Abstract

The commercialization of perovskite solar cells (PSCs) is bottlenecked by inefficient trial-and-error approaches reliant on human expertise in both materials discovery and device fabrication 1 , 2 , 3 . Here we introduce an autonomous closed-loop framework that integrates machine learning (ML)-driven materials discovery with an automated manufacturing platform. The system uses active learning and quantum modelling to rapidly identify high-performance molecules and the platform uses Bayesian optimization and symbolic regression in a feedback loop to continuously refine the fabrication process. This integrated approach enabled the discovery of a passivation molecule, 5-(aminomethyl)nicotinonitrile hydroiodide (5ANI), which yielded 0.05-cm 2 solar cells with a power conversion efficiency (PCE) of 27.22% (certified maximum power point tracking (MPPT) efficiency of 27.18%) and 21.4-cm 2 mini-modules with a PCE of 23.49%. Moreover, the devices exhibited long-term operational stability, retaining 98.7% of their initial efficiency after 1,200 h of continuous operation under the ISOS-L-1I protocol. Crucially, the automated platform achieved an efficiency reproducibility nearly five times that of manual fabrication. This work establishes an automated closed-loop system that synergizes ML-powered discovery with the high-fidelity data from automated manufacturing, setting a benchmark for autonomous discovery and manufacturing in photovoltaics and materials.

Main

PSCs represent a pivotal next-generation photovoltaic technology capable of achieving high PCEs with cost-effective, solution-processed active layers 4 , 5 , 6 , 7 , 8 . However, their development remains critically hindered by human-expertise-dependent trial-and-error fabrication and optimization. This approach results in inefficient materials discovery, poor device reproducibility and inconsistent manufacturing yield, impeding the lab-to-fab transition 9 , 10 , 11 , 12 , 13 , 14 . ML-driven autonomous experimentation offers a transformative pathway to address these issues through closed-loop materials discovery and automated fabrication 15 , 16 , 17 , 18 , 19 , 20 . Although ML models with physics-informed learning show promise for accelerated discovery, a unified framework fully integrating interpretable ML-guided materials design with autonomous device optimization remains unexplored 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 .

Here we introduce an autonomous closed-loop research framework that integrates ML-driven materials discovery with automated device manufacturing to enable the reproducible fabrication of state-of-the-art PSCs performance. On the discovery side, the system uses interpretable ML algorithms, coupled with density functional theory (DFT) calculations, to rapidly screen more than 20,000 chemical species. This process established quantitative structure–property–performance relationships for targeted passivation molecules to improve the PCEs of PSCs. On the automated manufacturing side, the platform uses Bayesian optimization and symbolic regression in a feedback loop to continuously refine the fabrication process. This integrated approach guided the discovery of a passivation molecule, 5ANI, and autonomously optimized its application in the fabrication workflow. This enabled the fabrication of 0.05-cm 2 solar cells with a PCE of 27.22% (certified MPPT efficiency of 27.18%) and 21.4-cm 2 mini-modules with a PCE of 23.49%, demonstrating substantial scalability. Moreover, the resulting devices exhibited long-term operational stability, retaining 98.7% of their initial efficiency after 1,200 h of continuous operation under the ISOS-L-1I protocol. Underscoring its process reproducibility, the automated platform achieved an efficiency reproducibility nearly five times that of manual fabrication. This high-precision process control is critical for generating the high-fidelity data necessary for an iterative ML model. Ultimately, this work establishes a broadly generalizable and transferable ML-driven, closed-loop model that accelerates the entire pipeline from materials discovery and screening to high-reproducibility manufacturing, setting a benchmark for automated research in photovoltaics.

