# Fileset

[Gracheva_2026_Mach._Learn.__Sci._Technol._7_045006.pdf](https://mdr.nims.go.jp/filesets/345a8416-8bbb-4fbc-a76f-75daff30f77b/download)

## Creator

[Ekaterina Gracheva](https://orcid.org/0000-0002-9704-5939), Shin-ichi Yamazaki, [Guillaume Lambard](https://orcid.org/0000-0003-0275-4079), [Keitaro Sodeyama](https://orcid.org/0000-0002-9228-0729), Tsutomu Ioroi, [Masafumi Asahi](https://orcid.org/0000-0002-2122-3073)

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Enhancing oxygen reduction reaction mass activity in polymer electrolyte fuel cells via molecular machine learning](https://mdr.nims.go.jp/datasets/55a97262-d049-4c9d-ba25-dd35fe09230c)

## Fulltext

Enhancing oxygen reduction reaction mass activity in polymer electrolyte fuel cells via molecular machine learning            PAPER • OPEN ACCESSEnhancing oxygen reduction reaction mass activityin polymer electrolyte fuel cells via molecularmachine learningTo cite this article: Ekaterina Gracheva et al 2026 Mach. Learn.: Sci. Technol. 7 045006 View the article online for updates and enhancements.You may also likeTransformer-based pulse shapediscrimination in hpge detectors withmasked autoencoder pre-trainingMarta Babicz, Saúl Alonso-Monsalve,Alain Fauquex et al.-Real-time classification of petawatt laserbeam profilesVlad Gaciu, Ioan Dncu, Mihai Caragea etal.-Conformalized-KANs: uncertaintyquantification with coverage guaranteesfor Kolmogorov–Arnold Networks (KANs)in scientific machine learningAmirhossein Mollaali, Christian BolivarMoya, Amanda A Howard et al.-This content was downloaded from IP address 144.213.253.16 on 26/08/2026 at 07:13https://doi.org/10.1088/2632-2153/ae7f7chttps://iopscience.iop.org/article/10.1088/2632-2153/ae7ec9https://iopscience.iop.org/article/10.1088/2632-2153/ae7ec9https://iopscience.iop.org/article/10.1088/2632-2153/ae7ec9https://iopscience.iop.org/article/10.1088/2632-2153/ae7e3chttps://iopscience.iop.org/article/10.1088/2632-2153/ae7e3chttps://iopscience.iop.org/article/10.1088/2632-2153/ae7495https://iopscience.iop.org/article/10.1088/2632-2153/ae7495https://iopscience.iop.org/article/10.1088/2632-2153/ae7495https://iopscience.iop.org/article/10.1088/2632-2153/ae7495Mach. Learn.: Sci. Technol. 7 (2026) 045006 https://doi.org/10.1088/2632-2153/ae7f7cOPEN ACCESSRECEIVED9 February 2026REVISED16 June 2026ACCEPTED FOR PUBLICATION18 June 2026PUBLISHED3 July 2026Original content fromthis work may be usedunder the terms of theCreative CommonsAttribution 4.0 licence.Any further distributionof this work mustmaintain attribution tothe author(s) and the titleof the work, journalcitation and DOI.PAPEREnhancing oxygen reduction reaction mass activity in polymerelectrolyte fuel cells via molecular machine learningEkaterina Gracheva1, Shin-ichi Yamazaki2,∗, Guillaume Lambard1,∗, Keitaro Sodeyama1,Tsutomu Ioroi2 and Masafumi Asahi21 Data-driven Materials Design group, Center for Basic Research on Materials (CBRM), National Institute for Materials Science(NIMS), 1-1 Namiki, Tsukuba 305-0044, Ibaraki, Japan2 Research Institute of Electrochemical Energy, Department of Energy and Environment, National Institute of Advanced IndustrialScience and Technology (AIST), 1-8-31 Midorigaoka, Ikeda 562-8577, Osaka, Japan∗ Authors to whom any correspondence should be addressed.E-mail: s-yamazaki@aist.go.jp, lambard.guillaume@nims.go.jp, gracheva.ekaterina@nims.go.jp, sodeyama.keitaro@nims.go.jp,ioroi-t@aist.go.jp andm.asahi@aist.go.jpKeywords:materials science, machine learning, chemoinformatics, polymer electrolyte fuel cell (PEFC), oxygen reduction reaction,active learningabstractThe search for efficient fuel cell additives is a critical challenge for improving their oxygen reduc-tion reaction activity. Here, we present the first successful application of active learning (AL) toguide the experimental discovery of organic molecules that enhance the mass activity (MA) ofPt cathode catalysts in polymer electrolyte fuel cells (PEFCs), a class of fuel cells often referred toin the literature as proton-exchange membrane fuel cells. Using our in-house simplified molec-ular input line entry system-X molecular characterization tool, we trained neural networks,autonomously tailored for our data, on an initial set of 96 organic compounds, of which 84 passedRDKit validity checks for modeling. Model performance improved over three AL cycles, withthe mean absolute error decreasing from 0.202 to 0.165 and the root mean square error decreas-ing from 0.271 to 0.218. The cycles were guided by predictive ranking, experimental feasibilityconstraints, and LSTM-based molecule generation. Among the additives identified, melam, amelamine dimer, achieved a 67% improvement in MA over the no-additive baseline. These resultsdemonstrate the feasibility of AL in low-data regimes for PEFC additive discovery and highlightthe importance of data consistency, model retraining, and close machine learning–experimentcoordination.1. IntroductionAs the global community accelerates efforts to achieve carbon neutrality, hydrogen-based energy systemsare gaining significant attention due to their potential to provide clean and efficient power. Polymerelectrolyte fuel cells (PEFCs) are a leading clean energy technology due to their high efficiency, lowoperating temperature, and potential for zero emissions [1–4]. PEFCs are being widely studied forapplications ranging from automotive transportation to portable power systems and stationary energysources. Despite extensive research, several challenges remain before PEFCs can be deployed at a largescale, including issues related to durability, cost, and catalyst performance [5–12].One key performance metric in PEFCs is the oxygen reduction reaction (ORR) activity at the cath-ode. Despite decades of study, the ORR on Pt under operating conditions remains the rate-limiting steprelative to the hydrogen oxidation at the anode, constraining overall PEFCs efficiency. Modifying thelocal chemical environment at Pt, via adsorbed organic additives, is a promising route to enhance ORRkinetics.Prior works have explored various classes of organic molecules, such as ionic liquid [13–18],oleylamine [19], pyrene amine with octylamine [20], tetraazaporphyrin [21, 22], melamine (polymer)© 2026 The Author(s). Published by IOP Publishing