# Fileset

[Advanced Science - 2026 - Fajar - Generative AI‐Driven Accelerated Discovery of Passivation Molecules for Perovskite Solar (1).pdf](https://mdr.nims.go.jp/filesets/961cc7b1-dc12-4702-913f-bd37edafb373/download)

## Creator

Adroit T. N. Fajar, [Guillaume Lambard](https://orcid.org/0000-0003-0275-4079), Jessie Manopo, Ruili Guo, Kevin Septioga, Rizfi F. Pari, Toshinori Matsushima, Zhanglin Guo

## Rights

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

## Other metadata

[Generative AI‐Driven Accelerated Discovery of Passivation Molecules for Perovskite Solar Cells](https://mdr.nims.go.jp/datasets/3f546f65-79be-4db8-b890-a661bcda7166)

## Fulltext

Generative AI‐Driven Accelerated Discovery of Passivation Molecules for Perovskite Solar CellsAdvanced Science www.advancedscience.comRESEARCH ARTICLEGenerative AI-Driven Accelerated Discovery of Passivation Molecules for Perovskite Solar Cells Adroit T. N. Fajar1 , 2 Guillaume Lambard3 Jessie Manopo1 , 2 Ruili Guo1 Kevin Septioga4 Rizfi F. Pari4 Toshinori Matsushima1 , 2 , 4 , 5 Zhanglin Guo1 , 2 , 5 1 International Institute for Carbon-Neutral Energy Research (WPI-I2 CNER), Kyushu University, Fukuoka, Japan 2 Center for Energy Systems Design (CESD), International Institute for Carbon-Neutral Energy Research (WPI-I2 CNER), Kyushu University, Fukuoka, Japan 3 Data-Driven Materials Design Group, Center for Basic Research on Materials, National Institute for Materials Science, Tsukuba, Japan 4 Department of Applied Chemistry, Graduate School of Engineering, Kyushu University, Fukuoka, Japan 5 Department of Automotive Science, Graduates School of Integrated Frontier Sciences, Kyushu University, Fukuoka, Japan Correspondence: Adroit T. N. Fajar ( adroit@i2cner.kyushu-u.ac.jp) Zhanglin Guo ( guo.zhanglin.903@m.kyushu-u.ac.jp) Received: 13 November 2025 Revised: 26 March 2026 Accepted: 27 March 2026 Keywords: generative AI | inverse design | perovskite solar cells | passivation molecules | SMILES ABSTRACT Molecular passivation is an effective strategy to suppress interfacial defects in perovskite solar cells (PSCs), yet the discovery of new passivation molecules remains limited by empirical design and narrow chemical libraries. Here, for the first time, we present an AI- driven framework integrating discriminative and generative language models to accelerate the discovery of effective passivators. A SMILES-X classifier trained on literature data achieved high predictive performance (F1 = 0.80, ROC–AUC = 0.88), while a GPT-2-based generative model iteratively produced over 100 000 novel molecules, more than 80% of which were predicted to be effective. Multi-criteria filtering reduced this pool to ∼ 8000 high-quality candidates, from which clustering analysis identified ten diverse representatives. Three molecules, including a surrogate analog, were prioritized for experimental testing, and all exhibited a clear passivation effect. In particular, 4-maleimidobutyric acid increased the average open-circuit voltage from 1.08 to 1.12 V and improved the average power conversion efficiency from 19.3% to 22.2%, while markedly reducing hysteresis. This study demonstrates that generative AI can autonomously propose synthetically accessible, functionally effective molecules for PSC passivation, offering a powerful paradigm for accelerated materials discovery beyond conventional chemical space exploration. 1H  m  t  c  t  e  w  [  e           Tc©Ah Introduction alide perovskite solar cells (PSCs) have emerged as one of theost promising next-generation photovoltaic technologies owingo their exceptional optoelectronic properties, facile solution pro-essability, and tunable bandgaps. Since their initial demonstra-ion with a modest power conversion efficiency (PCE) of 3.8%, thefficiency of single-junction PSCs has rapidly advanced to 27.0%,hile perovskite-based tandem devices are now approaching 35% 1, 2 ]. Despite this remarkable progress, further improvements infficiency and operational lifetime are hindered by interfacial andhis is an open access article under the terms of the Creative Commons Attribution License, which permited. 2026 The Author(s). Advanced Science published by Wiley-VCH GmbH dvanced Science , 2026; 13:e23042 ttps://doi.org/10.1002/advs.202523042surface defects that promote nonradiative recombination and ionmigration. Surface defect passivation using small organic molecules hasproven to be one of the most effective strategies to mitigate theseissues. A representative example is phenethylammonium iodide,which binds to undercoordinated Pb2 + or halide vacancies andcan form quasi-2D capping layers, thereby improving efficiency[ 3, 4 ]. Numerous other functional molecules bearing amine, thiol,carbonyl, or carboxyl groups have also been reported to interactwith the perovskite lattice and suppress defects [ 5 ]. However,its use, distribution and reproduction in any medium, provided the original work is properly 1 of 10http://www.advancedscience.comhttps://doi.org/10.1002/advs.202523042https://orcid.org/0000-0003-3168-2588mailto:adroit@i2cner.kyushu-u.ac.jpmailto:guo.zhanglin.903@m.kyushu-u.ac.jphttp://creativecommons.org/licenses/by/4.0/https://doi.org/10.1002/advs.202523042http://crossmark.crossref.org/dialog/?doi=10.1002%2Fadvs.202523042&domain=pdf&date_stamp=2026-04-02t  o  c  aM  a  r  v  r  s  S  D  c  d  n  l  a  b  g  c  a  i  d  r  sG  p  m  t  m  m  d  S  i  s  s  s  a  s  [  d  b  r  hI  e  f  i  2  t  p  g  t  o  c  m  A  v                                                     2 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creathe discovery of new passivation molecules has relied heavilyn empirical trial-and-error chemistry, which is time-consuming,ostly, and often lacks clear design principles. A more efficientnd systematic approach is therefore urgently needed. achine learning (ML) has been used as a powerful tool toccelerate molecule discovery by learning structure–propertyelationships from large datasets. Early examples include super-ised models trained on 21 organic–halide capping layers, whichevealed that a low hydrogen-bond donor count and small polarurface area could extend perovskite film lifetime fourfold [ 6 ].ubsequent studies employing random forest regressors andFT-derived binding energy filters