Closed-loop automated discovery workflow

Figure 1 illustrates our integrated workflow combining chemical space exploration, construction of an interpretable ML model for PCEs and a closed-loop automated manufacturing platform for the discovery and implementation of optimized passivating molecules in PSCs. To move beyond incremental optimization of a few well-known aromatic amines and to systematically uncover design principles that generalize across the chemical space, we developed a two-stage RDKit-based 31 workflow (Fig. 1a). Starting from phenylalkylamine scaffolds, the algorithm systematically substitutes aromatic hydrogens with nine functional groups (-CN, -CF 3 , -CH 3 , -OCH 3 , -COOH, -OH, -F, -Cl, -Br), generates isomers and, finally, sanitizes the selection to ensure chemical validity 9 , 32 . The generated molecules undergo geometry optimization, resulting in a comprehensive library of 18,264 compounds that define the chemical space for exploration (Supplementary Fig. 1 and Supplementary Note 1). Figure 1b illustrates the ML-driven workflow for interfacial materials exploration to improve the PCEs. To train the interpretable ML model, the initial training set, consisting of 20 known passivating molecules from the literature (Supplementary Fig. 2), was systematically evaluated by combining DFT calculations with experimentation to extract structure–property relationships that correlate strongly with device performance. The structure–property relationship for the PCEs of PSC devices constructed by the ML model is applied to the virtual library of compounds and accurately predict device performance with the various aromatic amine derivatives. Finally, four predicted high-performing passivation molecules were synthesized and experimentally validated. Figure 1c presents the closed-loop automated manufacturing platform using Bayesian optimization to iteratively optimize device efficiency through a fabrication, testing and experimental planning feedback loop 33 , 34 . The precision afforded by the rigorous automated process control generates high-fidelity data, a fundamental prerequisite for the success of ML-driven materials discovery. Building on these high-quality data, symbolic regression is then used to establish robust correlations between perovskite film fabrication conditions, passivation layer preparation strategies and device performance. The analysis results are fed back in real time to optimize subsequent fabrication parameters, enabling fully automated, closed-loop experimentation.

a, Exploration of chemical space through molecular design, integrating key features such as molecular weight, carbon rings, substitution and carbon chain length. b, AI-based active learning cycle for interface passivation. The process begins with an initial training set established by high-throughput DFT calculations, followed by a closed-loop sequence of model training, sample acquisition, property labelling and iterative dataset expansion to identify optimal passivation materials. c, Schematic of the closed-loop automated fabrication platform making use of Bayesian optimization and symbolic regression for integrated device fabrication, efficiency testing and process optimization.

Interfacial materials discovery through ML

Traditional materials discovery models for interfacial passivators in perovskite research predominantly rely on chemical intuition and trial-and-error experimentation, which not only constrain the accessible chemical space and result in exceedingly low throughput but also further undermine the scientific interpretability of the discovery process. Critically, passivation materials potentially offering performance may remain overlooked, even in heavily studied molecular classes such as phenylalkyl amines, as shown below.

We developed an interpretable ML-driven framework for perovskite passivation materials design (Fig. 2a), combining high-throughput calculations and an automated experimental platform. The optimization of PCEs in PSCs involves the synergistic interplay of several complex processes, encompassing photon absorption, charge transport, carrier recombination and extraction. These factors are all strongly affected by the choice of interfacial passivation material. To explain the intrinsic physical mechanisms underlying interfacial passivation, the explainable ML algorithm, established using the sure independence screening and sparsifying operator (SISSO), was used to quantitatively predict the PCE of PSCs across the chemical space 35 . This symbolic regression technique is based on sparsification operators and independence screening principles, capable of autonomously discovering complex mathematical patterns embedded in data without previous functional form constraints, generating analytical expressions with clear physical importance. The training dataset comprises PCE measurements from 20 sets of PSC samples fabricated under standardized experimental conditions (Supplementary Figs. 3–27 and Supplementary Table 1). Each sample was characterized by six atomic descriptors derived from first-principles DFT calculations, encompassing both molecular properties (highest occupied molecular orbital energy (HOMO), lowest unoccupied molecular orbital energy (LUMO), dipole moment (P m ) and molecular volume (V m )) and interfacial properties of the passivated perovskite layer (for example, work function (E WF ) and formation energy (E c )); more details are provided in Supplementary Note 2. SHapley Additive exPlanations (SHAP) analysis 36 was used to analyse the relative importance of the respective descriptors in predicting PCE (Fig. 2b). The results reveal that the V m and the E WF correlate most strongly with the attainable PCE. The correlation heatmap shows low pairwise correlation among descriptors, indicating high orthogonality (Fig. 2c). The final PCE prediction model was established through SISSO model training with feature space reconstruction based on key atomic descriptors identified by means of SHAP analysis (detailed methodology provided in Supplementary Note 3). The SISSO algorithm systematically searched through this reconstructed space to generate mathematically simple, interpretable expressions that best map the descriptors to the experimental PCE values. The optimal expression equation (1) was selected as the final model because it minimizes prediction error while capturing the coupled effects of molecular and interfacial properties on device performance.

