM-recommended parameters led to successful purification, reflecting appropriate consideration of solvent polarity and NP solubility. In (c), when PVP was used as the surfactant, the initial washing attempt failed. However, following feedback-based adjustment, the updated solvent conditions enabled successful precipitation and purification.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/dd47b14ece100ede4d0af9f2.png"},{"id":87814228,"identity":"8b4be69a-7ca4-4c49-959a-e1673b1db32d","added_by":"auto","created_at":"2025-07-29 09:52:36","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":675160,"visible":true,"origin":"","legend":"\u003cp\u003eValidation results of our automated washing system across three classes of nanomaterials. (a, b) NiFe-based LDH, (c, d) IrRu NPs, and (e, f) CdSe/CdS QDs. For each material, we compared the effects of three conditions: no washing, manual washing by domain experts, and an automated system using parameters recommended by an LLM. (a, c, e) Electrochemical and PL data highlighting the functional performance of each sample. (b, d, f) SEM and TEM images showing morphological and structural differences before and after washing.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/7ca77bcf0dedff211825ec98.png"},{"id":90091959,"identity":"d3062610-1760-4840-9d57-0e23d37d5bd6","added_by":"auto","created_at":"2025-08-28 11:23:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4140791,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/b8929f89-9c23-4a9f-8eb2-f688aef92a27.pdf"},{"id":87813903,"identity":"67317ec1-a197-4584-83af-232347c9dd1a","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"mp4","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":30323802,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Video_1\u003c/p\u003e","description":"","filename":"SupplementaryVideo1.mp4","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/ff2ce5511aa2b6c5d423226a.mp4"},{"id":87813889,"identity":"477ad239-c189-4e59-a7a0-0a110e2418f0","added_by":"auto","created_at":"2025-07-29 09:44:36","extension":"mp4","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":13882726,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Video_2\u003c/p\u003e","description":"","filename":"SupplementaryVideo2.mp4","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/268a7d8b61e5a91e4d6067f2.mp4"},{"id":87813906,"identity":"975c4358-0964-4e9f-b456-6f55e0278dbe","added_by":"auto","created_at":"2025-07-29 09:44:38","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":79500270,"visible":true,"origin":"","legend":"Supplementary infromation","description":"","filename":"SupplementaryInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-7060706/v1/ba82d39fcd41b8979133a1ac.docx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e there is potential Competing Interest.\nWe have a patent pending titled “System and method for automated nanoparticle washing” (Korean Patent Application No. 10-2024-0170078).","formattedTitle":"You Only Put Your Nanoparticle: A Fully Automated System for Nanoparticle Washing Enabled by Vision and Language AI","fulltext":[{"header":"Main","content":"\u003cp\u003eAutomation in chemical experiments has gained significant attention for its potential to minimize human labor and enhance experimental reproducibility.\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e With rapid advancements in artificial intelligence (AI), robotics, and machine learning, the concept of fully autonomous laboratories that integrate AI and automated experimental systems is becoming increasingly feasible.\u003csup\u003e\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10 CR11 CR12\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e While automation has demonstrated considerable success in predefined repetitive tasks such as chemical reaction screening and material synthesis, real experimentation still requires sophisticated decision-making and adaptability to dynamic conditions, which are challenges that traditional rule-based approaches often fail to address.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e For example, chemical laboratories operate in highly dynamic physical environments, where simple XYZ-based actuators or preprogrammed robotic arms are often inadequate. Factors such as object positioning, lighting conditions, vessel orientation, and the efficiency of material separation can vary significantly between experimental trials. These uncertainties call for an intelligent system that can perceive, reason, and adapt dynamically to the experimental conditions by integrating physical motion planning, visual recognition, and cognitive decision-making.\u003csup\u003e\u003cspan additionalcitationids=\"CR15 CR16 CR17\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eIn the development of materials, the preprocessing of synthesized material samples is an essential step for accurate characterization and evaluation of their properties. Among these preprocessing steps, centrifuge-based washing of as-synthesized nanoparticles (NPs) is particularly critical, as it directly affects the accuracy and reliability of the subsequent characterization processes.\u003csup\u003e\u003cspan additionalcitationids=\"CR20 CR21 CR22 CR23 CR24\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e However, automating this washing process presents significant challenges because of the need for high adaptability. For instance, an automated system must be capable of detecting variable hole positions in the centrifuge after each operation and distinguishing between phases within falcon tubes (e.g., supernatant vs. precipitate). These requirements suggest that, under conditions of high environmental variability, fully automating the washing process necessitates not only advanced visual adaptability but also cognitive capabilities, particularly the ability to reason about and respond intelligently to dynamic and uncertain experimental contexts\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan additionalcitationids=\"CR27 CR28 CR29\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eTo address these challenges, we propose an intelligent system that integrates vision and language AI models to fully automate nanoparticle washing workflows. Our platform leverages the combined capabilities of computer vision, a large language model (LLM), and robotic motion planning to execute experimental procedures with minimal human intervention. Together, these components form a visually and cognitively adaptive framework capable of recognizing experimental setups, dynamically adjusting motion strategies on the basis of real-time visual input, and recommending optimal washing parameters\u0026mdash;such as solvent selection and centrifugation conditions\u0026mdash;by extracting and interpreting contextual chemical knowledge from natural language instructions and reagent information.\u003c/p\u003e\u003cp\u003eAs illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, our automated washing platform begins operation in response to a simple user command (e.g., \u003cem\u003e\u0026lsquo;Please wash my nanoparticles\u0026rsquo; and \u0026lsquo;Nanoparticles synthesized with the following reagents\u0026rsquo;\u003c/em\u003e). Upon receiving such a request, the system autonomously retrieves relevant protocols on the basis of the LLM, plans the necessary sequence of actions, and executes the entire washing procedure\u0026mdash;including solution addition, centrifugation, and supernatant removal. An integrated vision system accurately identifies the location of Falcon tubes, including the synthesized NPs, and evaluates the success of phase separation, whereas the LLM extracts contextual knowledge from the scientific literature to generate on-demand experimental recipes. This entire workflow is implemented as a retrieval-augmented generation (RAG) process, with each iteration incorporating experimental feedback.