identified NH3 + -terminatedations such as 2-phenylpropane-1-aminium iodide, boostingevice efficiency to 24.5% [ 7 ]. More advanced physics-informedeural networks trained on full-DFT datasets screened five mil-ion pseudo-halide anions and discovered sodium thioglycolate,chieving 24.6% PCE with 96% retention after 900 h [ 8 ]. Ensem-le learning and complexity-weighted descriptors, respectively,uided the selection of sulfonated and carboxyphthalide-basedandidates that enhanced device efficiency [ 9, 10 ]. Despite thesedvances, conventional ML approaches remain inherently lim-ted by their reliance on predefined chemical libraries and labeledatasets, restricting exploration to known molecular spacesather than enabling the generation of entirely new chemicaltructures with unconventional functionalities. enerative artificial intelligence (AI) offers a transformativearadigm shift, from screening known compounds to inverseolecular design. By leveraging sequence- or graph-based archi-ectures, generative AI can navigate vast chemical spaces, esti-ated to contain more than 1060 synthetically accessible smallolecules [ 11 ], and propose novel compounds beyond existingatabases. Recent reinforcement learning frameworks, such asyntheMol, have successfully generated and experimentally val-dated new antibiotic candidates [ 12 ]. Similarly, in materialscience, diffusion-based models like MatterGen have demon-trated more than a twofold increase in discovering stable crystaltructures compared with traditional generators [ 13 ]. However,pplications of generative AI to the discovery of molecular pas-ivators for perovskite photovoltaics remain largely unexplored 14, 15 ]. Establishing such AI-driven pipelines could enable theesign of previously unknown molecular scaffolds with superiorinding affinities, defect-healing capabilities, and photothermalobustness, which are key requirements for achieving stable,igh-efficiency PSCs. n this study, we present a language model (LM)-based gen-rative framework for the discovery of molecular passivatorsor perovskite solar cells. Specifically, for the first time, wentegrate fine-tuned large language models (LLMs, e.g., GPT- and LLaMA-2) with an LM-based molecular characterizationool (SMILES-X) and employ an iterative generation–screeningipeline to produce over 100 000 novel candidate molecules. Theenerated molecules are filtered using seven physicochemical cri-eria and grouped into ten fingerprint-based clusters, from whichne representative molecule is sampled per cluster to ensurehemical diversity. After expert evaluation, three representativeolecules were experimentally tested in PSC devices. All threeI-generated molecules effectively improved the open-circuitoltage ( VOC ) of inverted PSCs, demonstrating their ability toof 10passivate perovskite surface defects. Notably, 4-maleimidobutyricacid (MBA) not only enhanced VOC but also optimized interfacialenergy alignment by eliminating electron-transport barriers,thereby minimizing energy losses and reducing J –V hysteresis.These results demonstrate the potential and reliability of AI-driven molecular generation for accelerated materials discoveryin perovskite photovoltaics, offering a powerful new pathway toexplore the vast, uncharted chemical space for next-generationsolar energy materials. 2 Results and Discussion The workflow for discovering novel passivation molecules usingthe generative AI framework is illustrated in Scheme 1 . Itbegins with the construction of an initial database (Data T0,314 molecules) consisting of experimentally reported moleculescollected from the literature, each labeled according to its rela-tive (normalized) improvement in power conversion efficiency( ΔPCEnorm , see Equation S1 ). This dataset was used to train adiscriminative model, SMILES-X [ 16 ], which directly interpretsmolecular SMILES representations to classify molecules as eithereffective (class 1) or ineffective (class 0) passivators. The trainedmodel was then employed to predict new candidates with highstructural similarity to the top-performing molecules, therebyaugmenting the original training data. In the second stage,a generative language model (GPT-2) was fine-tuned on thesubset of effective molecules from the augmented dataset (DataT1) to autonomously generate new molecular structures withpotential passivation capability. Through three iterative trainingand inference cycles, the model produced over 100 000 molecules,of which more than 80% were predicted to belong to class 1.The generated molecules were subsequently filtered based onphysicochemical and structural criteria, yielding a refined setof ∼ 8000 promising candidates. These were further groupedinto ten clusters according to structural similarity. Finally, onerepresentative molecule was selected from each cluster, and basedon expert evaluation, three were chosen for experimental testing.The following sections describe the detailed processes of thisstudy and the results obtained. 2.1 Existing Passivation Molecules Understanding existing data is a crucial first step in buildingan AI system for any specific application. In the field of PSCs,hundreds of passivation molecules have been reported overthe past decade (see Data T0). Data T0 was constructed bysystematically mining previously reported passivation moleculesfrom curated review articles and related primary studies, followedby manual verification of the original publications. The datasetwas rigorously preprocessed to remove duplicates, convert allstructures to canonical SMILES, and extract paired initial/finalPCE values, yielding 314 unique labeled molecules (see Sup-porting Information). As shown in Figure 1a , most reportedmolecules result in moderate improvements in device perfor-mance (normalized PCE improvement, ΔPCEnorm ≈ 0.10). A fewmolecules exhibit negative effects (typically reported as controlsin the original studies), while some demonstrate exceptionallypositive effects. Among these molecules, electronegative atoms(i.e., O, N, F, S) are commonly utilized to induce defect passivationAdvanced Science, 2026ive Commons LicenseSCHEME 1 Workflow of AI-driven novel passivation molecules discovery and verification. (  c  t  t  c  w  F  r  m  rT  y  t  m  f  s  T  s  t  w  m  p  p  T  a  m  i  r  d  l                       