The first term (α 1 = −15.020) shows that the E WF /HOMO ratio, exponentially scaled by E c 2 , negatively correlates with PCE. This complex interaction suggests that optimal E WF tuning relative to HOMO levels requires careful balance with the formation energy E c . The second term (α 2 = −0.466 × 10 −3 ) reveals a small inverse relationship between the squared P m relative to LUMO energy, modulated by work function E WF . Although dipolar passivation molecules typically enhance performance through built-in electric fields that suppress interfacial recombination, this term suggests that, when dipole strength becomes disproportionate to the electron-accepting capability of the molecule (LUMO), it can induce unfavourable electrostatic perturbations at the perovskite interface. The exponential E WF dependence indicates that this effect is amplified in high-work-function regimes, in which strong dipoles may disrupt optimal band alignment and create localized trap states. The third term (α 3 = 4.320) shows that larger V m combined with more favourable (larger) E c enhances PCE. This reflects that thermodynamically favoured passivation interactions with appropriate molecular size promote better surface coverage and defect healing. The constant term C = 25.887 represents the baseline PCE contribution from intrinsic perovskite properties and device architecture independent of the passivating molecule’s properties. The formula indicates that, although individual properties such as P m and E c generally improve passivation, their complex nonlinear interdependencies create optimization landscapes that require balanced molecular design strategies. Finally, as shown in Fig. 2f, the SISSO-derived PCE prediction model demonstrates agreement with experimental data, achieving a coefficient of determination R 2 = 0.921. We also performed t-tests on all estimated parameters (α 1 , α 2 , α 3 ) and the constant term (C) in equation (1) to verify that the coefficients are very different from zero. The results in Supplementary Table 2 show that all estimated parameters exhibit high t-statistics and P-values well below the standard threshold of 0.05. This confirms that every descriptor term in equation (1) makes a statistically significant contribution to the prediction of PCE. We also benchmarked SISSO against widely used off-the-shelf algorithms. As shown in Supplementary Fig. 28, although these models can fit the training data, they performed poorly (R 2 < 0.15) owing to the limited sample size (N = 20). We also examined simple linear relationships between the target property (PCE) and individual molecular features (for example, dipole moment). These univariate analyses showed poor correlation (R 2 < 0.5), confirming that the passivation mechanism is governed by the complex interplay of several physicochemical properties rather than any single factor. Here equation (1) should be considered an interpretable regression model rather than a general physical law. The analytical form identified by SISSO is not singular, as different subsets of descriptors or alternative training protocols can yield comparably predictive relations. Its value lies in providing a compact representation of how molecular and interfacial descriptors jointly influence device PCE within the chemical space explored, thereby offering an interpretable surrogate that guides understanding and materials design.

a, Symbolic-regression-derived functional relationship between molecular and interfacial passivation properties and device PCE, showing both the schematic representation and analytical results. b, SHAP analysis quantifying the relative importance of interfacial passivation properties (green) and molecular properties (grey) on the PCE of PSCs. Abbreviations explained in the main text. c, Heatmap showing pairwise Pearson correlation coefficients (r) between descriptors in the training set of PCEs, with colour intensity representing correlation strength (r-values ranging from −1 to +1). d, e, Active learning model performance in predicting formation energy and work function for perovskite surfaces passivated with different molecular species across the chemical space. Grey and green points represent training data and validation data, respectively. f, Comparison between PCE function predictions and experimentally measured PCE, demonstrating high accuracy with R 2 = 0.921. g, Statistical distribution of predicted PCE values for interface passivation across the chemical space using the interpretable PCE ML model. h, Violin plot illustrating the relationship between different functional groups and predicted PCE. The width of each shaded area represents the probability density of the data at different values. Inside each violin, the white line indicates the median PCE, the thick solid bar represents the interquartile range and the thin black lines (whiskers) extend to the interquartile range to show the rest of the distribution. i, Hierarchical screening approach for identifying candidates with high synthetic feasibility and high PCE potential from the chemical space of 18,264 potential molecules. RMSE, root mean square error.

The prediction model was subsequently applied to explore the entire chemical space with different molecular candidates. Molecular properties (HOMO, LUMO, P m and V m ) for all 18,264 candidate molecules were obtained using high-throughput first-principles calculations (detailed methodology provided in Supplementary Note 4), with statistical distributions shown in Supplementary Fig. 29. Given the computational complexity of first-principles calculations for perovskite systems with diverse molecular passivation layers, an active learning framework based on the NGBoost algorithm 37 was developed to reduce computational cost (Supplementary Fig. 30; detailed methodology is provided in Supplementary Note 5). Compared with conventional first-principles calculations, which typically require about two days to evaluate the properties of a single material, the trained ML model can predict the properties of about 18,000 candidate materials within minutes. This framework uses first-principles calculations for 212 representative perovskite–molecule systems to determine E WF and E c (Supplementary Fig. 31 and Supplementary Table 3) before predicting the respective properties for all remaining candidates in the chemical space. The predictive models demonstrated performance with R 2 = 0.920 for E WF and R 2 = 0.954 for E c (Fig. 2d,e).