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e If the initial parameters fail to yield effective separation of the liquid and solid phases in Falcon tubes, the system engages a closed-loop refinement cycle: the experimental results are fed back to the LLM, which then updates its recommendations and proposes a revised protocol. This adaptive loop enables continuous system improvement without requiring further human intervention.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe also demonstrate the implementation and validation of this platform using three representative nanomaterials: NiFe-layered double hydroxide (LDH), IrRu bimetallic NPs, and CdSe/Cds quantum dots (QDs). Comparative evaluations against manual washing and untreated controls show that our system achieves equivalent or superior reproducibility, enhanced removal of residual organics/inorganics, and improved material performance, highlighting the promise of AI-driven automation in next-generation laboratory workflows.\u003c/p\u003e"},{"header":"Results and Discussion","content":"\u003cp\u003e\u003cb\u003eDevelopment of Automated Hardware Platform for Nanoparticle Washing\u003c/b\u003e\u003c/p\u003e\u003cp\u003eVarious techniques have been suggested for washing synthesized NPs, including centrifugation, filtration, dialysis, and magnetic separation. The efficiency of these methods depends largely on factors such as the NP size, surface properties, and aggregation tendencies.\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan additionalcitationids=\"CR22 CR23 CR24\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan additionalcitationids=\"CR34 CR35\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e Among these methods, centrifugation is particularly effective because of its high separation efficiency, which is based on differences in NP size and density. This makes it suitable for applications requiring high-purity NPs, such as catalysts and quantum dots. Based on these advantages, this study employs a centrifuge-based washing method.\u003c/p\u003e\u003cp\u003eTo enhance both efficiency and consistency, we designed a fully automated NP washing system by systematically integrating multiple laboratory automation components. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the front and top views of the system, whereas Supplementary Video S1 shows the operation of each hardware component, facilitating a clearer understanding of the system\u0026rsquo;s functionality. The automated platform comprises a 6-axis robotic arm, a multichannel dispenser, container holder. The robotic arm (Doosan Robotics, M0609; Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) significantly enhances the system\u0026rsquo;s degrees of freedom, enabling a broader range of automated tasks. It is equipped with a tool changer (Supplementary Figure S2(a)) that supports autonomous switching between a 2-finger gripper (OnRobot RG6) and a vacuum gripper (VCG10), enabling secure handling and precise placement of chemical containers, even within the confined space of the centrifuge. Additionally, a webcam mounted on the robotic arm (Supplementary Figure S2(b)) provides real-time visual recognition, enabling the arm to identify containers and dynamically adjust its position. This feature further enhances the accuracy and reliability of the system\u0026rsquo;s operations.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe multidispenser is equipped with a liquid pump (NEXT 50FJ) for removing supernatants and syringe pumps (Cavro XCalibur) for the precise dispersal and extraction of solvents (Supplementary Figure S3). The system comprises linear stages, 3D-printed components, and servo motors and accommodates up to five dispensing tips for the simultaneous handling of multiple solutions. Compared with manual techniques, pump-based solvent removal minimizes the risk of precipitate loss. The dispensing tips are precisely aligned within conical tubes to ensure high positional accuracy during operation. After each use, the tips are rinsed with ethanol to prevent contamination and maintain experimental consistency. Additionally, the actuator is used to laterally transfer the chemical vessels to the dispensing unit, which is located outside the direct reach of the robotic arm (Supplementary Figure S4).\u003c/p\u003e\u003cp\u003eNP washing was performed using a conventional benchtop centrifuge (Hanil Scientific CombiR515, Supplementary Figure S5). To automate lid operation, a custom 3D-printed component was integrated with a wall-mounted linear actuator, providing operational stability. As the centrifuge lacks native support for external communication, we developed a control system utilizing an Arduino microcontroller and a relay module. This setup enables serial communication between the centrifuge and the central control system, allowing it to be integrated into the automated process.\u003c/p\u003e\u003cp\u003eAn ultrasonic bath (BANDELIN SONOREX Digitec; Supplementary Figure S6) was used to disperse the precipitates after the addition of solvent, promoting the dissolution and removal of residual impurities. A 3D-printed holder within the bath allows simultaneous processing of up to six samples, improving throughput. Additional hardware specifications are detailed in the supplementary information (SI).\u003c/p\u003e\u003cp\u003eReliability is a key consideration in the development of automated chemical systems. To evaluate the performance of the liquid handling module, we measured its injection precision. The system consistently dispensed 5 mL of various solvents\u0026mdash;water, ethanol, acetone, and toluene\u0026mdash;with standard deviations below 0.004 mL, as illustrated in Supplementary Figure S9.\u003c/p\u003e\u003cp\u003e\u003cb\u003eVision-based Robot Motion Planning: Centrifugation Task\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe developed a visual feedback system to enable automated experimentation by adaptively recognizing the positions and states of samples within a dynamic centrifuge environment. After each configuration cycle, the position of the rotor inside the centrifuge varies, whereas the robotic arm operates from a fixed, predefined location. This positional inconsistency limits the accuracy of robotic sample placement. Furthermore, due to the bilaterally symmetric design of the rotor, the postcentrifugation arrangement of samples is also symmetric, making it difficult to distinguish between individual sample locations. To overcome this challenge, three color stickers (red, blue, and yellow) were attached inside the rotor. These colors were selected to contrast effectively with the commonly used cap colors of chemical vessels in laboratories. A color recognition algorithm was implemented to automatically assign pick-and-place sequences for each sample, thereby ensuring precise handling and preventing misidentification.\u003c/p\u003e\u003cp\u003eThe system operates through three main stages: (1) perception, (2) robot motion planning, and (3) robot execution. The overall system architecture is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. During the perception phase, object detection and instance segmentation were performed to classify three key elements within the centrifuge: the \u003cem\u003eempty hole, the filled hole, and the rotor center\u003c/em\u003e. For this task, we employed YOLACT, a real-time instance segmentation model known for its balance between speed and accuracy.\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e As shown in Supplementary Figure S10, the model achieved a mean average precision (mAP) of 99.99% at an intersection over union (IoU) threshold of 0.5, indicating highly reliable visual recognition performance.