A 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable CreatFigure 1b ). For example, functional groups with a higher oxygenontent have been reported to exhibit stronger binding energyoward undercoordinated Pb2 + defects on the perovskite surfacehrough the Lewis acid–base coordination [ 8 ]. Similarly, nitrogen-ontaining functional groups can interact via hydrogen bondingith FA+ /MA+ cations, contributing to defect suppression [ 17 ].igure 1c presents the distribution of atom counts in the dataset,eflecting molecular size. Small molecules ( < 20 atoms) areost frequently used for surface passivation, whereas the largesteported molecules contain up to 60 atoms. o further characterize the dataset, principal component anal-sis (PCA) and uniform manifold approximation and projec-ion (UMAP) were applied to the Morgan fingerprints of theolecules. In these plots, PCA-1 and PCA-2 correspond to theirst and second principal components, which capture the mostignificant directions of variation in the molecular feature space.hey help visualize how molecules differ based on their overalltructural characteristics. Similarly, UMAP-1 and UMAP-2 arehe two main dimensions obtained from the UMAP projection,hich preserves both local and nonlinear relationships betweenolecules, making it easier to observe potential clusteringatterns in the dataset [ 18 ]. Figure 1d shows the 2D PCArojection of the molecules, labeled by their ΔPCEnorm values.his analysis provides insight into global variance relationshipsmong molecules. However, molecular structures with low,oderate, and high passivation effects are not clearly separatedn the chemical space, suggesting that the structure–propertyelationship governing passivation is highly complex. Figure 1eisplays the 2D UMAP projection with binary classificationabels, where class 0 represents ineffective passivation ( ΔPCEnormdvanced Science, 2026< 0.10) and class 1 represents effective passivation ( ΔPCEnorm ≥0.10). Although UMAP better preserves local structures and non-linear relationships than PCA, distinct class groupings are stillnot observed. These analyses indicate that identifying suitablemolecular scaffolds or substructures for perovskite passivationis inherently challenging, even for human experts (see FigureS1 ), and thus strongly motivates the application of ML-basedapproaches. 2.2 Expanding the Chemical Space Building accurate ML prediction models for molecular structure–property relationships typically requires rigorous feature extrac-tion by domain-specific experts or quantum chemical calcula-tions [ 19, 20 ], which are time-consuming and resource-intensive.Following the principle of natural language processing, SMILES-X bypasses this step by directly using the SMILES string asinput (optionally augmented with a few simple descriptors) andmapping it to property labels [ 16 ]. Here, a binary classificationmodel was developed to distinguish ineffective (class 0) andeffective (class 1) passivation molecules by training SMILES-Xon Data T0. To evaluate model performance, both the precision–recall (PR) curve and confusion matrix were analyzed based onthe final out-of-sample predictions. As shown in Figure 2a , the PRcurve illustrates the trade-off between precision and recall acrossa range of probability thresholds (mean scores). The optimalthreshold of 0.47 yielded the maximum F1 score of 0.80. The corresponding confusion matrix (Figure 2b ) provides adetailed breakdown of prediction outcomes aggregated over five-3 of 10ive Commons LicenseFIGURE 1 Characteristics of the existing passivation molecules. Distribution of (a) passivation effects ( ΔPCEnorm values), (b) atom types, and (c) atom counts in the dataset (Data T0). Chemical space visualization of the dataset using (d) PCA labeled with ΔPCEnorm values, and (e) UMAP labeled with binary classification results. f  (  r  b  r  0  aA  s  p  p  f  R  X  w  m  e  dB  r  u  s  f  b                         4 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creatold cross-validation: 157 true negatives (TN), 102 true positivesTP), 31 false negatives (FN), and 24 false positives (FP). Theseesults indicate that the model performs well in distinguishingetween the two classes, with relatively low misclassificationates. Furthermore, the precision, recall, and F1-score averaged.82 based on cross-validation, confirming robust performancecross folds. The area under the precision–recall curve (PRC–UC) was 0.86, and the ROC–AUC reached 0.88, highlightingtrong discriminative ability even for an imbalanced dataset. Thiserformance is comparable to that of a binary classification modelrepared using the random forest (RF) algorithm with Morganingerprints as input features (see Figure S2 ). However, unlike theF approach, which requires explicit feature extraction, SMILES- is computationally efficient and can be seamlessly integratedith LM–based generative frameworks. Therefore, the developedodel achieves both high predictive accuracy and computationalfficiency, making it suitable for guiding AI-driven molecularesign. ecause molecular structures encoded as SMILES can beegarded as a chemical language [ 21 ], they can also be directlysed to fine-tune LLMs. However, the number of reported pas-ivation molecules in Data T0 is relatively small and insufficientor LLM fine-tuning. To address this, the dataset was augmentedy retrieving molecules from the PubChem database with at leastof 1080% Tanimoto similarity to the top-performing class 1 molecules( ΔPCEnorm > 0.16), yielding 15 540 additional entries (Data T-aug).As shown in Figure 2c , approximately 70% of these moleculeswere classified as class 1 by the SMILES-X model. A new datasetcombining class 1 molecules from Data T0 and Data T-aug wasthen defined as Data T1 and was used to fine-tune both GPT-2 andLLaMA-2 models. As proven in the following, this method largelyimproved the reliability and effectiveness of generative AI-drivenmolecule discovery. The test loss distributions (Figure S3a,b )indicate that both models achieved low average losses (GPT-2:0.13; LLaMA-2: 0.15), confirming that they successfully learnedthe atomic sequence patterns associated with effective passivationmolecules. Despite comparable performance, LLaMA-2 requiredsubstantially greater computational resources (five times longerfor training and over 100 times slower for inference) due to itslarge parameter count (7 billion), which is designed for moreexpressive human conversation (see Methods). Consequently, theGPT-2 model was selected for subsequent analysis. To further expand the chemical search space of perovskite pas-sivation molecules, the GPT-2 model was fine-tuned iterativelyby incorporating class 1 molecules generated from each previouscycle into the training dataset (details in Methods). The iterativefine-tuning strategy was adopted as a practical guided-explorationapproach for domain-specific molecular discovery, where theAdvanced Science, 2026ive Commons LicenseFIGURE 2 Expanding the chemical space of PSC passivation molecules using language models. (a) Precision–recall (PR) curve and (b) confusion matrix of the binary classification model prepared with SMILES-X trained on Data T0. (c) UMAP projection of Data T-aug labeled with binary classification results predicted by SMILES-X. (d) Number of molecules and (e) chemical space visualization across three cycles of iterative training and generation with the fine-tuned GPT-2 model. a  F  c  i  g  p  t  e  1  t  t  o  r  A  p  t  s  s  v  i  a  (  o  i                     A 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creatvailable high-quality training data are limited. As shown inigure 2d , one-shot generation (cycle 1) yields only ∼ 30 000hemically valid, unique, and novel (CUN) molecules, which isnsufficient to reach the targeted chemical space size. Iterativeeneration progressively enriches the training set with model-rioritized candidates, allowing the number of CUN moleculeso exceed 100 000 by the third cycle. The generative modelxpanded the dataset by approximately tenfold—from about0 000 molecules in Data T1 to over 100 000 in Data G-all. Morehan 80% (87 750) of the generated molecules were predicted byhe SMILES-X classifier to have a high model-predicted likelihoodf belonging to the effective passivation class (Data G-class1),eflecting enrichment toward chemically promising candidates. concise sensitivity analysis for the classifier threshold isrovided in Table S1 . These predictions serve as a probabilis-ic prioritization signal to guide exploration of the chemicalpace, rather than as ground-truth labels. Consequently, furthercreening based on physicochemical criteria and experimentalalidation serves as the final validation steps. Although additionalterations could have been performed, the process was terminatedfter the third cycle to maintain a manageable dataset size ∼ 100 000 molecules) and prevent model collapse, a phenomenonbserved in our previous work on ionic liquid generation andn other general-purpose generative models [ 22, 23 ]. The atom-dvanced Science, 2026type distribution for the cumulative generated molecules (DataG-all) is shown in Figure S3c , while Figure 2e illustrates theprogressive expansion of the explored chemical space acrossthree iterative cycles relative to the original training data. Toquantitatively assess structural novelty, we computed the nearest-neighbor Tanimoto similarity between the cumulative generatedlibrary and the initial GPT-2 fine-tuning dataset (Figure S4 ).The resulting distribution indicates substantial chemical diver-sification beyond direct scaffold replication, with the majorityof generated molecules exhibiting moderate similarity to thetraining set rather than near-identical structures. Overall, thisiterative fine-tuning strategy effectively broadened the accessiblemolecular space for perovskite passivation, enabling explorationof previously uncharted chemical regions. 2.3 From AI Design to Laboratory With numerous perovskite passivation molecule candidates avail-able, the next step was to select a manageable subset for laboratoryanalysis. To narrow down the pool, seven filtering criteria wereapplied. Molecules with synthetic accessibility (SA) ≤ 6 wereretained to ensure practical synthesizability, while substructuresknown as Pan-Assay Interference Compounds (PAINS), which5 of 10ive Commons LicenseFIGURE 3 Filtering the generated molecules for laboratory assessment. (a) Schematic illustration of the filtering process based on seven physicochemical criteria. (b) Visualization of the ten agglomerative clusters of the filtered molecules, with one representative molecule randomly selected from each cluster. o  s  a  0  s  p  t  i  a  p  i  t  w  h  a  p  c  s  t  f  TT  b  d  m  t  a  t                          6 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creatften yield false-positive results due to nonspecific reactivity orignal artifacts, were excluded. Hydrogen-bond donor (HBD)nd hydrogen-bond acceptor (HBA) counts were restricted to–2 and 2–5, respectively, to balance interfacial interactiontrength and molecular stability. Molecules with a topologicalolar surface area (TPSA) between 50 and 120 Å2 were selectedo maintain moderate polarity suitable for thin-film compat-bility. Additionally, molecules with energy gaps between 1.5nd 5.0 eV and dipole moments between 1.5 and 4.0 D wererioritized to favor electronic stability and directional surfacenteractions. As illustrated in Figure 3a , this multi-criteria fil-ering retained fewer than 10% of the most desirable candidateshile preserving chemical diversity, yielding approximately 8000igh-quality molecules suitable for PSC passivation. Althoughny molecule from this filtered set could potentially exhibitassivation activity, the molecules were further grouped into tenlusters using an agglomerative clustering algorithm to maximizetructural diversity and minimize redundancy. One represen-ative molecule was then randomly selected from each clusteror further analysis (Figure 3b ), with details summarized inable S2 . ranslating AI-generated molecules into experimentally accessi-le compounds remains a key bottleneck in modern moleculariscovery. Because novelty is an inherent goal of generativeodeling, most generated molecules are new (i.e., absent fromhe training data) and their synthetic feasibility is often uncertain,lthough SA scores offer preliminary guidance. Importantly, sincehe training dataset focuses on domain-relevant molecules ratherof 10than the full chemical universe, generated candidates can occa-sionally