To mitigate the issues of dimensionality and overfitting inherent in ML approaches, feature engineering was used to optimize an initial feature set comprising 130 descriptors derived from first-principles calculations and RDKit computations (detailed methodology outlined in Supplementary Note 6, Supplementary Table 4 and Supplementary Fig. 32). Feature importance analysis revealed that P m were the most critical descriptors in predicting E c , whereas the HOMO–LUMO gap and dipole moment dominated E WF predictions (Supplementary Figs. 33 and 34). The distributions of NGBoost-predicted E WF and E c are presented in Supplementary Fig. 35.

The resulting PCE distribution is shown in Fig. 2g, with functional-group-specific performance comparisons presented as violin plots in Fig. 2h. The analysis reveals that pyridine-containing molecules and cyano (-CN)-functionalized compounds exhibit the highest predicted PCEs, suggesting that these chemical motifs are particularly promising for perovskite passivation applications 38 , 39 , 40 . Beyond confirming expected trends, the model enables systematic comparison across a large number of molecules, highlighting motifs with the greatest potential for experimental validation. Having determined the key molecular properties of high-performing passivating materials, we used a step-by-step screening protocol to identify synthetically tractable passivators (Fig. 2i). Numerical optimization results indicate that the maximum PCE is attained at the boundaries of the parameter space, with pronounced sensitivities to E WF and E c . These findings substantiate the use of E WF and P m as effective screening criteria for identifying high-PCE materials, in which P m serves as a robust predictor of E c . Consequently, the descriptors P m and E WF in the SISSO formula were selected as the screening criteria, screening targeted molecules with P m and E WF values exceeding those of the phenylethylammonium iodide (PEAI), which was established as the reference. This approach identified 57 top-performing candidates from the molecular library (Supplementary Table 5). From this refined set, four compounds were selected for experimental validation: 4AEI, 4AMI, 3-(aminomethyl)benzonitrile hydroiodide (3AMI) and 5ANI. The selection of four materials for experimental validation was based on the high predicted efficiency and rationalizable presence of strongly electron-withdrawing substituents. First, these compounds exhibited the top device performance among the 57 screened candidates. Second, Fig. 2h demonstrates that cyano (-CN)-functionalized and fluorine (-F)-functionalized compounds achieve the highest PCEs, providing a rational basis for their prioritization in the validation study. First-principles calculations of the selected molecules confirm their interfacial properties, with electrostatic potential maps revealing distinct charge distributions (Supplementary Fig. 36). Detailed structural analysis of FAPbI 3 (100) surface passivation, including band structure, projected density of states and work function modifications, validates the predicted enhancement in electronic properties (Supplementary Figs. 37 and 38).

Automated fabrication and optimization

Figure 3 illustrates the device structure, process optimization workflow and performance analysis of PSCs fabricated using the automated platform (Supplementary Fig. 39), as discussed previously. The device structure, shown in Fig. 3a, consists of a glass substrate, indium tin oxide (ITO) electrode, self-assembled monolayer (SAM), perovskite active layer, passivation layer, electron transport layer and top electrode. The optimization of the manufacturing process on the fully automated PSC fabrication platform follows the closed-loop Bayesian optimization logic illustrated in Fig. 3b and Supplementary Fig. 40 and detailed in Supplementary Note 7. This logic uses a Gaussian process model as a probabilistic surrogate for the device efficiency, which in turn guides a q-expected improvement acquisition function to select the next batch of experiments. This iterative process efficiently navigates the multidimensional parameter space to maximize the PCE. Through a combination of automated experiments (Supplementary Video 1), training data collection and generation of new experimental conditions, the platform achieves iterative optimization of the fabrication process for PSCs. Figure 3c illustrates the relationship between the batch number and the PCE of unpassivated devices during the optimization of automated PSC fabrication conditions. Through nine rounds of process optimization (Supplementary Table 6), the performance of devices based on non-passivated perovskite films was greatly enhanced, achieving a PCE of 23.87%.

a, Schematic diagram of the PSC device structure. b, Workflow logic of process optimization in the closed-loop automated manufacturing platform based on Bayesian optimization. c, Relationship between batch number and device performance of iteratively optimized automated PSC manufacturing without surface passivation. d, Relationship between batch number and device performance of iteratively optimized automated PSC manufacturing featuring 3AMI and 5ANI passivation layers. e, J–V curves of best-performing PSCs fabricated with three different passivation materials. f, J–V curves (forward and reverse scans) of the champion device fabricated with 5ANI passivation. g, J–V curves (forward and reverse scans) of the mini-module (active area: 21.4 cm 2 ) fabricated with 5ANI passivation. h, Comparison of reproducibility between traditional manual manufacturing and automated platform fabrication of 5ANI-passivated PSCs across several batches. The central line represents the median, the box bounds indicate the 25th and 75th percentiles and the whiskers denote the maximum and minimum values. n = 24 independent devices were tested for each condition. i, Long-term operational stability of PSCs fabricated with three different passivation materials under the ISOS-L-1I protocol.