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo evaluate the operational reliability of the system, the centrifuge was run at 1000 rpm for 1 minute, and randomized robotic pick-and-place tasks were subsequently performed. The experiment was repeated 10 times with 6 samples per trial, yielding a 100% task success rate across all repetitions (Supplementary Video S2). These results demonstrate the robustness and reliability of the system, even under visually variable centrifugation conditions induced by centrifugation.\u003c/p\u003e\u003cp\u003e\u003cb\u003eVision-based Robot Motion Planning: Supernatant Removal\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFollowing centrifugation, the removal of the supernatant is a critical step to ensure experimental reproducibility. To automate this process, we developed a vision-guided robotic system that incorporates motion planning. Specifically, the latent mask R-CNN algorithm was employed as the vision model due to its ability to detect and quantify uncertainty. While YOLACT offers the advantage of fast detection for static objects with simple geometries such as the rotor component of a centrifuge, it is less effective in the supernatant removal process, where precipitates often exhibit complex and irregular morphologies. These challenging features limit the applicability of YOLACT in such scenarios.\u003c/p\u003e\u003cp\u003eThe overall workflow for the supernatant removal task is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(a). Before initiating the removal process, the system first determines whether proper precipitation and phase separation have occurred; this step is referred to as the \u0026ldquo;detect the precipitate\u0026rdquo; stage. As shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(b) and (c), cases in which the precipitate and liquid layers are visually distinct are classified as successful separations. Conversely, if only the liquid phase is observed without any discernible precipitate, as shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(d) and 4(e), the system concludes that the centrifugation has failed. When clear separation between the precipitate and supernatant is observed, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(c), the robotic system proceeds with the aspiration of the supernatant.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eHowever, due to distortions caused by the fixed camera angle and the variable orientation of the chemical vessels after centrifugation (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(b)), the vision model may occasionally misclassify a correctly separated sample as a failure. This misclassification arises from two types of uncertainty: epistemic uncertainty, due to insufficient training data covering diverse vessel orientations, and aleatoric uncertainty, resulting from visual distortions caused by the camera perspective and environmental factors. Such errors may necessitate experimental repetition or parameter adjustment, thereby disrupting downstream procedures and reducing overall operational efficiency. To address this challenge, we implemented a feedback mechanism based on uncertainty estimation from the latent mask R-CNN model.\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e When high classification uncertainty is detected as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(a), the model triggers an active feedback loop: the robotic arm rotates the sample by 60\u0026deg; to capture an alternative visual perspective, which is then reanalyzed. This process enhances the robustness and accuracy of sample assessment by enabling a more comprehensive evaluation of the separation state. The feedback mechanism not only improves classification accuracy but also reduces the need for repeated full rotations. By minimizing unnecessary motion, it significantly enhances experimental throughput and mitigates solvent evaporation risks, particularly for volatile washing agents such as acetone, ethanol, and isopropyl alcohol (IPA).\u003c/p\u003e\u003cp\u003eTo evaluate the effectiveness of the latent mask R-CNN, we conducted a comparative study using 45 images, as presented Supplementary Figure S13. Among the 25 images where centrifugation was successful, the latent mask R-CNN correctly identified all cases, while the conventional mask R-CNN\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e misclassified two cases. For samples lacking visible precipitates, both models accurately recognized the absence of separation. However, in the 10 visually ambiguous cases, the Latent Mask R-CNN consistently reported high uncertainty, whereas the mask R-CNN erroneously classified five of them as lacking precipitates. These results demonstrate the superior performance of the latent mask R-CNN in handling complex or ambiguous morphologies, particularly when subtle visual features such as precipitate boundaries must be accurately distinguished.\u003c/p\u003e\u003cp\u003e\u003cb\u003eAdaptive Washing Parameter Recommendation via LLMs with Cognitive Failure Feedback\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a) presents the architecture of the LLM-based washing parameter recommendation system, which includes an integrated cognitive failure feedback loop. The system is designed not only to recommend optimal washing parameters\u0026mdash;such as solvent type and volume, and centrifugation RPM and duration\u0026mdash;based on basic reagent input, but also to adaptively ensure successful precipitation outcomes, particularly when initial attempts fail. This is achieved through an intelligent feedback mechanism that guides experimental modifications in real time. At the core of the system is a Retrieval-Augmented Generation (RAG) framework. It utilizes reagent information provided by the user, typically associated with NP synthesis, to query a user-specific literature database and identify appropriate washing protocols.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eIn Step 1, the system formulates a search query using the input reagents, framed as \"\u003cem\u003enanoparticles synthesized with the following reagents\u003c/em\u003e \".\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIn Step 2, OpenAI\u0026rsquo;s Assistant is used to retrieve the most relevant document from a literature repository in an OpenAI vector store. Relevance is determined via the cosine similarity between embedding vectors, and the document with the highest similarity score is selected for further analysis.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e A key advantage of this system is its scalability: user documents (e.g., scientific papers) can be seamlessly incorporated without requiring retraining or regeneration of embeddings\u0026mdash;an overhead typically encountered in local vector search systems.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIn Step 3, a contextual compression retriever\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e is used to refine the search and extract specific washing protocol information from the selected document.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIn Step 4, the extracted washing parameters are incorporated into a RAG-based prompt (see Supplementary Figure S18), which is then processed by the LLM to generate optimal washing recommendations.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eWhen relevant literature is unavailable, or RAG is not applied, the system defaults to generate washing parameters based solely on the LLM\u0026rsquo;s internal knowledge (Supplementary Figure S19). If the recommended parameters lead to a failed outcome, the system triggers the cognitive adaptive system for failure feedback. This mechanism enables the LLM to suggest an alternative recipe by analyzing the previous attempt and its associated failure mode\u0026mdash;without relying on retraining or external datasets. The corresponding prompt structure for this process is detailed in Supplementary Figure S20.