coincide with existing compounds known from otherresearch areas. This phenomenon, often termed molecular “re-discovery,” has been widely discussed in drug discovery [ 24, 25 ].In our randomly selected ten candidates, two known compounds:DL-mandelic acid (S9), commonly used in antimicrobial andcosmetic formulations [ 26, 27 ], and 4-maleimidobutyric acid(S10), a reagent for drug and protein conjugation [ 28 ], wereidentified. These two molecules were commercially available andthus directly selected for experimental verification. The thirdselected molecule, S3, appeared to be novel and required customsynthesis. To enable quick validation, a structural similaritysearch was conducted, identifying maleic acid monoamide (Tani-moto similarity = 85%) as a commercially available analog, whichwas adopted as a surrogate for S3. Details including the molecularstructure of all three molecules selected for laboratory assessmentare provided in Table S3 . 2.4 Passivation Effects on PSCs To experimentally verify the effect of the selected three represen-tative molecules, PSCs were fabricated with an inverted config-uration (Scheme 1 ). This structure was chosen because it offersfacile fabrication and is compatible with tandem solar cell archi-tecture, which are key for achieving higher performance. Previ-ous studies have shown that defects, particularly those located atthe perovskite upper surface and the perovskite/electron trans-port layer (ETL) interface [ 29, 30 ], remain the primary factorsAdvanced Science, 2026ive Commons Licensel  t  v  e  a  r  mA  p  m  s  a  c  m  d  a  (  a  b  s  dT  i  t  d  c  d  t  r  u  o  s  d  (  f  F  o  s  s  r  e  w  1  s  s  f  0  r  J  s  a  p  d  a  d  e  r  h                                                        A 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Cimiting the efficiency of inverted PSCs. Therefore, demonstratinghe effectiveness of the selected molecules in real devices providesaluable insight for developing highly efficient PSCs. Note thatxperimental validation presented here should be regarded as proof-of-concept demonstration of AI-generated passivators,ather than implying broad effectiveness across all generatedolecules. s detailed in the Experimental Studies section, a 1 mg/mLassivation solution (a typical concentration for surface treat-ent) was prepared in 2-propanol (IPA) and dynamicallypin-coated on the perovskite surface, followed by annealingt 100◦C for 5 min. The representative current–voltage ( J –V )urves of the PSCs are shown in Figure S5a . Notably, all threeolecules improved the open-circuit voltage ( VOC ), indicatingefect passivation with suppressed nonradiative recombinationt the perovskite surface. Among them, 4-maleimidobutyric acidMBA) yielded the most significant improvement, exhibiting pronounced reduction in hysteresis (efficiency discrepancyetween J –V curves obtained under forward scan and reversecan) compared with both the other passivators and the controlevice. he optimal concentration of MBA was next examined. As shownn Figure S5b , the devices passivated with 1 mg/mL MBA achievedhe highest VOC and PCE, whereas higher concentrations slightlyecreased performance, confirming 1 mg/mL as the optimalondition. These results suggest that further optimization ofeposition parameters could allow other AI-designed passiva-ion molecules to deliver even better performance. To assesseproducibility, additional batches of devices were fabricatednder this optimized condition. In the champion device, PCEsf 24.13% and 23.39% were achieved under reverse and forwardcans, respectively (Figure S6a ). The integrated photocurrentensity derived from the incident photon-to-current efficiencyIPCE) spectrum (Figure S6b ) agrees well with that obtainedrom the J –V measurements, with a minimal mismatch of 2.9%.igure 4a–c summarizes the statistical photovoltaic parametersf PSCs with and without MBA passivation. The control deviceshowed a large difference in VOC between forward and reversecans, consistent with the hysteresis behavior observed in theepresentative J –V curves. In contrast, MBA-passivated devicesxhibited nearly identical VOC values in both scan directions,ith the average VOC (reverse scan) increasing from 1.08 to.12 V. Due primarily to this VOC improvement, along with alight enhancement in fill factor (FF), the average PCE (con-idering both forward and reverse scan directions) increasedrom 19.3% to 22.2%, corresponding to an average ΔPCEnorm of.15, a value comparable to or exceeding that of previouslyeported passivation molecules (see Data T0). Furthermore, the –V hysteresis index ( ℎ𝑦𝑠 𝑡𝑒 𝑟 𝑒 𝑠 𝑖𝑠 𝑖𝑛𝑑 𝑒 𝑥 = 𝑃 𝐶𝐸reverse − 𝑃 𝐶𝐸forward 𝑃𝐶𝐸reverse ) wasignificantly reduced, from 0.160 in control devices to 0.036fter MBA treatment. We further performed maximum poweroint tracking (MPPT) measurements on two representativeevices exhibiting average efficiencies (19.2% for the referencend 22.4% for the passivated device). As shown in Figure S7 ,uring 350 s of continuous tracking, the passivated devicexhibits constant efficiency while the reference one decreasesemarkably. This behavior is consistent with the reduced J –Vysteresis, as stabilized MPPT and suppressed hysteresis sharedvanced Science, 2026reaa common origin in diminished interfacial trapping and ionicmigration. These results indicate that MBA treatment not onlypassivates surface defects but also enhances interfacial carriertransport. To further elucidate the role of MBA at the perovskite interface,perovskite films and devices with and without MBA treatmentwere systematically investigated. The interaction between MBAmolecules and surface Pb2 + sites was examined by X-ray photo-electron spectroscopy (XPS). As shown in Figure S8 , the Pb 4 fcore-level peaks shift toward higher binding energies ( ∼ 0.20 eV)after passivation, indicating strong coordination between thecarboxylate groups of MBA and under-coordinated Pb2 + sites.This interaction modifies the local electronic environment ofPb, directly confirming effective chemical passivation of surfacedefects [ 31 ]. Steady-state photoluminescence (PL) and time-resolved PL (TRPL) measurements were conducted on perovskitefilms deposited on glass substrates to isolate surface effects. Asshown in Figure 4d,e , MBA-passivated films