Guided by ML-driven materials identification, we successfully designed and synthesized several previously unreported passivation molecules, including 3AMI and 5ANI (as shown in Supplementary Figs. 41–52 and detailed in Supplementary Note 8). On the basis of the control device with an efficiency of 23.87%, the process optimization of the passivation layer was then conducted on the automated device fabrication platform. After several rounds of surface passivation optimization, the devices passivated with PEAI, 3AMI and 5ANI exhibited concurrent improvements in open-circuit voltage (V OC ) and fill factor (FF), achieving PCEs of 25.75%, 26.63% and 27.22%, respectively (Supplementary Table 7). Devices using 5ANI as the passivation material demonstrated greater potential for PCE improvement (Fig. 3d,e, Supplementary Figs. 53 and 54 and Supplementary Table 8). Figure 3f shows the champion device fabricated with 5ANI passivation, which exhibits forward and reverse scan consistency, with PCEs of 26.98% and 27.22%, respectively. For the reverse scan, the device achieved a V OC of 1.196 V, FF of 86.70% and short-circuit current density (J SC ) of 26.25 mA cm − 2 . The certified device MPPT efficiency is 27.18% (Supplementary Figs. 55 and 56 and Supplementary Table 9). The external quantum efficiency (EQE) spectra of devices with different surface passivation are presented in Supplementary Fig. 57, showing similar integrated currents. As shown in Supplementary Fig. 58, the V OC and FF of the device substantially improved after 5ANI passivation and the distribution of all parameters became more uniform.

To validate the feasibility of this passivation approach in large-area modules, a mini-module (active area: 21.4 cm 2 ) was fabricated with 5ANI passivation, achieving PCEs of 23.07% and 23.49% for the forward and reverse scans, respectively, as shown in Fig. 3g, Supplementary Fig. 59 and Supplementary Table 10. These results demonstrate the feasibility of the passivation material and process for large-area module fabrication. The reproducibility of device performance across several experimental batches between traditional manual fabrication and the automated platform using 5ANI passivation is compared in Fig. 3h. The fabrication reproducibility was quantitatively compared using the coefficient of variation (CV; defined as standard deviation/mean × 100%). The manual process yielded a high CV of 4.99% (1.29/25.84), which was greatly reduced to just 1.05% (0.28/26.68) on the automated platform. This nearly five times reduction in CV underscores the process consistency and reproducibility achieved through automation (Supplementary Fig. 60 and Supplementary Table 11). It is precisely this level of reproducibility that ensures the generation of high-fidelity data—a fundamental prerequisite for effective, continuous AI-driven device optimization. Figure 3i shows the long-term operational stability of PSCs fabricated with different passivation materials, tested under the ISOS-L-1I protocol. The results demonstrate that devices with 5ANI passivation only degraded by 1.3% of their initial efficiency after 1,200 h of continuous operation and demonstrated the best efficiency retention rate under conditions of damp heat (85 °C and 85% relative humidity) storage (as shown in Supplementary Fig. 61), further validating the passivation performance and practical applicability of the material.

Explaining passivation mechanisms

To further explain the underlying mechanism for the improved performance, we analysed the physical properties of both perovskite films and devices following surface passivation. As presented in Fig. 4a and Supplementary Tables 12–14, quasi-Fermi-level splitting (QFLS) statistics were obtained for perovskite films in various device configurations, including pristine films and those interfaced with electron and hole transport layers. For the control samples, the QFLS values were determined to be 1.196 eV for glass/polyvinylcarbazole (PVK), 1.188 eV for glass/ITO/SAM/PVK and 1.162 eV for glass/ITO/SAM/PVK/C 60 . On treatment with PEAI, 3AMI and 5ANI, a progressive enhancement in QFLS was observed across all configurations, which indicates a marked suppression of non-radiative recombination losses owing to surface passivation.

a, QFLS statistics of perovskite films in different device configurations under 1-sun-equivalent excitation. Glass/PVK, glass/ITO/SAM/PVK and glass/ITO/SAM/PVK/C 60 under different surface treatments. b, EL for devices with different passivation molecules. c–e, Optical PCE maps under 1-sun-equivalent excitation of PSCs treated with different passivation molecules. Optical PCE is defined as the product of maximum power point voltage (V MPP ) and charge extraction efficiency at MPP (Φ PL (V MPP )) ascertained from voltage-dependent broadband PL measurements. Scale bars, 20 μm. f, Corresponding optical PCE distributions of differently passivated devices. g, Statistical analysis of the contact potential distribution of perovskite films with different passivation materials obtained by KPFM. h, tDOS extracted from thermal admittance spectroscopy for different devices.