Ltdhttps://doi.org/10.1088/2632-2153/ae7f7chttps://crossmark.crossref.org/dialog/?doi=10.1088/2632-2153/ae7f7c&domain=pdf&date_stamp=2026-7-3https://creativecommons.org/licenses/by/4.0/https://creativecommons.org/licenses/by/4.0/https://orcid.org/0000-0002-9704-5939https://orcid.org/0000-0003-0275-4079https://orcid.org/0000-0002-2122-3073mailto:s-yamazaki@aist.go.jpmailto:lambard.guillaume@nims.go.jpmailto:gracheva.ekaterina@nims.go.jpmailto:sodeyama.keitaro@nims.go.jpmailto:ioroi-t@aist.go.jpmailto:m.asahi@aist.go.jpMach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et al[23–26], and guanamine [27] as activity-enhancing additives for electrocatalysis in PEFCs. These stud-ies have demonstrated that modification of Pt surface with appropriate molecules significantly enhancesthe ORR activity, and that changes in molecular structure can significantly impact interfacial chemistryand catalytic performance. However, the selection process in these works has typically relied on expertjudgment or limited chemical libraries, without systematic data-driven exploration of the broader chem-ical space. Although the underlying mechanisms are not fully understood, these additives are thought tomodulate the interfacial chemistry, alter adsorption behavior, or destabilize reactive intermediates, ulti-mately improving catalytic activity [28].Selecting the right additives is, however, a complex problem. The chemical space of potential organicmolecules is vast, and traditional trial-and-error experimental screening is both time-consuming andcostly. In many cases, materials scientists rely on chemical intuition or structural similarity to knownadditives to select candidates. This approach, while practical, can easily overlook unconventional orstructurally diverse compounds that might exhibit strong performance. As a result, the exploration ofadditive space remains incomplete, and more efficient, systematic methods are needed to accelerate dis-covery and reduce experimental overhead.To the best of our knowledge, no prior study has applied a closed-loop active-learning frameworkto systematically discover organic additives that enhance ORR mass activity (MA) in PEFCs. Such anapproach enables efficient navigation of the additive space and offers a foundation for systematic dis-covery of high-performing, structurally novel candidates.Molecular machine learning (MML) has emerged as a powerful tool to assist in materials discov-ery and design, particularly in domains where large datasets are unavailable and each experiment isexpensive [29]. In recent years, MML has been successfully applied to many tasks, including catalystoptimization [30, 31], polymer design [32–35], and molecular generation for pharmaceuticals [36, 37].However, its adoption in PEFC additive screening has been limited [38, 39]. This is partly due to thecomplexity of the problem, molecular structure-property relationships are often non-linear and requiremeaningful molecular representations, and partly due to the lack of high-quality experimental data.To address the challenge of identifying effective organic additives that can enhance the ORR activityof Pt catalysts, we employ an active learning (AL) framework. In AL, a machine learning (ML) model isused not only to learn from existing data but also to guide the selection of new experiments [40–42].It is an iterative process that aims to improve the model’s accuracy while minimising the number ofrequired experiments. At each step, the model selects the most informative or promising candidates totest next, based on an acquisition strategy such as uncertainty or expected improvement. This approachhas shown promise in chemistry and materials science, particularly in small-data scenarios where effi-cient use of experimental resources is critical [43–46].In the present study, we combine prediction-based candidate ranking, molecular generation, andexperimental feasibility filtering within a closed-loop active-learning workflow. This strategy allows thesearch to focus on candidate additives that are not only predicted to improve ORR MA, but are alsorealistic for experimental validation.With this work, we aimed to demonstrate the applicability of AL to PEFC additive discovery. Ourprimary objective was to improve MA by optimizing the choice of organic additives adsorbed on the Ptsurface. We employed the simplified molecular input line entry system (SMILES-X) framework, usingLSTM-based molecular representations to support both property prediction and candidate generation.Through experimental feedback cycles, the integration of AL enabled us to improve model accuracyand identify promising new molecular candidates. Among the top-performing candidates generated bythe model and selected for validation, melam, a dimer of melamine, demonstrated a 67% improvementin MA compared to the baseline, pure solvent. This result illustrates the potential of the approach foraccelerating materials discovery. More broadly, our results suggest that AL can guide molecular discoveryeven in low-data regimes and provide a reproducible framework for accelerating fuel cell development.Moreover, the methods presented here are generalizable and can be applied to a wide range of materialsscience problems involving molecular structure–property relationships.Our key contributions are as follows:• We demonstrate a closed-loop active-learning workflow for improving Pt-catalyst MA using organicadditives.• We develop a predictive ML model, trained and refined over multiple experimental feedback cycles.• We integrate a generative component to propose new candidate molecules, leading to several high-performing suggestions.2Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et al• We highlight the importance of data consistency, experimental feedback, and uncertainty-aware explo-ration in small-data settings.The remainder of this paper is organized as follows:Section 2 outlines the ML and experimental procedures. Section 2.1 provides a detailed descriptionof the molecular model, dataset, and learning strategy. Section 2.2 describes the processing of organiccompounds and the MA measurements, respectively. In section 3, we present the results across threeAL cycles. Section 4 discusses these findings, and section 5 concludes the paper with final remarks andpotential future directions.2. Methods2.1. MMLTo accelerate the discovery of effective organic additives for PEFCs, we employed an AL frameworkdriven by predictive modeling and molecular generation. For both tasks, we used the SMILES-Xtool [47], a MML model that learns structure–property relationships directly from SMILES strings, thesimplified