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b) presents a case study involving the washing of IrRu NPs. When CTAB was used as a surfactant during synthesis, the LLM\u0026rsquo;s initial recommendation (water\u0026thinsp;+\u0026thinsp;ethanol) successfully induced NP precipitation. However, under the same washing conditions, replacing CTAB with PVP led to precipitation failure (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c)). In response, the failure feedback loop proposed an alternative solvent combination\u0026mdash;ethanol and acetone\u0026mdash;with lower polarity, which led to successful precipitation. This result is consistent with established chemical principles: the lower dielectric constant of acetone diminishes electrostatic repulsion among NPs, thereby facilitating effective sedimentation.\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e These findings demonstrate the ability of the LLM to incorporate fundamental solvent properties into experimental decision-making, extending its role beyond simple prompt-response applications. Furthermore, the cognitive failure feedback system offers more targeted and efficient optimization of experimental conditions than traditional approaches, such as merely increasing centrifugation speed or time.\u003c/p\u003e\u003cp\u003e\u003cb\u003eValidation of the Automated Washing System\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo evaluate the generality and reliability of the proposed automated washing system, we conducted experiments on three representative classes of nanomaterials, namely, NiFe-based LDH, bimetallic IrRu NPs, and CdSe/CdS quantum dots (QDs), as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. For each material, we compared the effects of three conditions: no washing (baseline), manual washing by domain experts, and an automated system using parameters recommended by an LLM. This comparison allows us to evaluate the impact of washing on the structural, chemical, and functional properties of the nanomaterials.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003eNiFe-Based LDH\u003c/em\u003e\u003c/p\u003e\u003cp\u003eNiFe-based LDH is a prototypical layered structure widely used as a catalyst.\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e,\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e Following synthesis, residual surfactants such as PVP and precursor-derived byproducts often remain on the surface, potentially impairing the catalytic performance of the material. Thus, we investigated how both the presence and method of washing affect the structural integrity and electrochemical characteristics of the material. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(a), we compared the overpotentials for the oxygen evolution reaction (OER) among the three conditions. The unwashed sample exhibited a relatively high overpotential, indicative of poor catalytic efficiency. In contrast, both the manually washed and automatically washed samples showed significantly reduced overpotentials, with no discernible difference between them. This finding demonstrates that the automated system can match the washing effectiveness of a human expert. The observed reduction in overpotential highlights the role of washing in removing surface impurities and electrochemically inactive species, ultimately enhancing the active surface area and improving the catalytic activity.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(b) presents scanning electron microscopy (SEM) images of the washed sample, revealing the characteristic sheet-like morphology of LDH on the surface. This structural feature is absent in the unwashed sample, which suggests that surface contaminants obscure or inhibit the formation of the typical LDH morphology. The results demonstrate that the washing process not only purifies the surface but also facilitates the manifestation of intrinsic nanostructural features.\u003c/p\u003e\u003cp\u003e\u003cem\u003eIrRu NPs\u003c/em\u003e\u003c/p\u003e\u003cp\u003eIrRu alloy catalysts are well recognized for their excellent electrocatalytic performance in the OER.\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(c), both the manually washed and system-washed IrRu NPs exhibited markedly improved OER activity compared with that of the unwashed sample. The unwashed catalyst showed a significantly higher overpotential and lower current density, indicative of surface contamination by insulating organic residues. These contaminants, likely residual surfactants and synthesis byproducts, hinder effective charge transfer and catalytic activity. In contrast, the catalysts washed either by a human expert or through the automated system\u0026mdash;both using LLM-recommended parameters\u0026mdash;demonstrated comparable and enhanced electrocatalytic activity. This result highlights the effectiveness of the LLM-guided automated washing protocol in replicating expert-level outcomes, confirming its reliability in the preparation of clean and active catalyst surfaces.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(d) presents transmission electron microscopy (TEM) images of the IrRu NP catalysts before and after washing. In the unwashed sample, densely packed surface residues are clearly observed, indicating the presence of unremoved organic species. After washing, the NP surfaces appeared significantly cleaner, with a more uniform dispersion of individual particles. This morphological improvement suggests the successful elimination of organic impurities, further supporting the observed increase in electrochemical performance.\u003c/p\u003e\u003cp\u003e\u003cem\u003eCdSe/CdS QDs\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFor CdSe/CdS QDs, achieving uniform particle size is essential, as their optical and electronic properties are highly sensitive to size distribution.\u003csup\u003e\u003cspan additionalcitationids=\"CR48\" citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e Accordingly, purification is one of the most critical steps in the synthesis workflow, as it directly influences the homogeneity, purity, and reproducibility of the final product. The effectiveness of this purification process is strongly governed by the washing parameters applied. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(e) shows the photoluminescence (PL) spectra of CdSe/CdS QDs subjected to three different washing conditions. The unwashed sample exhibited two distinct emission peaks at 535 nm and 605 nm, indicating the presence of two distinct size populations within the QD ensemble. This result confirms that, without purification, the synthesized QDs consisted of a heterogeneous mixture of particle sizes. Upon applying the initial washing protocol generated by the LLM, the 535 nm peak was substantially reduced, suggesting partial removal of the smaller QDs. However, the persistence of this peak indicated incomplete purification. Through an automated failure feedback loop, the system subsequently refined the washing parameters, ultimately eliminating the 535 nm peak entirely. This indicates the successful removal of size heterogeneity and highlights the efficacy of iterative protocol optimization via feedback. Notably, the PL spectrum of the system-washed sample closely mirrored that of the manually washed QDs, validating the automated system\u0026rsquo;s ability to match expert-level purification performance. This result illustrates the key capability of the autonomous experimental platform: it can not only induce physical separation processes such as selective precipitation, but also interpret spectral features to refine protocols based on material-specific feedback. As shown in Supplementary Figure S25, the washing recipe generated through the failure feedback loop demonstrates the system\u0026rsquo;s capacity to detect suboptimal outcomes and autonomously redesign experiments for improved results\u0026mdash;exemplifying property-aware optimization in nanomaterial synthesis.