This surface passivation did not form a 2D perovskite phase, as demonstrated by 2D grazing-incidence wide-angle X-ray scattering (Supplementary Fig. 62), but acted as a surface-specific passivation layer that may exist in the form of a thin molecular interface. Notably, perovskite films treated with 5ANI exhibited the highest QFLS values, which are 1.231 eV, 1.223 eV and 1.204 eV for the three respective configurations, highlighting the passivation capability and its pronounced effect on carrier recombination dynamics. The effectiveness of the passivation against contact-induced losses was quantified by the QFLS drop (ΔQFLS) on C 60 deposition. The 5ANI-treated film exhibited the smallest ΔQFLS of 19 meV, greatly outperforming the PEAI-treated (22 meV) and control (26 meV) samples (Supplementary Tables 10 and 11). This minimal loss demonstrates the ability of 5ANI to protect the perovskite surface and mitigate recombination pathways introduced by the electron transport layer 41 , 42 , 43 , 44 . Figure 4b and Supplementary Table 15 show electroluminescence (EL) measurements on eight statistically independent devices for each treatment condition. The average EL values for the control, PEAI-treated, 3AMI-treated and 5ANI-treated devices were 3.2%, 8.8%, 10.7% and 11.2%, respectively, further confirming that 5ANI treatment effectively reduces non-radiative losses, thus facilitating more efficient photon emission under electrical injection. Optical PCE maps, using the spatial photoluminescence (PL) response of the device as a function of bias, was used to assess changes in both recombination losses and charge extraction 41 , 45 .

Figure 4c–e presents the optical PCE maps for PSCs passivated with PEAI, 3AMI and 5ANI. These maps show that devices treated with 3AMI and 5ANI not only exhibit a higher mean optical PCE than the PEAI control but also have substantially more uniform spatial distributions. This enhanced homogeneity is quantitatively captured in the PCE histograms shown in Fig. 4f. By defining the ‘optical PCE disorder’ as the full width at half maximum of the distribution 41 , we can see that the full width at half maximum for 3AMI (0.026) and 5ANI (0.020) is much smaller than that of the PEAI control (0.116). These results therefore confirm that the molecular passivation strategies yield films with much lower optical PCE disorder. To further examine the interfacial electronic properties after molecular passivation, Kelvin probe force microscopy (KPFM) measurements were conducted to map the surface potential of perovskite films following the different surface treatments (Supplementary Fig. 63). A clear decrease in surface work function was observed following PEAI, 3AMI and 5ANI treatment (Fig. 4g). Assuming that the perovskite bandgap remains unchanged, this reduction suggests a favourable shift in energy alignment that promotes electron extraction from the perovskite to C 60 , while suppressing recombination between electrons in C 60 and holes in the bulk perovskite after Fermi-level alignment. Moreover, the 5ANI-treated film exhibited the most uniform surface potential distribution, greatly outperforming the PEAI-treated and 3AMI-treated counterparts. This homogeneity is further corroborated by its more uniform PL emission (Supplementary Fig. 64) and optical PCE response. Further evidence of effective passivation is provided by thermal admittance spectroscopy, which shows a reduced trap density of states (tDOS). Specifically, the tDOS data for the 5ANI-based device (Fig. 4h) indicate a marked suppression of deep-level traps (0.4–0.5 eV). This demonstrates the ability of 5ANI to neutralize performance-limiting defects, which are a primary source of recombination losses in PSCs.

Conclusion

In conclusion, we have developed an interpretable, ML-driven, closed-loop framework that synergizes ML algorithms with automated manufacturing. This platform accelerates the development of materials and devices while minimizing reliance on manual expertise. By explaining atomic-level structure–property relationships, the interpretable model bridges the gap between theory and experiment, enabling the rational design of high-performance surface passivators. These design principles—integrating both molecular and interfacial properties—reveal clear pathways to substantially enhance PCEs. Ultimately, this work establishes a generalizable pattern for integrating intelligent algorithms with automated experimentation, unlocking frontiers in photovoltaics and materials research.