molecular input line entry system [48], a widely used text-based representation of molecu-lar structures. Although SMILES representations are not unique and can lead to syntactically invalidmolecules during generation, these issues were addressed in this work through data augmentation,canonicalization, and post-generation validity checks.Graph neural networks (GNNs) provide an attractive alternative because they operate directly onmolecular graphs and avoid some limitations of text-based molecular representations [49, 50]. However,standard message-passing GNNs propagate information locally along graph edges and may thereforerequire multiple layers to capture dependencies between distant atoms. This can be problematic in low-data settings, where shallow architectures are often preferred, whereas deeper GNNs may suffer fromover-squashing or over-smoothing effects that limit long-range information propagation [51, 52]. Inmolecular systems, this limitation is particularly relevant for long-range or directional interactions, suchas electrostatic effects, which are not explicitly represented in standard message passing [53]. In contrast,the LSTM architecture used in SMILES-X was originally developed to model long-distance dependenciesin sequential data and can therefore exploit non-local patterns within SMILES strings [54]. This makesSMILES-X a practical and defensible choice for the present low-data experimental setting. In addition,one of the secondary aims of this study was to evaluate the applicability of our previously developedmethod to the present PEFC additive discovery problem.For molecular generation, SMILES-X uses a second LSTM-based model trained to generate SMILESstrings. Molecules are constructed token by token, using statistical likelihood and property-based heuris-tics to select the next character. Notably, SMILES-X reuses the property prediction model during gen-eration to estimate the properties of partially built molecules, guiding the generation toward promisingcandidates even before the full structure is completed [55].Our AL loop consisted of three cycles. After training on the initial dataset, the model was used toevaluate a list of organic molecules proposed by the experimental team based on domain knowledge andpractical feasibility. These compounds were ranked by the SMILES-X prediction model according to theirexpected MA, and the top candidates were selected for experimental validation.In the second cycle (Cycle 2), the model was retrained using the newly acquired experimental data,which improved its predictive accuracy and enabled its use not only for prediction but also for molecu-lar generation. SMILES-X was then used to generate new molecules predicted to act as effective additives.These generated candidates were prioritized based on predicted performance and chemical feasibility,and were considered alongside additional researcher-suggested candidates.In the third cycle (Cycle 3), the model was retrained using the full set of available data to furtherimprove its reliability. This final step provided a more robust predictor for evaluating additive perfor-mance and concluded the AL process.The training dataset was expanded after each cycle by incorporating newly measured and re-measured compounds, as detailed in section 3.The AL workflow is summarized as a flowchart in figure 1, showing the progression from Cycle 1 toCycle 3, including experimental data acquisition, model retraining, candidate evaluation, molecular gen-eration, and feedback into subsequent cycles. The corresponding evolution of the ML model perfor-mance is summarized in table 3.3Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alFigure 1. Overview of the active learning workflow combining SMILES-X-based prediction and generation with experimentalvalidation.Table 1.Hyperparameters selected for each dataset in the three active learning cyclesa.Cycle Batch size Learning rate Embedding size LSTM units Dense layer sizeCycle 1 16 10−3.5 2 8 8Cycle 2 8 10−3.5 2 8 256Cycle 3 8 10−2.5 1 16 256aEmbedding sizes of 1–2 are sufficient here due to character-level SMILES tokenization and the small-dataregime; increasing capacity led to overfitting.Throughout all cycles, we focused on predicting improvements in MA at 0.9 V relative to a base-line electrolyte. Molecules were encoded into latent vectors using the LSTM encoder of SMILES-X, anda regression model was trained to estimate MA. Model performance was assessed using 10-fold cross-validation, where the data is split into ten parts, and each subset is used once as a test set while theremaining nine serve for training. For each fold, we performed five independent training runs with dif-ferent random seeds to account for variability due to model initialization and training dynamics. Wereport the average prediction across these five runs for each test fold, and use the standard deviationacross runs to indicate uncertainty, represented as vertical error bars on the prediction plots.The model architecture and training process were governed by several hyperparameters, includ-ing batch size and learning rate, which influence training stability and convergence, as well as embed-ding size, number of LSTM units, and the size of the time-distributed dense layer, which determinethe model’s representational capacity. The architecture-related hyperparameters of the SMILES-X modelwere selected using the zero-shot neural architecture search algorithm epsinas [56]. This procedure wasapplied separately in each AL cycle to tailor the geometry of the LSTM-based architecture to the avail-able experimental dataset, enabling dataset-specific model design in the low-data regime considered inthis work. The selected hyperparameter values for each cycle are listed in table 1.2.2. CharacterizationThe ORR activity of Pt catalyst (TEC10E50E, Pt content: 46.5%, Tanaka Kikinzoku Kogyo Co. Ltd) wasmeasured by a method described in [21, 23, 24]. Briefly, electrochemical measurements of a Pt catalyst-modified electrode (evaluation of the ORR activity and electrochemically active surface area (ECSA))were conducted in the absence of additives. Then, the modified electrode was pulled up from the elec-trolyte solution (0.1 M HClO4), and the electrode was immersed in the solution of the additives for10 min. Then, the electrode was retrieved from the solution of the additives, rinsed with purified water,and re-immersed in the electrolyte solution (0.1 M HClO4). The same