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(f) shows low-magnification TEM images of the CdSe/CdS QD films before and after washing. In the unwashed sample, numerous dark patches and aggregated domains are evident, indicative of residual surfactants and poor particle dispersion. After washing, the film displays a notably cleaner surface with reduced contamination and improved uniformity. High-resolution inset images further confirm the presence of well-dispersed, size-uniform QDs, underscoring the effectiveness of the purification protocol employed by the LLM-guided method.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn this study, we developed an adaptive automation system for centrifuge-based NP washing by integrating computer vision and an LLM. The vision module of this system employs YOLACT for real-time detection and localization of centrifuge tubes. After centrifugation, the latent mask R-CNN is used to classify the supernatant and precipitate regions and to quantify the segmentation uncertainty. When the segmentation uncertainty is high, the system initiates a visual feedback loop by reorienting the sample to reduce ambiguity and improve the classification accuracy. To support decision-making, the system incorporates the GPT-4o-mini LLM in combination with a RAG pipeline to suggest initial centrifugation parameters. These parameters are iteratively refined through experimental feedback, particularly in response to failure cases identified during operation. This system was validated across three representative nanomaterial classes: NiFe-LDHs, IrRu NPs, and CdSe/CdS QDs. The results demonstrate that the integration of perception (via computer vision) and cognition (via LLM reasoning) enables a transition from rigid, rule-based procedures to adaptive, intelligent experimental workflows. By facilitating robust operation under uncertainty and enabling data-driven refinement of protocols, our system paves the way for intelligent, reproducible, and scalable experimental design. The success of this system highlights the broader potential of autonomous workflows in both material synthesis and characterization.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cb\u003eTraining datasets for computer vision models\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003eYOLACT for the Centrifugation Task\u003c/em\u003e\u003c/p\u003e\u003cp\u003eA YOLACT instance segmentation model was developed to detect key components within the centrifugation environment, including the rotors, Falcon tubes, and empty holes. The training dataset comprised 295 original images, which were augmented using color transformations including red and purple overlays, resulting in a total of 885 images. Model training was based on a customized yolact_base_config\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e derived from the original COCO\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e configuration, featuring a ResNet-101 backbone, an input image size of 550×550 pixels, and a maximum of 1,000 iterations. The model was trained on an NVIDIA RTX A4000 GPU, with the dataset split into 80% for training and 20% for validation. Additional training parameters, including learning rate steps and IOU thresholds, followed the default base configuration.\u003c/p\u003e\u003cp\u003e\u003cem\u003eLatent Mask R-CNN for removing the supernatant\u003c/em\u003e\u003c/p\u003e\u003cp\u003eTo classify and detect the supernatant removed after centrifugation, a latent mask R-CNN model was trained using 1,926 manually annotated experimental images. The model utilized the mask_rcnn_R_50_FPN_1x architecture from the Detectron2\u003csup\u003e51\u003c/sup\u003e framework. Training was conducted with standard hyperparameters, including an initial learning rate of 1x10\u003csup\u003e-4\u003c/sup\u003e and a batch size of 4, on an NVIDIA RTX 3060 Ti GPU. The dataset was divided to achieve a training-to-validation ratio of 80:20. The incorporation of latent representations enhanced the model's generalization capability across diverse experimental variations, including differences in supernatant appearance, precipitate shape, sample holder, and Falcon tube positioning.\u003c/p\u003e\u003cp\u003e\u003cb\u003eRobot motion planning for the centrifugation task\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo enable precise robotic manipulation within the centrifuge, key objects, including the chemical vessels, rotor holes, and rotor center, were first detected from the input image and used to generate processed data suitable for motion planning. The image coordinates of these objects were then computed. Because the robot arm observes the centrifuge interior from a fixed position, the rotor center serves as a consistent reference point. This reference was mapped to the robot’s coordinate system, enabling the calculation of relative positions for the chemical vessels and holes. To determine the correct processing sequence, color-based object recognition was employed using distinct stickers (red, yellow, and blue) affixed inside the rotor. These stickers were carefully selected to avoid color overlap with Falcon tube caps. For example, if a Falcon tube with a red cap is used, blue or yellow stickers are selected to prevent recognition errors. Once the center coordinates of the color-coded stickers were identified, the objects were sorted in a clockwise order based on their angular positions relative to the reference sticker. The angle 𝜃 between the reference sticker and each object was computed using the following equations:\u003c/p\u003e\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\theta\\:=\\text{atan}2({y}_{color}-{y}_{object},{x}_{color}-{x}_{object})$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{\\theta\\:}_{deg}=\\theta\\:\\:\\times\\:\\left(\\frac{180}{{\\pi\\:}}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIf the calculated angle was negative, 360° was added to convert it to a positive value. The angular difference, Δ𝜃, between the reference sticker and each object was then calculated as follows:\u003c/p\u003e\u003cp\u003eΔ𝜃 = \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\theta\\:}_{color}-{\\theta\\:}_{object}\\)\u003c/span\u003e\u003c/span\u003e(3)\u003c/p\u003e\u003cp\u003eThis normalized angular difference allows for consistent sorting of objects in a clockwise sequence. Once the object sequence is determined, the positional data of each object are transformed into the robot’s coordinate system. By using the rotor center previously mapped in the robot’s coordinate space as a fixed anchor, the relative image-based coordinates of each object were added to the robot center’s robot-frame coordinates (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{center\\:of\\:rotor},\\:\\:{y}_{center\\:of\\:rotor})\\)\u003c/span\u003e\u003c/span\u003e. This transformation enables the robot to accurately position and manipulate each sample.