Methods

Materials

Caesium iodide (CsI) and formamidinium iodide (FAI) were purchased from Dysol. Lead iodide (PbI 2 ) and (4-(7H-dibenzo[c,g]carbazol-7-yl)butyl)phosphonic acid (4PADCB) were purchased from TCI. 4-(2-aminoethyl)benzonitrile, 4-(aminomethyl)benzonitrile, 3-(aminomethyl)benzonitrile, 5-(aminomethyl)nicotinonitrile hydrochloride and hydriodic acid (HI; 55–58%) were purchased from Adamas (Shanghai Titan Scientific Co., Ltd.). C 60 , 4-fluoro-phenylammonium iodide (p-F-PEAI), methylammonium iodide (MAI), methylammonium chloride (MACl), propylammonium chloride (PACl), 2-phenylethylammonium iodide (PEAI), 3-(3-fluorophenyl)ethylammonium iodide (m-F-PEAI), 4-(4-fluorophenyl)ethylammonium iodide (p-F-PEAI), 2-(2-fluorophenyl)ethylammonium iodide (o-F-PEAI), 4-(trifluoromethylphenyl)ethylammonium iodide (p-CF 3 PEAI), 4-(chlorophenyl)ethylammonium iodide (p-ClPEAI), 4-(fluorophenyl)methylammonium iodide (p-F-PMAI), 4-(methoxyphenyl)ethylammonium iodide (p-MeO-PEAI), 4-(methylphenyl)ethylammonium iodide (p-MePEAI), 1,4-phenylenediethylammonium diiodide (PDADI), piperidinium iodide (PipI), phenylmethylammonium iodide (PMAI), 4-(trifluoromethylphenyl)methylammonium iodide (p-CF 3 PMAI), 3-(aminomethyl)pyridinium iodide (3-PyAI), 5-aminopentanoic acid (5-APA), octylammonium iodide (OAI), oleylammonium iodide (OAmI) and bathocuproine (BCP) were purchased from Xi’an Polymer Light Technology Co., Ltd. 2-naphthalen-2-ylethanamine hydrochloride (NEAI) and (1S)-1-naphthalen-1-ylethanamine (S(1-1)NEA) were purchased from Shanghai Macklin Biochemical Technology Co., Ltd. The solvents, including dimethylformamide (DMF), dimethyl sulfoxide (DMSO), isopropanol (IPA) and chlorobenzene (CB), were purchased from J&K and used as received. Methanol was purchased from TCI. High-purity silver was purchased from commercial sources. 1.1-mm glass substrates patterned with ITO (15 Ω sq −1 ) were received from Mishi Tech. Co., Ltd. Detailed information on the raw materials is summarized in Supplementary Table 16.

Initial experimental process: manual operation fabrication of PSCs

Glass/ITO substrates (15 Ω sq −1 ) were sequentially cleaned by sonication with detergent (detergent:deionized water = 1:2 v/v), deionized water, acetone and isopropyl alcohol for 20 min, respectively. The substrates were thoroughly rinsed with deionized water after the detergent sonication step. Then the glass/ITO substrates were dried at 100 °C in an oven, treated with oxygen plasma for 10 min and finally transferred into a N 2 -filled glovebox before use. To ensure surface activity, the SAM coating was completed within 90 min after plasma treatment; otherwise, the substrates were re-treated.

4PADCB was dissolved in methanol with an optimized concentration of 0.4 mg ml −1 . The as-prepared SAM solution was dropped on the ITO substrate and allowed to spread for 5 s. Subsequently, the substrate was accelerated at 2,000 rpm s − 1 to a spin speed of 4,000 rpm for 30 s. The substrate was immediately transferred to a hotplate for annealing at 100 °C for 10 min.

For perovskite precursor preparation, a 1.52 M precursor solution was prepared by mixing 19.7 mg of CsI, 12.1 mg of MAI, 235.3 mg of FAI and 721.8 mg of PbI 2 (3% excess PbI 2 ) in 1 ml DMF:DMSO (4:1 v/v) to form a Cs 0.05 FA 0.90 MA 0.05 PbI 3 solution. The precursor solutions were prepared in a dedicated, nitrogen-filled glovebox, maintaining identical environmental conditions to those of the fabrication glovebox (O 2 < 5 ppm, H 2 O < 0.5 ppm). On completion of the formulation, the solutions were hermetically sealed and transferred to the fabrication glovebox. All perovskite precursor solutions were used within 96 h of preparation. Immediately before spin-coating, 15 mol% MACl, 3.5 mol% PACl and 0.2 mol% OAmI were added and the mixture was stirred for 15 min at room temperature using a 4-mm magnetic stir bar at 600 rpm. The additive-containing precursor was used within 3 h.