electrochemical measurementswere conducted again. By comparing the electrochemical properties after and before the treatment withadditives, effects of the additives on the catalytic activity can be revealed. The ORR MA was calculatedat 0.9 V and 0.95 V (vs a reversible hydrogen electrode). Relative MA is defined as the ratio of the MAafter/before the treatment of the additives, and is discussed as an indicator of the effect of additives.4Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alSpecific activity is determined by dividing MA with ECSA. Most of the additives were dissolved in ace-tone at 0.7 mM. The other compounds were dissolved in water, 0.1 M HClO4, or dimethylsulfoxide. Weselected the solvent based on the solubility of the additives. Concentrations of some compounds werechanged from 0.7 mM to other values (0.02 mM, 0.1 mM, 0.85 mM, 3 mM, 15 mM, or 20 mM).3. ResultsWe first constructed a curated molecular dataset for model training and active-learning-driven candidateselection. For modeling consistency, we selected the fixed-concentration subset (0.7 mM; 68 samples) forCycle 1 training (table 2). After adding 17 new measurements in Cycle 2 and several re-measured points,the training set grew to 85. Cycle 3 incorporated 12 additional measurements (8 expert-suggested and4 generated, commercially available), yielding 93 training samples used for the final predictor (figure 5and table 3).Each compound was characterized thanks to experimentally measured properties relevant to fuelcell performance. These included MA at 0.9 V and 0.95 V, and ECSA, measured as dimensionless ratiosbefore and after additive adsorption. Initial analysis showed strong correlations between these properties(see appendix, figure A1). Since slightly more data was available for MA at 0.9 V, we selected it as theprimary prediction target for model training.Cycle 1: initial model and candidate evaluationTo build the first model, all candidate compounds were validated using RDKit [57], an open-sourcecheminformatics toolkit widely used for molecular structure handling. RDKit was used to parse theSMILES strings and ensure that they cORResponded to chemically valid molecules, i.e. structures thatcould be interpreted without errors and complied with basic valence and bonding rules. Of the 96 com-pounds considered, 84 passed this validation step and were used for model training. These samplesspanned various solvent and concentration combinations. To ensure consistency in the training data, wecompared four filtering strategies based on whether solvent type and additive concentration were fixed:• Use all data regardless of additive concentration and solvent• Fix additive concentration, disregard solvent type• Fix solvent, disregard additive concentration• Fix both concentration and solventTable 2 summarizes these dataset sizes. From a chemical standpoint, additive concentration is expectedto influence performance more than solvent choice. To confirm that filtering did not adversely affectmodel accuracy, we trained a separate model for each filtering strategy using the same set of hyperpa-rameters, which were optimized on the full dataset. As shown in figure 2, the differences in performancewere minimal. We therefore chose the dataset with fixed concentration (68 samples) to balance size andconsistency.Using this model (Model 1), we evaluated a list of 34 candidate molecules provided by the experi-mental collaborators. These predictions helped prioritize compounds for experimental validation in thenext cycle.Cycle 2: updated model and candidate generationFigure 3 compares the predictions of Model 1 with the experimental measurements obtained for the18 expert-suggested candidate additives. This step served as a prospective evaluation of the initial modelon newly proposed compounds, rather than as an assessment of a retrained model on an expandeddataset. The comparison showed moderate predictive agreement: while compound-level deviationsremained, the prospective error was comparable to the cross-validation error, indicating that the initialmodel could provide useful trend-level guidance while still benefiting from additional consistent exper-imental data. These newly acquired measurements were therefore added to the original dataset to createan expanded training set. In addition, several samples were re-evaluated to improve data consistency andoverall data quality.We retrained the SMILES-X model on this updated data, consisting of 85 data points, to produceModel 2, which showed improved accuracy due to the increased sample size and more targeted data5Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alTable 2. Number of valid RDKit entries under different datafiltering strategies. Fixing additive concentration whileallowing solvent variation provided a good balance betweendata consistency and size.Filtering strategy Number of samplesNo filtering (all data) 84Fixed concentration only 68Fixed solvent only 61Fixed concentration and solvent 48Figure 2. Performance of initial SMILES-X models trained on datasets filtered by different additive concentration and solventcriteria.(figure 4). Model 2 was also used to guide a generative model trained on the same data. Together, theysuggested new candidate molecules predicted to improve MA. Notably, four of these generated moleculeswere found to be commercially available, allowing direct experimental validation without additional syn-thesis steps.Cycle 3: final model trainingIn the final cycle, eight more compounds from the chemists’ candidate list were evaluated, along withthe four commercially sourced molecules generated in Cycle 2, bringing the total dataset to 93 datapoints. These new measurements were added to the training set, and the final predictive model wastrained using the full dataset (Model 3). As shown in figure 5, the final model achieves the highest pre-dictive accuracy among all training cycles. Red circles mark the SMILES-X–enerated molecules, all ofwhich exhibit substantial performance gains, with MA improvements exceeding 20% relative to the no-additive baseline. The top-performing compound in this group is melam, a dimer of melamine, whichboosts MA by 67%.6Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alFigure 3. Predicted (gray) and experimental (black) relative mass activities for 17 expert-selected candidates measured after Cycle1. Predictions come from the fixed-concentration model.Figure 4. Cross-validation performance of Model 