exhibit strongerPL intensity and prolonged carrier lifetimes, confirming sup-pressed nonradiative recombination and reduced trap density.These findings are further supported by transient photovoltage(TPV) measurements (Figure S9a ), where enhanced photovolt-age amplitude and extended decay lifetimes are observed afterpassivation, indicating suppressed interfacial recombination andimproved quasi-Fermi level splitting [ 32 ], consistent with theobserved VOC enhancement. The surface energetics of the perovskite films were charac-terized by UV–vis absorption spectroscopy and photoelectronyield spectroscopy (PYS) (Figure S10a–c ). While the opticalbandgap remains unchanged, the perovskite conduction bandshifts downward from 5.39 to 5.46 eV after MBA treatment(Figure S10c ), which might be attributed to the molecular dipolemoment of MBA [ 33, 34 ]. The electronic properties of the PCBMelectron transport layer were also determined by UV–vis andPYS measurements (Figure S9d–f ). Based on these results, theenergy-level alignment diagram was constructed (Figure 4f ).Consequently, the energy offset between the perovskite andPCBM layers is reduced from 0.07 eV to nearly zero, mini-mizing interfacial energy loss and accounting for the enhancedVOC . Moreover, this favorable alignment promotes faster elec-tron extraction and suppresses interfacial charge accumulation,thereby reducing J –V hysteresis [ 32, 35, 36 ]. Consistently, tran-sient photocurrent (TPC) measurements (Figure S9b ) show fastercurrent decay after passivation, confirming accelerated chargeextraction [ 32 ]. To further support the experimental observations, density func-tional theory (DFT) calculations were performed to examinethe interaction between representative passivation molecules andthe perovskite surface (Figure S11 ). Among the three moleculesconsidered, MBA exhibits the strongest interaction with the FAI-terminated FAPbI3 (001) surface, with an adsorption energy of− 1.462 eV and the largest charge transfer (0.146 e). Charge-densitydifference plots reveal pronounced charge accumulation local-ized around the three oxygen atoms in the terminal functionalgroup of MBA. In contrast, maleic acid monoamide and DL-mandelic acid contain only two terminal oxygen atoms, resultingin weaker charge redistribution and smaller adsorption energies.7 of 10tive Commons LicenseFIGURE 4 Passivation effects of 4-maleimidobutyric acid on PSCs. (a) Open-circuit voltage ( VOC ), (b) power conversion efficiency (PCE), and (c) J –V hysteresis index of PSCs with and without MBA passivation. (d, e) Steady-state and time-resolved photoluminescence (PL and TRPL) spectra of perovskite films with and without passivation. (f) Energy-level diagrams of the perovskite with and without passivation and their alignment with the PCBM electron transport layer. T  M  t  r  pI  d  i  o  f  T  m  g3T  t  i  e  =  m  l  t  r  r  p  t  m  1                               8 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creathese results suggest that the enhanced interaction strength ofBA arises from its molecular functionality and contributeso more effective defect passivation, consistent with previouseports correlating stronger adsorption with improved interfacialassivation [ 37 ]. n summary, combined experimental and theoretical resultsemonstrate that the proposed AI-driven pipeline successfullydentifies effective passivation candidates, such as MBA, capablef reducing surface defects in perovskite films, improving inter-acial charge transfer, and enhancing overall device efficiency.hese findings validate the reliability of the generative AI-assistedolecular design strategy and highlight its strong potential foruiding future PSC optimization.  Conclusions his work establishes an AI-driven framework that accelerateshe discovery of molecular passivators for PSCs. A discrim-native model based on SMILES-X successfully distinguishedffective from ineffective molecules with high accuracy (F1 0.80, ROC-AUC = 0.88), while a GPT-2–based generativeodel iteratively produced over 100 000 chemically valid andargely effective candidates. Integrating these two models, forhe first time, enabled exploration of previously inaccessibleegions of chemical space. Multi-criteria filtering and clusteringeduced the candidates to ∼ 8000 high-quality molecules whilereserving chemical diversity. Among the three experimentallyested molecules, 4-maleimidobutyric acid (MBA) delivered theost significant improvement, increasing the VOC from 1.08 to.12 V and boosting the average power conversion efficiency fromof 1019.3% to 22.2%, with an impressive ΔPCEnorm of 0.15. These resultsvalidate the reliability of the proposed AI-assisted moleculardesign strategy for PSCs. By reducing reliance on trial-and-errorchemistry and enabling exploration of uncharted chemical space,this framework exemplifies how generative AI can complementtraditional design methodologies and shorten the pathway frommolecular concept to functional optoelectronic materials. While the present framework demonstrates effective enrichmentof candidate passivation molecules, several limitations should benoted. The discriminative model is trained on literature-derivedΔPCEnorm values and therefore primarily captures empiricalstructure–performance correlations reflected in reported deviceoutcomes. As such, it may combine transferable chemical motifswith interpolation within chemistries represented in the trainingdata, rather than explicitly learning physics-based descriptors.Incorporating systematic theoretical quantities as additionaltraining labels would further enhance mechanistic interpretabil-ity and generalizability in future developments. In addition,because Data T0 was constructed from published studies, it mayreflect inherent publication bias toward improved passivationstrategies, while fully negative or null results remain underrep-resented. Although the dataset spans a range of performanceoutcomes and includes cases treated as class 0, incorporatingsystematically collected negative samples would sharpen classboundaries and reduce potential label noise. Finally, the presentAI workflow is efficiency-oriented, as it is trained exclusively onefficiency-based labels. The MPPT tracking characterizes opera-tional efficiency rather than