\u003c/p\u003e\u003cp\u003e\u003cb\u003eUncertainty quantification using the latent mask R-CNN\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe utilized latent mask R-CNN to mitigate image distortions encountered during the supernatant removal process and to improve the segmentation accuracy between the liquid (supernatant) and precipitate regions. Unlike traditional deterministic segmentation methods, latent mask R-CNN incorporates a latent variable \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{z}\\)\u003c/span\u003e\u003c/span\u003e into the mask prediction process, enabling a probabilistic framework. This allows the model to generate multiple plausible segmentation masks, thereby capturing the inherent distributional uncertainty present in the data. To quantify segmentation uncertainty, we sampled \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{n}\\)\u003c/span\u003e\u003c/span\u003e masks from the latent variable \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{z}\\)\u003c/span\u003e\u003c/span\u003e and computed the pixelwise mean \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mu\\:}\\)\u003c/span\u003e\u003c/span\u003e and variance \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{{\\sigma\\:}}^{2}\\)\u003c/span\u003e\u003c/span\u003e. The coefficient of variation (CV) for each pixel is defined as follows:\u003c/p\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:\\text{C}\\text{V}\\:=\\frac{{\\sigma\\:}}{{\\mu\\:}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIt was then computed separately for the supernatant and precipitate classes. Higher CV values correspond to regions of greater segmentation uncertainty, indicating lower confidence in the predicted mask boundaries. This uncertainty-aware approach facilitates more reliable downstream decision-making in automated experimental workflows.\u003c/p\u003e\u003cp\u003e\u003cb\u003eRecipe generator via LLM\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo generate optimal washing conditions, we used the GPT-4o-mini model from OpenAI within the LangChain framework. The model was configured with a temperature setting of 0 to ensure deterministic and consistent outputs. Input prompts were constructed from structured experimental metadata, including reagent names, volumes, concentrations, relevant literature on washing protocols (retrieved text through the retrieval process), and records of prior experimental failures when available. To tailor prompt generation based on the available context, we used different prompt augmentation strategies: RAG-based augmentation was applied when relevant literature could be retrieved; failure-based augmentation was used when a previous washing attempt had failed; and metadata-only prompts were used in the absence of both literature and failure history. The specific prompt templates used for each case are provided in Supplementary Figures S18 ~ S20. Model outputs were returned in JSON format and processed programmatically following a preprocessing step to ensure correct formatting and interpretability. In cases of failure, a feedback loop was implemented: the failure conditions were incorporated into the next prompt, allowing the model to refine its recommendations iteratively. This iterative framework enables continuous improvement of washing protocols without requiring additional fine-tuning of the model or repeated manual intervention.\u003c/p\u003e\u003cp\u003e\u003cb\u003eImplementation of the retriever system\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo build the retrieval system, we utilized OpenAI’s vector store to create the embedding database. PDF documents were uploaded and automatically embedded via the OpenAI API (Supplementary Figure S16). For efficient and context-preserving retrieval, the documents were segmented into overlapping chunks—each consisting of 800 tokens with a 400-token overlap. The retrieval process consisted of two stages:\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eInitial Retrieval: Using a list of synthesized reagents as input, the GPT Assistant API queried the embedded document database in the OpenAI vector store to identify literature relevant to the synthesis and washing process.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eContextual Compression: In the second stage, we applied ContextualCompressionRetriever to filter the initial results and extract only the segments specifically related to the washing protocols. This component integrated a base retriever with an LLM-based compressor, LLMChainExtractor, powered by the GPT-4o-mini model. The model was executed with a temperature setting of 0 to maintain consistency and determinism.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThis two-tiered retrieval system enabled accurate identification of procedural content, which was subsequently used to inform and enhance the generation of washing condition recommendations. By isolating contextually relevant information from broader scientific literature, the retriever significantly improved the precision of downstream prompt engineering for the recipe generation module.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData and code availability\u003c/h2\u003e\n\u003cp\u003eSeveral examples of our result data, the codes, and related explanations are provided in the following GitHub repository (https://github.com/KIST-CSRC/washingModule). All codes are written in Python 3.9 and all environments could be created via requirements.txt file.\u003c/p\u003e\n\u003ch2\u003eLead contact\u003c/h2\u003e\n\u003cp\u003eFurther information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Sang Soo Han ([email protected]).\u003c/p\u003e\n\u003ch2\u003eAcknowledgements\u003c/h2\u003e\n\u003cp\u003eThis work was supported by the National Research Foundation of Korea funded by the Ministry of Science and ICT [NRF-2022M3H4A7046278 and RS-2024-00450102].\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003cbr\u003e\u003cstrong\u003eS.S.H. and H.S.L.(Heeseung Lee)\u003c/strong\u003e conceived the idea. S.S.H. and S.S.S. supervised the project. \u003cstrong\u003eH.S.L., D.K, N.K, and H.J.Y\u003c/strong\u003e developed the hardware, and H.S.L. integrated vision and language AI into the system. \u003cstrong\u003eH.L.(Heyin Lee) and T.Y.\u003c/strong\u003e synthesized the LDH and IrRu samples and conducted their electrochemical measurements. N.Y.K. and N.O. synthesized the CdSe/Cds quantum dot samples and evaluated their optical properties. All authors contributed to the analysis of the results and the writing of the manuscript.\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eWe have a patent pending titled \u0026ldquo;System and method for automated nanoparticle washing\u0026rdquo; (Korean Patent Application No. 10-2024-0170078).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eShi Y, Prieto PL, Zepel T, Grunert S, Hein JE (2021) Automated Experimentation Powers Data Science in Chemistry. Acc Chem Res 54:546\u0026ndash;555. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1021/acs.accounts.0c00736\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1021/acs.accounts.0c00736\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDu H, Corkan LA, Yang K, Kuo PY, Lindsey JS (1999) An automated microscale chemistry workstation capable of parallel, adaptive experimentation. Chemometr Intell Lab Syst 48:181\u0026ndash;203\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKing RD et al (2009) The Automation of Science. Science 324:85\u0026ndash;89. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1126/science.1165620\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1126/science.1165620\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelekman JA et al (2017) High-throughput automation in chemical process development. Annual Rev Chem Biomol Eng 8:525\u0026ndash;547\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGromski PS, Henson AB, Granda JM, Cronin L (2019) How to explore chemical space using algorithms and automation. Nat Reviews Chem 3:119\u0026ndash;128\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJiang Y et al (2022) An artificial intelligence enabled chemical synthesis robot for exploration and optimization of nanomaterials. Sci Adv 8:eabo2626. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1126/sciadv.abo2626\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1126/sciadv.abo2626\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYoo HJ et al (2024) Bespoke Metal Nanoparticle Synthesis at Room Temperature and Discovery of Chemical Knowledge on Nanoparticle Growth via Autonomous Experimentations. Adv Funct Mater 34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1002/adfm.202312561\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1002/adfm.202312561\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBurger B et al (2020) A mobile robotic chemist. Nature 583:237\u0026ndash;. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41586-020-2442-2\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41586-020-2442-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGranda JM, Donina L, Dragone V, Long DL, Cronin L (2018) Controlling an organic synthesis robot with machine learning to search for new reactivity. Nature 559:377\u0026ndash;381. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s41586-018-0307-8\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s41586-018-0307-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRapp JT, Bremer BJ, Romero PA (2024) Self-driving laboratories to autonomously navigate the protein fitness landscape. Nat Chem Eng 1:97\u0026ndash;107. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1038/s44286-023-00002-4\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1038/s44286-023-00002-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTao HC et al (2021) Self-Driving Platform for Metal Nanoparticle Synthesis: Combining Microfluidics and Machine Learning. \u003cem\u003eAdv Funct Mater\u003c/em\u003e 31 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 2106725. 10.1002/adfm.202106725\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYoo HJ, Lee KY, Kim D, Han SS (2024) OCTOPUS: operation control system for task optimization and job parallelization via a user-optimal scheduler. \u003cem\u003eNat Commun\u003c/em\u003e 15 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 9669.10.1038/s41467-024-54067-7\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoch LM et al (2020) ChemOS: An orchestration software to democratize autonomous discovery. PLoS ONE 15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN e0229862\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN e0229862\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0229862\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0229862\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDu Toit NE, Burdick JW (2011) Robot motion planning in dynamic, uncertain environments. IEEE Trans Robot 28:101\u0026ndash;115\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTaherimakhsousi N et al (2021) A machine vision tool for facilitating the optimization of large-area perovskite photovoltaics. \u003cem\u003eNpj Computational Materials\u003c/em\u003e 7 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 190.10.1038/s41524-021-00657-8\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDarvish K et al (2025) ORGANA: A robotic assistant for automated chemistry experimentation and characterization. \u003cem\u003eMatter-Us\u003c/em\u003e 8 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 101897. 10.1016/j.matt.2024.10.015\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLisca G, Nyga D, B\u0026aacute;lint-Bencz\u0026eacute;di F, Langer H, Beetz M in \u003cem\u003e2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).\u003c/em\u003e 5202\u0026ndash;5208 (IEEE)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEppel S, Xu H, Bismuth M, Aspuru-Guzik A (2020) Computer Vision for Recognition of Materials and Vessels in Chemistry Lab Settings and the Vector-LabPics Data Set. ACS Cent Sci 6:1743\u0026ndash;1752. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1021/acscentsci.0c00460\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1021/acscentsci.0c00460\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChiba T et al (2017) High-Efficiency Perovskite Quantum-Dot Light-Emitting Devices by Effective Washing Process and Interfacial Energy Level Alignment. ACS Appl Mater Interfaces 9:18054\u0026ndash;18060. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1021/acsami.7b03382\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1021/acsami.7b03382\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNyongombe G, Kabongo GL, Noto LL, Dhlamini MS (2022) Investigating the Impact of the Washing Steps of Layered Double Hydroxides (LDH) on the Electrochemical Performance. \u003cem\u003eNanomaterials\u003c/em\u003e 12, 578\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAdkins H, Billica HR (1948) The preparation of Raney nickel catalysts and their use under conditions comparable with those for platinum and palladium catalysts. J Am Chem Soc 70:695\u0026ndash;698\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHoshi K et al (2018) Purification of perovskite quantum dots using low-dielectric-constant washing solvent diglyme for highly efficient light-emitting devices. ACS Appl Mater Interfaces 10:24607\u0026ndash;24612\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou S (2021) Rapid separation and purification of lead halide perovskite quantum dots through differential centrifugation in nonpolar solvent. RSC Adv 11:28410\u0026ndash;28419\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang J et al (2022) Surface enrichment of Ir on the IrRu alloy for efficient and stable water oxidation catalysis in acid. Chem Sci 13:12114\u0026ndash;12121\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi D et al (2012) Surfactant removal for colloidal nanoparticles from solution synthesis: the effect on catalytic performance. ACS Catal 2:1358\u0026ndash;1362\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEl-Khawaldeh R et al (2024) Keeping an eye on the experiment: computer vision for real-time monitoring and control. Chem Sci 15:1271\u0026ndash;1282. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1039/d3sc05491h\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1039/d3sc05491h\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTiong LCO et al (2024) Machine vision-based detections of transparent chemical vessels toward the safe automation of material synthesis. npj Comput Mater 10:42\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKendall A, Gal Y (2017) What uncertainties do we need in bayesian deep learning for computer vision? Adv Neural Inf Process Syst 30\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXia Y, Zhang J, Jazdi N, Weyrich M (2024) Incorporating Large Language Models into Production Systems for Enhanced Task Automation and Flexibility. \u003cem\u003earXiv preprint arXiv:2407.08550\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXia Y, Jazdi N, Zhang J, Shah C, Weyrich M (2024) Control Industrial Automation System with Large Language Models. \u003cem\u003earXiv preprint arXiv:2409.18009\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTang Y, Yang Y (2024) Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries. \u003cem\u003earXiv preprint arXiv:2401.15391\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLewis P et al (2020) Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Adv Neur In 33\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei L et al (2008) Effect of centrifugation on the purity of single-walled carbon nanotubes from MCM-41 containing cobalt. J Phys Chem C 112:17567\u0026ndash;17575\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang Y, Li D, Tian XA, Highly Efficient (2024) Fluorescent Turn-Off Nanosensor for Quantitative Detection of Teicoplanin Antibiotic from Humans, Food, and Water Based on the Electron Transfer between Imprinted Quantum Dots and the Five-Membered Cyclic Boronate Esters. Molecules 29:4115\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeeranoppanant N, Adamo A (2020) In-Line Purification: A Key Component to Facilitate Drug Synthesis and Process Development in Medicinal Chemistry. ACS Med Chem Lett 11:9\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1021/acsmedchemlett.9b00491\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1021/acsmedchemlett.9b00491\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLordi V, Yao N, Wei J (2001) Method for supporting platinum on single-walled carbon nanotubes for a selective hydrogenation catalyst. Chem Mater 13:733\u0026ndash;737\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBolya D, Zhou C, Xiao FY, Lee YJ (2019) YOLACT Real-time Instance Segmentation. Ieee I Conf Comp Vis 9156\u0026ndash;9165. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1109/Iccv.2019.00925\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1109/Iccv.2019.00925\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu YX, Mishra N, Abbeel P, Chen X (2023) Distributional Instance Segmentation: Modeling Uncertainty and High Confidence Predictions with Latent-MaskRCNN. Ieee Int Conf Robot 7069\u0026ndash;7075. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1109/Icra48891.2023.10160812\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1109/Icra48891.2023.10160812\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe K, Gkioxari G, Dollar P, Girshick R, Mask R-CNN (2020) IEEE Trans Pattern Anal Mach Intell 42:386\u0026ndash;397. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1109/TPAMI.2018.2844175\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1109/TPAMI.2018.2844175\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOpenAI (2025) \u003cem\u003eOpenAI assistant\u003c/em\u003e, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://platform.openai.com/docs/assistants/tools\u003c/span\u003e\u003cspan address=\"https://platform.openai.com/docs/assistants/tools\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVerma S (2024) Contextual Compression in Retrieval-Augmented Generation for Large Language Models: A Survey. \u003cem\u003earXiv preprint arXiv:2409.13385\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVan der Hoeven PC, Lyklema J (1992) Electrostatic stabilization in non-aqueous media. Adv Colloid Interface Sci 42:205\u0026ndash;277\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGu Y et al (2024) NiFe layered double hydroxides synthesized based on solvent properties as anode catalysts for enhanced oxygen evolution reaction. \u003cem\u003eChem Eng J\u003c/em\u003e 480 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 147789. 10.1016/j.cej.2023.147789\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHou CM et al (2021) Rapid large-scale synthesis of ultrathin NiFe-layered double hydroxide nanosheets with tunable structures as robust oxygen evolution electrocatalysts. RSC Adv 11:37624\u0026ndash;37630. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1039/d1ra05045a\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1039/d1ra05045a\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark Y et al (2025) Atomic-level Ru-Ir mixing in rutile-type (RuIr)O for efficient and durable oxygen evolution catalysis. Nat Commun 16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-025-55910-1\u003c/span\u003e\u003cspan address=\"10.1038/s41467-025-55910-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBornet A et al (2024) Beyond RDE characterisation - Unveiling IrRu/ATO OER catalyst stability with a GDE setup. Electrochim Acta 501. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.electacta.2024.144773\u003c/span\u003e\u003cspan address=\"10.1016/j.electacta.2024.144773\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. https://doi.org:ARTN 144773\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu YC, Bose S, Fan WJ (2018) Effect of size and shape on electronic and optical properties of CdSe quantum dots. Optik 155:242\u0026ndash;250. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1016/j.ijleo.2017.10.165\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1016/j.ijleo.2017.10.165\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSamadi-Maybodi A, Tirbandpay R (2021) Synthesis, optical properties and tuning size of CdSe quantum dots by variation capping agent. \u003cem\u003eSpectrochim Acta A\u003c/em\u003e 250 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:ARTN\u003c/span\u003e\u003cspan address=\"https://doi.org:ARTN\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 119369.10.1016/j.saa.2020.119369\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGwak N et al (2025) Comparative Study of Two Different Quasi-Type II Heterostructured Quantum Dots for Enhanced Electroluminescence. Acs Photonics 12:944\u0026ndash;951. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:10.1021/acsphotonics.4c01995\u003c/span\u003e\u003cspan address=\"https://doi.org:10.1021/acsphotonics.4c01995\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLin TY et al (2014) Microsoft COCO: Common Objects in Context. Lect Notes Comput Sc 8693:740\u0026ndash;755. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org:Doi 10.1007/978-3-319-10602-1_48\u003c/span\u003e\u003cspan address=\"https://doi.org:Doi 10.1007/978-3-319-10602-1_48\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGirshick W-Y (2019) L. a. R. \u003cem\u003eDetectron2\u003c/em\u003e, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/facebookresearch/detectron2\u003c/span\u003e\u003cspan address=\"https://github.com/facebookresearch/detectron2\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7060706/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7060706/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSelf-driving laboratories are reshaping materials discovery by combining automated experimentation with AI-driven decision-making. However, the lack of automation in key preprocessing steps such as nanoparticle (NP) washing\u0026mdash;remains a major barrier to achieving full experimental autonomy. Effective automation of NP washing requires visual adaptivity, to detect subtle changes in the appearance of dispersions or precipitates, and cognitive adaptivity, to handle failure cases like incomplete sedimentation or phase separation. These demands make NP washing uniquely challenging despite its apparent simplicity. We introduce a fully integrated NP washing platform that combines computer vision with a large language model (LLM) to enable intelligent, end-to-end preprocessing in self-driving labs. The system employs YOLACT for real-time robotic manipulation and latent mask R-CNN for uncertainty-aware image segmentation, achieving 100% task success across 60 trials and accurately processing 45 diverse precipitate images. A retrieval-augmented LLM autonomously generates and continuously refines washing protocols based on cognitive feedback from failure detection. The platform was validated on NiFe layered double hydroxides, IrRu nanoparticles, and CdSe/CdS quantum dots. Electrochemical and photoluminescence analyses demonstrated that the automated washing matches or exceeds the quality of expert-level manual processing. This work represents a critical step toward fully autonomous, closed-loop experimentation by bridging synthesis and characterization through intelligent preprocessing.\u003c/p\u003e","manuscriptTitle":"You Only Put Your Nanoparticle: A Fully Automated System for Nanoparticle Washing Enabled by Vision and Language AI","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-29 09:44:31","doi":"10.21203/rs.3.rs-7060706/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3d1ff4a9-2818-45ca-a3d2-4e10d2ca3929","owner":[],"postedDate":"July 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":52120819,"name":"Physical sciences/Materials science/Techniques and instrumentation/Design, synthesis and processing"},{"id":52120820,"name":"Physical sciences/Materials science/Techniques and instrumentation/Imaging techniques"}],"tags":[],"updatedAt":"2025-08-28T11:15:41+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-29 09:44:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7060706","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7060706","identity":"rs-7060706","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}