2 trained on the expanded dataset after Cycle 2. Black points represent the orig-inal dataset; dark gray points indicate re-measured data; large light-gray circles show results for chemist-suggested compounds.4. DiscussionThe development of next-generation energy materials increasingly depends on the ability to search largemolecular design spaces efficiently. In this context, ML and AL offer powerful strategies to accelerate dis-covery, reduce experimental cost, and help navigate the trade-offs between performance, stability, andmanufacturability. While much attention has been paid to models achieving state-of-the-art accuracy,our study highlights a complementary and often overlooked benefit: the ability of even modest, data-limited models to provide meaningful experimental guidance when integrated into a human-in-the-loopworkflow.Our results show that ML models, even when trained on small and noisy datasets, can effectively pri-oritize experimental targets. In this study, a model trained on only 68 initial measurements was used toselect 17 promising additives from an expert-curated list. As shown in figure 3, the model’s predictionsfor these new candidates showed useful trend-level agreement with the experimental outcomes, with test-set RMSE and MAE values comparable to those obtained during cross-validation. Specifically, the model7Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alFigure 5. Final SMILES-X model performance using 93 data points. Black points represent the original dataset; dark gray pointsindicate re-measured data; large light-gray circles show results for chemist-suggested compounds; red circles correspond to gen-erated compounds.Table 3. Performance of ML models across three active learning cycles. All metrics are calculated by 10-foldcross-validation.Cycle Training samples RMSE MAE Pearson r R2Cycle 1 68 0.271 0.202 0.48 0.14Cycle 2 85 0.226 0.180 0.57 0.29Cycle 3 93 0.218 0.165 0.61 0.33trained on the initial 68 samples achieved an out-of-sample RMSE of 0.271, while predictions on the17 new samples yielded an RMSE of 0.231. This level of agreement indicates that even models trainedon limited experimental datasets can detect performance trends, enabling researchers to focus effort onthe most promising candidates and thereby reduce screening time and material waste.Model performance improved steadily across three AL cycles, each consisting of retraining with newexperimental data. As summarized in table 3, every cycle added new, informative samples that reducedprediction error and strengthened the model’s ability to generalize. This effect is especially pronouncedin small-sample settings typical of materials science, where each measurement carries significant infor-mational weight. By Cycle 3, the model trained on the expanded dataset achieved the best overall per-formance, as reflected by the lowest RMSE and MAE and the highest Pearson correlation reported intable 3. These results confirm that the integration of human expertise, model-guided selection, and itera-tive feedback can progressively enhance predictive power.The final model was further used to screen novel molecules generated using SMILES-X. While gener-ative models can produce unrealistic or synthetically inaccessible structures, we focused on commerciallyavailable candidates to ensure experimental viability. Four of the top-ranked molecules were tested, allshowing improved MA in line with the model’s predictions, though in some cases the model slightlyoverestimated performance.The strong performance of melam is particularly informative because it suggests that motifs relatedto melamine-based chemistry remain valuable targets for ORR enhancement, while also pointing tonearby molecular variants that may offer improved stability, cost, or process compatibility. Althoughmelam suffers from stability and cost limitations, its performance demonstrates the model’s ability tosuggest useful directions for exploration. A second generated candidate, melamine, achieved a 51%. It isapproximately three orders of magnitude less expensive than the best-performing compound, Co-tBuTAP(90% MA enhancement [21]), as of August 2025 [58, 59], and contains no transition metals, makingit a highly practical alternative. These findings exemplify how even simple generative strategies, when8Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alcoupled with informed filtering and predictive models, can yield promising new candidates that balanceperformance with real-world constraints.Taken together, these results illustrate the multi-faceted role ML can play in early-stage materialsdiscovery. Beyond maximizing model accuracy, the integration of ML with expert insight and itera-tive experimental feedback enables more informed decisions, better data quality, and the discovery ofalternative compounds with desirable properties. Our workflow shows that progress does not requireperfect models, only useful ones, continuously improved by the right experiments. As data grows andtools evolve, such frameworks are likely to become standard practice in the search for new functionalmolecules.5. ConclusionWe demonstrated the successful application of AL to optimize fuel cell performance through the screen-ing of organic electrolyte additives. Our results show that ML models, even when trained on limited andnoisy data, can meaningfully guide the selection of candidates for experimental validation. By integratingexpert knowledge with predictive modeling and molecular generation, we were able to prioritize promis-ing compounds.Model performance improved steadily with each iteration, and the final model was used to screennew, commercially available molecules. Among the generated candidates, melam (a melamine dimer)achieved a 67% increase in MA relative to the additive-free baseline. Melamine itself also showed a 51%improvement while offering advantages in cost, availability, and solvent compatibility. Notably, melaminehas been known to be a good ORR enhancer [23–26], but our ML model was not aware of this whengenerating the SMILES. These results illustrate the potential for ML-guided discovery to yield practical,sustainable alternatives even without large datasets or state-of-the-art models.Overall, this work highlights the value of ML tools in small-data regimes. By combining predictivemodels with experimental iteration, candidate generation, and domain expertise, our approach providesa viable strategy for accelerating materials