ISOS-compliant lifetime evaluation.Future extensions integrating standardized stability metrics astraining objectives would enable multi-objective optimization ofboth efficiency and long-term durability. Advanced Science, 2026ive Commons LicenseAA  t  a  e  p  c  d  vAT  (  (  n  s  (CTDT  s  dR HJ  d2  p3  F  h4  I  D  d5  S  16  CN  07  o  L8  S  19  S  C  11  v                                              A 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creatuthor Contributions .T.N.F., Z.G., and G.L. conceived the study. ATNF designed the computa-ional methods, prepared the code, and performed the data analysis. G.L.ssisted with data analysis and interpretation. Z.G. and R.G. conductedxperimental measurements and T.M. associated data analysis. J.M.erformed DFT calculations and associated data analysis. K.S. and R.F.P.ontributed to data collection and preprocessing. A.T.N.F. and Z.G.rafted the manuscript, and all authors reviewed and approved the finalersion. cknowledgements his work was supported by the Center for Energy Systems DesignCESD), the International Institute for Carbon-Neutral Energy ResearchWPI-I2 CNER), which was established by the World Premier Inter-ational Research Center Initiative (WPI), MEXT, Japan. Additionalupport was provided by the ROIS NII Open Collaborative Research251FP-22667). onflicts of Interest he authors declare no conflicts of interest. ata Availability Statement he data that support the findings of this study are available in theupplementary material of this article. All code used in this study has beeneposited at: https://github.com/adroitfajar/pvmol-gen . eferences 1 . A. Kojima, K. Teshima, Y. Shirai, and T. Miyasaka, “Organometalalide Perovskites as Visible-Light Sensitizers for Photovoltaic Cells,”ournal of the American Chemical Society 131 (2009): 6050–6051, https://oi.org/10.1021/ja809598r .  . NREL, Best Research-Cell Efficiency Chart, 2025, https://www.nlr.gov/v/cell-efficiency.  . Q. Jiang, Y. Zhao, X. Zhang, et al., “Surface Passivation of Perovskiteilm for Efficient Solar Cells,” Nature Photonics 13 (2019): 460–466,ttps://doi.org/10.1038/s41566- 019- 0398-2 .  . R. Zhi, C. Q. Yang, M. U. Rothmann, et al., “Direct Observation ofntragrain Defect Elimination in FAPbI3 Perovskite Solar Cells by Two-imensional PEA2 PbI4 ,” ACS Energy Letters 8 (2023): 2620–2629, https://oi.org/10.1021/acsenergylett.3c00685 .  . J. Xia, C. Liang, H. Gu, et al., “Surface Passivation Toward Efficient andtable Perovskite Solar Cells,”Energy & Environmental Materials 6 (2023):2296, https://doi.org/10.1002/eem2.12296 .  . N. T. P. Hartono, J. Thapa, A. Tiihonen, et al., “How Machine Learningan Help Select Capping Layers to Suppress Perovskite Degradation,”ature Communications 11 (2020): 4172, https://doi.org/10.1038/s41467-20- 17945- 4 .  . C. Zhi, S. Wang, S. Sun, et al., “Machine-Learning-Assisted Screeningf Interface Passivation Materials for Perovskite Solar Cells,” ACS Energyetters 8 (2023): 1424–1433, https://doi.org/10.1021/acsenergylett.2c02818 .  . J. Xu, H. Chen, L. Grater, et al., “Anion Optimization for Bifunctionalurface Passivation in Perovskite Solar Cells,”Nature Materials 22 (2023):507–1514, https://doi.org/10.1038/s41563- 023- 01705-y.  . X. Zhang, B. Ding, Y. Wang, et al., “Machine Learning for Screeningmall Molecules as Passivation Materials for Enhanced Perovskite Solarells,” Advanced Functional Materials 34 (2024): 2314529, https://doi.org/0.1002/adfm.202314529 . 0 . Z. Wu, L. Kang, T. Huang, et al., “Elevating Perovskite Efficiencyia Machine Learning-Assisted Screening of Passivators,” Chemicaldvanced Science, 2026Engineering Journal 499 (2024): 156391, https://doi.org/10.1016/j.cej.2024.156391 . 11 . J. L. Reymond, “The Chemical Space Project,” Accounts of ChemicalResearch 48 (2015): 722–730, https://doi.org/10.1021/ar500432k . 12 . K. Swanson, G. Liu, D. B. Catacutan, A. Arnold, J. Zou, and J. M.Stokes, “Generative AI for Designing and Validating Easily Synthesizableand Structurally Novel Antibiotics,”Nature Machine Intelligence 6 (2024):338–353, https://doi.org/10.1038/s42256-024- 00809- 7 . 13 . C. Zeni, R. Pinsler, D. Zügner, et al., “A Generative Model for InorganicMaterials Design,” Nature 639 (2025): 624–632, https://doi.org/10.1038/s41586- 025- 08628- 5 . 14 . A. Li and C. Yi, “From Data-Driven to Intelligence-Driven: InnovatingResearch Paradigms of Perovskite Solar Cells,” Solar RRL 9 (2025):202500681, https://doi.org/10.1002/solr.202500681 . 15 . C. Chen, A. Maqsood, Z. Zhang, et al., “The Use of ChatGPT toGenerate Experimentally Testable Hypotheses for Improving the SurfacePassivation of Perovskite Solar Cells,” Cell Reports Physical Science 5(2024): 102058, https://doi.org/10.1016/j.xcrp.2024.102058 . 16 . G. Lambard and E. Gracheva, “SMILES-X: Autonomous MolecularCompounds Characterization for Small Datasets Without Descriptors,”Machine Learning: Science and Technology 1 (2020): 025004, https://doi.org/10.1088/2632-2153/ab57f3 . 17 . D.-H. Kim, S.-J. Woo, C. P. Huelmo, et al., “Surface-Binding MolecularMultipods Strengthen the Halide Perovskite Lattice and Boost Lumines-cence,” Nature Communications 15 (2024): 6245, https://doi.org/10.1038/s41467- 024- 49751- 7 . 18 . L. McInnes, J. Healy, N. Saul, and L. Großberger, “UMAP: Uni-form Manifold Approximation and Projection,” Journal of Open SourceSoftware 3 (2018): 861, https://doi.org/10.21105/joss.00861 . 19 . D. S. Wigh, J. M. Goodman, and A. A. Lapkin, “A Review ofMolecular Representation in the Age of Machine Learning,” WIREsComputational Molecular Science 12 (2022): 1603, https://doi.org/10.1002/wcms.1603 . 20 . M. Gao, T. Huang, M. Li, Y. Zhang, L. Wang, and J. Ding, “Machine-Learning-Driven QSPR Models for Energetic Molecules: A Reviewon Safety and Energetic Properties Prediction,” Chemical EngineeringJournal Advances 23 (2025): 100804, https://doi.org/10.1016/j.ceja.2025.100804 . 21 . D. Weininger, “SMILES, A Chemical Language and InformationSystem. 1. Introduction to Methodology and Encoding Rules,” Journal ofChemical Information and Computer Sciences 28 (1988): 31–36, https://doi.org/10.1021/ci00057a005 . 22 . A. T. N. Fajar, G. Lambard, M. A. Islam, B. B. Saha, Z. D. Nurfajrin,and K. Septioga, “Generating Eco-Friendly Ionic Liquids With EnhancedCO2 Solubility Using Language Models,” Artificial Intelligence Chemistry3 (2025): 100089, https://doi.org/10.1016/j.aichem.2025.100089 . 