development in complex chemical systems.Following the completion of the three AL cycles, the experimental team identified an improved addi-tive application method that leads to significantly enhanced MA. This discovery opens a new directionfor optimization, and the data and models developed during the current study are already being reusedto inform the next set of experiments under the updated protocol.In parallel, the existing models are being applied to related experimental setups involving alternativecatalyst morphologies. These extensions highlight the broader applicability of the trained models andtheir underlying representations. Looking forward, the models developed here can be adapted to simi-lar tasks through a combination of strategies. When conditions change between experiments (such as incatalyst structure, measurement setup, or additive application method) domain adaptation techniquesbecome useful. For example, linear regression or other lightweight models can be employed to translatepredictions between related domains, leveraging the same molecular features. Alternatively, neural net-works can be fine-tuned via transfer learning by retraining only the final layers, allowing the core molec-ular understanding to be retained while adjusting to new conditions. These approaches extend the utilityof ML in data-scarce scientific environments and offer a flexible framework for accelerating discoveryacross related systems.Data availability statementThe data cannot be made publicly available upon publication because they contain commercially sensi-tive information. The data that support the findings of this study are available upon reasonable requestfrom the authors.Conflict of interestsThe authors declare that they have no competing interests.9Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alFundingThis work is supported by New Energy and Industrial Technology Development Organization (NEDO).Scientific contributionThis work establishes a closed-loop experimental workflow in which predictive modeling, candidategeneration, and feasibility-based experimental selection are combined to identify organic additives forimproving Pt-catalyst ORR MA under low-data conditions.Author contributionsEkaterina Gracheva  0000-0002-9704-5939Data curation (equal), Investigation (equal), Methodology (equal), Software (equal), Validation (equal),Visualization (equal), Writing – original draft (equal), Writing – review & editing (equal)Shin-ichi YamazakiConceptualization (equal), Data curation (equal), Formal analysis (equal), Investigation (equal),Validation (equal), Visualization (equal), Writing – original draft (equal), Writing – review &editing (equal)Guillaume Lambard  0000-0003-0275-4079Data curation (equal), Investigation (equal), Methodology (equal), Software (equal), Supervision (equal),Validation (equal), Visualization (equal), Writing – review & editing (equal)Keitaro SodeyamaConceptualization (equal), Funding acquisition (equal), Project administration (equal),Supervision (equal), Writing – review & editing (equal)Tsutomu IoroiConceptualization (equal), Data curation (equal), Formal analysis (equal), Funding acquisition (equal),Investigation (equal), Project administration (equal), Resources (equal), Supervision (equal),Validation (equal), Visualization (equal), Writing – review & editing (equal)Masafumi Asahi  0000-0002-2122-3073Conceptualization (equal), Data curation (equal), Formal analysis (equal), Investigation (equal),Methodology (equal), Project administration (equal), Resources (equal), Supervision (equal),Validation (equal), Visualization (equal), Writing – review & editing (equal)10https://orcid.org/0000-0002-9704-5939https://orcid.org/0000-0002-9704-5939https://orcid.org/0000-0003-0275-4079https://orcid.org/0000-0003-0275-4079https://orcid.org/0000-0002-2122-3073https://orcid.org/0000-0002-2122-3073Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et alAppendix. Correlationmatrix of measured propertiesFigure A1. Correlation matrix of measured properties across the initial dataset. Mass activity and Specific activity at 0.9 V, 0.95 V,show strong mutual Correlation, justifying the use of MA at 0.9 V as a primary prediction target.References[1] Majlan E, Rohendi D, Daud W, Husaini T and Haque M 2018 Electrode for proton exchange membrane fuel cells: a review Renew.Sustain. Energy Rev. 89 117–34[2] Gasteiger H A and Markovíc N M 2009 Just a dream—or future reality? Science 324 48–49[3] Zeng W, Guan B, Zhuang Z, Chen J, Zhu L, Ma Z, Hu X, Zhu C, Zhao S and Shu K 2025 Comprehensive review on the advancesand comparisons of proton exchange membrane fuel cells (pemfcs) and anion exchange membrane fuel cells (afcs): from funda-mental principles to key component technologies Int. J. Hydrog. Energy 102 222–46[4] Vallejo-Cervantes C, Espinoza-Andaluz M and Iranzo A 2025 Technical review of commercial lt-pemfc technologies: performance,applications and challenges Int. J. Hydrog. Energy 176 151480[5] Liu C-Y and Sung C-C 2012 A review of the performance and analysis of proton exchange membrane fuel cell membrane elec-trode assemblies J. Power Sources 220 348–53[6] Guo Y, Pan F, Chen W, Ding Z, Yang D, Li B, Ming P and Zhang C 2021 The controllable design of catalyst inks to enhance pemfcperformance: a review Electrochem. Energy Rev. 4 67–100[7] Borup R L, Kusoglu A, Neyerlin K C, Mukundan R, Ahluwalia R K, Cullen D A, More K L, Weber A Z and Myers D J 2020 Recentdevelopments in catalyst-related pem fuel cell durability Curr. Opin. Electrochem. 21 192–200[8] Hidayat E F, Zain S M and Noviyanti A R 2025 Advances in proton exchange membrane fuel cell (pemfc) materials: a review ofdevelopments from 2021 to 2025 J. Power Sources 657 238124[9] Zhang F, Zu B, Wang B, Qin Z, Yao J, Wang Z, Fan L and Jiao K 2025 Developing long-durability proton-exchange membrane fuelcells Joule 9 101853[10] Qasem N A 2024 A recent overview of proton exchange membrane fuel cells: fundamentals, applications and advances Appl.Therm. Eng. 252 123746[11] Chen S, Liu Y, Fu X and Wang W 2024 Recent advances in morphology control of platinum catalysts toward oxygen reductionreaction Front. Energy 18 330–55[12] Li J-R, Liu M-X, Liu X, Yu X-H, Li Q-Z, Sun Q, Sun T, Cao S and Hou C-C 2024 The recent progress of oxygen reduction electro-catalysts used at fuel cell level Small Methods 8 2301249[13] Snyder J, Fujita T, Chen M and Erlebacher J 2010 Oxygen reduction in nanoporous metal–ionic liquid composite electrocatalystsNat. Mater. 9 904–711https://doi.org/10.1016/j.rser.2018.03.007https://doi.org/10.1016/j.rser.2018.03.007https://doi.org/10.1126/science.1172083https://doi.org/10.1126/science.1172083https://doi.org/10.1016/j.ijhydene.2024.12.513https://doi.org/10.1016/j.ijhydene.2024.12.513https://doi.org/10.1016/j.ijhydene.2025.151480https://doi.org/10.1016/j.ijhydene.2025.151480https://doi.org/10.1016/j.jpowsour.2012.07.090https://doi.org/10.1016/j.jpowsour.2012.07.090https://doi.org/10.1007/s41918-020-00083-2https://doi.org/10.1007/s41918-020-00083-2https://doi.org/10.1016/j.coelec.2020.02.007https://doi.org/10.1016/j.coelec.2020.02.007https://doi.org/10.1016/j.jpowsour.2025.238124https://doi.org/10.1016/j.jpowsour.2025.238124https://doi.org/10.1016/j.joule.2025.101853https://doi.org/10.1016/j.joule.2025.101853https://doi.org/10.1016/j.applthermaleng.2024.123746https://doi.org/10.1016/j.applthermaleng.2024.123746https://doi.org/10.1007/s11708-024-0929-5https://doi.org/10.1007/s11708-024-0929-5https://doi.org/10.1002/smtd.202301249https://doi.org/10.1002/smtd.202301249https://doi.org/10.1038/nmat2878https://doi.org/10.1038/nmat2878Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et al[14] Zhang G-R, Munoz M and Etzold B J 2016 Accelerating oxygen-reduction catalysts through preventing poisoning with non-reactive species by using hydrophobic ionic liquids Angew. Chem., Int. Edn. 55 2257–61[15] Li Y, Hart J, Profitt L, Intikhab S, Chatterjee S, Taheri M and Snyder J 2019 Sequential capacitive deposition of ionic liquids forconformal thin film coatings on oxygen reduction reaction electrocatalysts ACS Catal. 9 9311–6[16] Li Y, Intikhab S, Malkani A, Xu B and Snyder J 2020 Ionic liquid additives for the mitigation of nafion specific adsorption on plat-inum ACS Catal. 10 7691–8[17] Avid A, Ochoa J L, Huang Y, Liu Y, Atanassov P and Zenyuk I V 2022 Revealing the role of ionic liquids in promoting fuel cell cat-alysts reactivity and durability Nat. Commun. 13 6349[18] Wang L, Morales-Collazo O, Brennecke J F and Jia H 2024 Impact of a protic dimeric ionic liquid on pem fuel cell performanceand durability Int. J. Hydrog. Energy 58 40–45[19] Chung Y-H, Chung D Y, Jung N and Sung Y-E 2013 Tailoring the electronic structure of nanoelectrocatalysts induced by a surface-capping organic molecule for the oxygen reduction reaction J. Phys. Chem. Lett. 4 1304–9[20] Miyabayashi K, Nishihara H and Miyake M 2014 Platinum nanoparticles modified with alkylamine derivatives as an active andstable catalyst for oxygen reduction reaction Langmuir 30 2936–42[21] Yamazaki S, Asahi M and Ioroi T 2019 Promotion of oxygen reduction on a porphyrazine-modified pt catalyst surface Electrochim.Acta 297 725–34[22] Yamazaki S, Asahi M, Taguchi N and Ioroi T 2019 Electrochemical analysis of the porphyrazine-induced enhancement of ORRactivity of pt catalysts for the development of porphyrazine-adsorbed pt catalysts J. Electroanal. Chem. 848 113321[23] Asahi M, Yamazaki S, Taguchi N and Ioroi T 2019 Facile approach to enhance oxygen reduction activity by modification of plat-inum nanoparticles by melamine-formaldehyde polymer J. Electrochem. Soc. 166 498[24] Yamazaki S, Asahi M, Taguchi N, Ioroi T, Kishimoto Y, Daimon H, Inaba M, Koga K, Kurose Y and Inoue H 2020 Creation of ahighly active pt/pd/c core–shell-structured catalyst by synergistic combination of intrinsically high activity and surface decorationwith melamine or tetra-(tert-butyl)-tetraazaporphyrin ACS Catal. 10 14567–80[25] Yamazaki S, Asahi M, Taguchi N and Ioroi T 2022 Sulphonated melamine polymer for enhancing the oxygen reduction reactionactivity and stability of a pt catalyst J. Electroanal. Chem. 908 116103[26] Wada N, Nakamura M and Hoshi N 2020 Structural effects on the oxygen reduction reaction on pt single-crystal electrodes modi-fied with melamine Electrocatalysis 11 275–81[27] Yamazaki S, Asahi M, Taguchi N and Ioroi T 2023 Triazine compounds for the high enhancement of the ORR activity and durabil-ity of Pt/C Electrochim. Acta 472 143417[28] Jinnouchi R and Minami S 2024 The melamine-driven solvation effect promotes oxygen reduction on a platinum catalyst:machine learning-aided free energy calculations J. Phys. Chem. Lett. 16 265–73[29] Butler K T, Davies D W, Cartwright H, Isayev O and Walsh A 2018 Machine learning for molecular and materials science Nature559 547–55[30] Goldsmith B R, Esterhuizen J, Liu J-X, Bartel C J, Sutton C 2018 Machine learning for heterogeneous catalyst design and discovery[31] McCullough K, Williams T, Mingle K, Jamshidi P and Lauterbach J 2020 High-throughput experimentation meets artificial intelli-gence: a new pathway to catalyst discovery Phys. Chem. Chem. Phys. 22 11174–96[32] Kim C, Batra R, Chen L, Tran H and Ramprasad R 2021 Polymer design using genetic algorithm and machine learning Comput.Mater. Sci. 186 110067[33] Sharma A, Mukhopadhyay T, Rangappa S M, Siengchin S and Kushvaha V 2022 Advances in computational intelligence of poly-mer composite materials: machine learning assisted modeling, analysis and design Arch. Comput. Methods Eng. 29 3341–85[34] Gracheva E, Lambard G, Samitsu S, Sodeyama K and Nakata A 2021 Prediction of the coefficient of linear thermal expansion forthe amorphous homopolymers based on chemical structure using machine learning Sci. Technol. Adv. Mater. 1 213–24[35] Wu S, Kondo Y, Kakimoto M-A, Yang B, Yamada H, Kuwajima I, Lambard G, Hongo K, Xu Y and Shiomi J 2019 Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm npj Comput. Mater.5 66[36] Yang X, Wang Y, Byrne R, Schneider G and Yang S 2019 Concepts of artificial intelligence for computer-assisted drug discoveryChem. Rev. 119 10520–94[37] Jiménez-Luna J, Grisoni F and Schneider G 2020 Drug discovery with explainable artificial intelligence Nat. Mach. Intell. 2 573–84[38] Ding R, Zhang S, Chen Y, Rui Z, Hua K, Wu Y, Li X, Duan X, Wang X and Li J 2022 Application of machine learning in optimizingproton exchange membrane fuel cells: a review Energy AI 9 100170[39] Ming W, Sun P, Zhang Z, Qiu W, Du J, Li X, Zhang Y, Zhang G, Liu K and Wang Y 2023 A systematic review of machine learningmethods applied to fuel cells in performance evaluation, durability prediction and application monitoring Int. J. Hydrog. Energy48 5197–228[40] Settles B 2012 Active Learning (Synthesis Lectures on Artificial Intelligence and Machine Learning) (Morgan & Claypool Publishers)[41] Chen C-T and Gu G X 2020 Generative deep neural networks for inverse materials design using backpropagation and active learn-ing Adv. Sci. 7 1902607[42] Lookman T, Balachandran P V, Xue D and Yuan R 2019 Active learning in materials science with emphasis on adaptive samplingusing uncertainties for targeted design npj Comput. Mater. 