23 . I. Shumailov, Z. Shumaylov, Y. Zhao, N. Papernot, R. Anderson, and Y.Gal, “AI Models Collapse When Trained on Recursively Generated Data,”Nature 631 (2024): 755–759, https://doi.org/10.1038/s41586-024-07566-y. 24 . M. H. S. Segler, T. Kogej, C. Tyrchan, and M. P. Waller, “GeneratingFocused Molecule Libraries for Drug Discovery With Recurrent NeuralNetworks,”ACS Central Science 4 (2018): 120–131, https://doi.org/10.1021/acscentsci.7b00512 . 25 . K. Handa, M. C. Thomas, M. Kageyama, T. Iijima, and A. Bender, “Onthe Difficulty of Validating Molecular Generative Models Realistically: ACase Study on Public and Proprietary Data,” Journal of Cheminformatics15 (2023): 112, https://doi.org/10.1186/s13321- 023- 00781- 1 . 26 . P. Egner, J. Pavlačková, J. Sedlaříková, P. Pleva, P. Mokrejš, and M.Janalíková, “Non-Alcohol Hand Sanitiser Gels With Mandelic Acid andEssential Oils,” International Journal of Molecular Sciences 24 (2023): 3855,https://doi.org/10.3390/ijms24043855 . 27 . S. Dayal, K. D. Kalra, and P. Sahu, “Comparative study of Efficacy andSafety of 45% Mandelic Acid versus 30% Salicylic Acid Peels in Mild-To-9 of 10ive Commons Licensehttps://github.com/adroitfajar/pvmol-genhttps://doi.org/10.1021/ja809598rhttps://www.nlr.gov/pv/cell-efficiencyhttps://doi.org/10.1038/s41566-019-0398-2https://doi.org/10.1021/acsenergylett.3c00685https://doi.org/10.1002/eem2.12296https://doi.org/10.1038/s41467-020-17945-4https://doi.org/10.1021/acsenergylett.2c02818https://doi.org/10.1038/s41563-023-01705-yhttps://doi.org/10.1002/adfm.202314529https://doi.org/10.1016/j.cej.2024.156391https://doi.org/10.1021/ar500432khttps://doi.org/10.1038/s42256-024-00809-7https://doi.org/10.1038/s41586-025-08628-5https://doi.org/10.1002/solr.202500681https://doi.org/10.1016/j.xcrp.2024.102058https://doi.org/10.1088/2632-2153/ab57f3https://doi.org/10.1038/s41467-024-49751-7https://doi.org/10.21105/joss.00861https://doi.org/10.1002/wcms.1603https://doi.org/10.1016/j.ceja.2025.100804https://doi.org/10.1021/ci00057a005https://doi.org/10.1016/j.aichem.2025.100089https://doi.org/10.1038/s41586-024-07566-yhttps://doi.org/10.1021/acscentsci.7b00512https://doi.org/10.1186/s13321-023-00781-1https://doi.org/10.3390/ijms24043855M32  S  G  J  i2  I  03  I  T  M3  M  C  F  23  U  P  13  f  S83  O  S  13  P  C  d3  o  P  j3  H  C  aSA  IS1 21983844, 2026, 36, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202523042 by National Institute For, Wiley Online Library on [01/07/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Ooderate Acne Vulgaris,” Journal of Cosmetic Dermatology 19 (2020): 393–99, https://doi.org/10.1111/jocd.13168 . 8 . M. Pribylova, M. Dvorakova, V. Hanusova, I. Nemethova, L.kalova, and T. Vanek, “Paclitaxel Conjugation With the Analog of theonadotropin-Releasing Hormone as a Targeting Moiety,” Internationalournal of Pharmaceutics 415 (2011): 175–180, https://doi.org/10.1016/j.jpharm.2011.05.072 . 9 . J. Xing, C. Liu, Q. Cao, et al., “Molecular-Perspective Insights Into Topnterface Passivation in Inverted Perovskite Solar Cells,” Small 21 (2025):6495, https://doi.org/10.1002/smll.202506495 . 0 . H. Wang, C. Liu, R. Xu, Y. Zhang, S. De Wolf, and K. Wang, “Uppernterface Engineering Between Perovskite and Electron Transport Layeroward Efficient and Stable Inverted Perovskite Solar Cells,” Advancedaterials 38 (2025): 13633, https://doi.org/10.1002/adma.202513633 . 1 . Z. Guo, S. Zhao, N. Shibayama, A. Kumar Jena, I. Takei, and T.iyasaka, “A Universal Method of Perovskite Surface Passivation forsPbX3 Solar Cells With VOC Over 90% of the S-Q Limit,” Advancedunctional Materials 32 (2022): 2207554, https://doi.org/10.1002/adfm.02207554 . 2 . Q. Jiang, L. Zhang, H. Wang, et al., “Enhanced Electron Extractionsing SnO2 for High-Efficiency Planar-Structure HC(NH2 )2 PbI3 -Basederovskite Solar Cells,” Nature Energy 2 (2017): 16177, https://doi.org/10.038/nenergy.2016.177 . 3 . J. Wang, L. Zheng, H.-B. Kim, et al., “Dipolar Carbazole Ammoniumor Broadened Electric Field Distribution in High-Performance Perovskiteolar Cells,” Journal of the American Chemical Society 147 (2025): 8663–671, https://doi.org/10.1021/jacs.4c18074 . 4 . Y. Peng, Y. Chen, J. Zhou, et al., “Enlarging Moment and Regulatingrientation of Buried Interfacial Dipole for Efficient Inverted Perovskiteolar Cells,” Nature Communications 16 (2025): 1252, https://doi.org/10.038/s41467-024-55653-5 . 5 . T. Chen, Z. Sun, M. Liang, and S. Xue, “Corr elating Hyster esishenomena With Interfacial Charge Accumulation in Perovskite Solarells,” Physical Chemistry Chemical Physics 22 (2020): 245–251, https://oi.org/10.1039/C9CP05381F . 6 . S. Ravishankar, S. Gharibzadeh, C. Roldán-Carmona, et al., “Influencef Charge Transport Layers on Open-Circuit Voltage and Hysteresis inerovskite Solar Cells,” Joule 2 (2018): 788–798, https://doi.org/10.1016/j.oule.2018.02.013 . 7 . T. Wu, T. B. Raju, J. Shang, et al., “Lattice Matching Anchoring ofole-Selective Molecule on Halide Perovskite Surfaces for n-i-p Solarells,” Advanced Materials 37 (2025): 2414576, https://doi.org/10.1002/dma.202414576 . upporting Information dditional supporting information can be found online in the Supportingnformation section. upporting File: advs75128-sup-0001-SuppMat.pdf. 0 of 10 Advanced Science, 2026nline Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttps://doi.org/10.1111/jocd.13168https://doi.org/10.1016/j.ijpharm.2011.05.072https://doi.org/10.1002/smll.202506495https://doi.org/10.1002/adma.202513633https://doi.org/10.1002/adfm.202207554https://doi.org/10.1038/nenergy.2016.177https://doi.org/10.1021/jacs.4c18074https://doi.org/10.1038/s41467-024-55653-5https://doi.org/10.1039/C9CP05381Fhttps://doi.org/10.1016/j.joule.2018.02.013https://doi.org/10.1002/adma.202414576 Generative AI-Driven Accelerated Discovery of Passivation Molecules for Perovskite Solar Cells 1 | Introduction 2 | Results and Discussion 2.1 | Existing Passivation Molecules 2.2 | Expanding the Chemical Space 2.3 | From AI Design to Laboratory 2.4 | Passivation Effects on PSCs 3 | Conclusions Author Contributions Acknowledgements Conflicts of Interest Data Availability Statement References Supporting Information