5 21[43] Xu P, Ji X, Li M and Lu W 2023 Small data machine learning in materials science npj Comput. Mater. 9 42[44] Lambard G, Sasaki T, Sodeyama K, Ohkubo T and Hono K 2022 Optimization of direct extrusion process for Nd-Fe-B magnetsusing active learning assisted by machine learning and bayesian optimization Scr. Mater. 209 114341[45] Kim C, Chandrasekaran A, Jha A and Ramprasad R 2019 Active-learning and materials design: the example of high glass transi-tion temperature polymersMrs Commun. 9 860–6[46] Ang S J, Wang W, Schwalbe-Koda D, Axelrod S and Gómez-Bombarelli R 2021 Active learning accelerates ab initio moleculardynamics on reactive energy surfaces Chem 7 738–51[47] Lambard G and Gracheva E 2020 Smiles-x: autonomous molecular compounds characterization for small datasets withoutdescriptorsMach. Learn.: Sci. Technol. 1 025004[48] Weininger D 1988 Smiles, a chemical language and information system. 1. Introduction to methodology and encoding rules J.Chem. Inf. Comput. Sci. 28 31–36[49] Wu X, Wang H, Gong Y, Fan D, Ding P, Li Q and Qian Q 2023 Graph neural networks for molecular and materials representationJ. Mater. Inform. 3 12[50] Li L, Zhang Y, Wang G and Xia K 2025 Kolmogorov–arnold graph neural networks for molecular property prediction Nat. Mach.Intell. 7 1346–5412https://doi.org/10.1002/anie.201508338https://doi.org/10.1002/anie.201508338https://doi.org/10.1021/acscatal.9b03157https://doi.org/10.1021/acscatal.9b03157https://doi.org/10.1021/acscatal.0c01243https://doi.org/10.1021/acscatal.0c01243https://doi.org/10.1038/s41467-022-33895-5https://doi.org/10.1038/s41467-022-33895-5https://doi.org/10.1016/j.ijhydene.2024.01.163https://doi.org/10.1016/j.ijhydene.2024.01.163https://doi.org/10.1021/jz400574fhttps://doi.org/10.1021/jz400574fhttps://doi.org/10.1021/la402412khttps://doi.org/10.1021/la402412khttps://doi.org/10.1016/j.electacta.2018.11.165https://doi.org/10.1016/j.electacta.2018.11.165https://doi.org/10.1016/j.jelechem.2019.113321https://doi.org/10.1016/j.jelechem.2019.113321https://doi.org/10.1149/2.0641908jeshttps://doi.org/10.1149/2.0641908jeshttps://doi.org/10.1021/acscatal.0c03124https://doi.org/10.1021/acscatal.0c03124https://doi.org/10.1016/j.jelechem.2022.116103https://doi.org/10.1016/j.jelechem.2022.116103https://doi.org/10.1007/s12678-020-00584-0https://doi.org/10.1007/s12678-020-00584-0https://doi.org/10.1016/j.electacta.2023.143417https://doi.org/10.1016/j.electacta.2023.143417https://doi.org/10.1021/acs.jpclett.4c03437https://doi.org/10.1021/acs.jpclett.4c03437https://doi.org/10.1038/s41586-018-0337-2https://doi.org/10.1038/s41586-018-0337-2https://doi.org/10.1039/D0CP00972Ehttps://doi.org/10.1039/D0CP00972Ehttps://doi.org/10.1016/j.commatsci.2020.110067https://doi.org/10.1016/j.commatsci.2020.110067https://doi.org/10.1007/s11831-021-09700-9https://doi.org/10.1007/s11831-021-09700-9https://doi.org/10.1080/27660400.2021.1993729https://doi.org/10.1080/27660400.2021.1993729https://doi.org/10.1038/s41524-019-0203-2https://doi.org/10.1038/s41524-019-0203-2https://doi.org/10.1021/acs.chemrev.8b00728https://doi.org/10.1021/acs.chemrev.8b00728https://doi.org/10.1038/s42256-020-00236-4https://doi.org/10.1038/s42256-020-00236-4https://doi.org/10.1016/j.egyai.2022.100170https://doi.org/10.1016/j.egyai.2022.100170https://doi.org/10.1016/j.ijhydene.2022.10.261https://doi.org/10.1016/j.ijhydene.2022.10.261https://doi.org/10.1002/advs.201902607https://doi.org/10.1002/advs.201902607https://doi.org/10.1038/s41524-019-0153-8https://doi.org/10.1038/s41524-019-0153-8https://doi.org/10.1038/s41524-023-01000-zhttps://doi.org/10.1038/s41524-023-01000-zhttps://doi.org/10.1016/j.scriptamat.2021.114341https://doi.org/10.1016/j.scriptamat.2021.114341https://doi.org/10.1557/mrc.2019.78https://doi.org/10.1557/mrc.2019.78https://doi.org/10.1016/j.chempr.2020.12.009https://doi.org/10.1016/j.chempr.2020.12.009https://doi.org/10.1126/science.aaa841https://doi.org/10.1126/science.aaa841https://doi.org/10.1021/ci00057a005https://doi.org/10.1021/ci00057a005https://doi.org/10.20517/jmi.2023.10https://doi.org/10.20517/jmi.2023.10https://doi.org/10.1038/s42256-025-01087-7https://doi.org/10.1038/s42256-025-01087-7Mach. Learn.: Sci. Technol. 7 (2026) 045006 E Gracheva et al[51] Alon U and Yahav E 2020 On the bottleneck of graph neural networks and its practical implications (arXiv: 2006.05205)[52] Di Giovanni F, Giusti L, Barbero F, Luise G, Lio P and Bronstein MM 2023 On over-squashing in message passing neural net-works: the impact of width, depth and topology Int. Conf. on Machine Learning (PMLR) pp 7865–85[53] Thürlemann M and Riniker S 2023 Anisotropic message passing: graph neural networks with directional and long-range interac-tions The 11th Int. Conf. on Learning Representations[54] Hochreiter S and Schmidhuber J 1997 Long short-term memory Neural Comput. 9 1735–80[55] Lambard G SMILES-X: a molecular machine learning tool (available at: https://github.com/Lambard-ML-Team/SMILES-X)(Accessed 03 July 2025)[56] Gracheva E 2025 Neural architecture search: two constant shared weights initialisations Artif. Intell. Rev. 58 296[57] RDKit: open-source cheminformatics (available at: www.rdkit.org/)[58] TCI Chemicals Melamine (CAS No. 108-78-1) - Product T0337 (available at: www.tcichemicals.com/JP/en/p/T0337) (Accessed20 August 2025)[59] Sigma-Aldrich 2,7,12,17-Tetra-tert-butyl-5,10,15,20-tetraaza-21H,23H-porphyrin (Dye Content 85) (available at: www.sigmaaldrich.com/JP/ja/product/sial/42836) (Accessed 7 September 2025)13https://arxiv.org/abs/2006.05205https://doi.org/10.1162/neco.1997.9.8.1735https://doi.org/10.1162/neco.1997.9.8.1735https://github.com/Lambard-ML-Team/SMILES-Xhttps://doi.org/10.1007/s10462-025-11238-2https://doi.org/10.1007/s10462-025-11238-2https://www.rdkit.org/https://www.tcichemicals.com/JP/en/p/T0337https://www.sigmaaldrich.com/JP/ja/product/sial/42836https://www.sigmaaldrich.com/JP/ja/product/sial/42836 Enhancing oxygen reduction reaction mass activity in polymer electrolyte fuel cells via molecular machine learning 1. Introduction 2. Methods 2.1. MML 2.2. Characterization 3. Results 4. Discussion 5. Conclusion Appendix. Correlation matrix of measured properties References