100 μl of the perovskite solution was dropped onto the ITO substrate and kept for 5 s. The substrate was then spin-coated at 1,000 rpm for 10 s (acceleration: 1,000 rpm s − 1 ), followed by 5,000 rpm for 40 s (acceleration: 2,000 rpm s − 1 ). 300 μl of CB was dripped onto the centre of the film 10 s before the end of spin-coating. After spinning, the films were immediately transferred to a hotplate and annealed at 100 °C for 45 min, followed by 125 °C for 5 min and then naturally cooled to room temperature.

For surface passivation, materials (including PEAI and others) were individually dissolved in a mixed solvent of IPA:DMF (150:1 v/v) at 2 mg ml −1 . The solutions were stirred for 2 h at 50 °C using 4-mm stir bars at 600 rpm and then stored at room temperature before use. 100 μl of the passivation solution was dynamically spin-coated onto the perovskite films after the substrate reached 5,000 rpm (acceleration: 2,000 rpm s − 1 ) for 30 s. The samples were transferred to a hotplate for annealing at 100 °C for 5 min and finally cooled to room temperature.

Finally, 30-nm-thick C 60 at a rate of 0.5 Å s −1 , 6-nm-thick BCP at a rate of 0.2 Å s −1 and a 100-nm-thick silver electrode at a rate of 1.0 Å s −1 were thermally evaporated, respectively, under high vacuum (< 5 × 10 − 6 Torr). To enhance transmittance, a 125-nm-thick magnesium fluoride (MgF 2 ) layer was deposited on the back side of the ITO substrate at a rate of 1 Å s − 1 . Two sizes of square substrates (1.5 cm × 1.5 cm and 2.5 cm × 2.5 cm) were used in this work, with the specific device layouts illustrated in Supplementary Fig. 63. The active device area was precisely defined by a metal shadow mask and characterized as 0.05 cm 2 .

The final optimized PSCs fabrication process by the automated platform

Using the same substrates, SAM preparation and perovskite solution as those prepared manually above, an automated fabrication platform with an intelligent robotic arm was used. 100 μl of the perovskite solution was dispensed onto the ITO substrate from a tip-to-substrate distance of 8 mm. After being left undisturbed for 5 s, the substrate was spin-coated at 1,000 rpm for 5 s using an acceleration of 1,000 rpm s − 1 and subsequently ramped up to 5,000 rpm for 30 s with an acceleration of 2,000 rpm s − 1 . 300 μl of CB was dripped onto the centre of the film from a height of 8 mm 5 s before the end of spin-coating. The deposited perovskite films were immediately transferred to a hotplate, a process requiring 6 s for the robotic arm to complete the pickup and placement, and then annealed at 110 °C for 30 min and 125 °C for 5 min.

For surface passivation, all solutions were prepared by stirring at 600 rpm using a 4-mm magnetic stir bar for 2 h at 50 °C.

PEAI was dissolved in IPA:DMF (200:1 v/v) at 1.5 mg ml −1 . The solution was dynamically spin-coated by accelerating the substrate at 2,000 rpm s −1 to 5,000 rpm. At 3 s after the rotation started, the solution was dispensed from a tip-to-substrate distance of 8 mm. The total spinning duration was 30 s, followed by an immediate transfer to the hotplate with a movement duration of 6 s, for annealing at 100 °C for 8 min.

4AEI and 4AMI were dissolved in IPA:DMSO (400:1 v/v) at 2 mg ml −1 . The solutions were dynamically spin-coated using the same protocol, in which the substrate was accelerated at 2,000 rpm s −1 to 5,000 rpm and the solution was dispensed from a height of 8 mm at the third second of the spinning process. The films were transferred to the hotplate over a duration of 6 s to be annealed at 110 °C for 5 min.

3AMI was dissolved in IPA:CB:DMSO (400:40:1 v/v) at 3 mg ml −1 . The solution was dynamically spin-coated with an acceleration of 2,000 rpm s −1 to 5,000 rpm, during which the solution was dispensed from a tip distance of 8 mm after a 3-s delay. The total spin time was 30 s, followed by a transfer step taking 6 s to anneal at 110 °C for 10 min. Subsequently, the film surface was rinsed with pure IPA. For this rinsing process, the substrate was accelerated at 2,000 rpm s −1 to 5,000 rpm. At 3 s after the rotation began, 150 μl of pure IPA was dispensed from a tip-to-substrate distance of 12 mm. After continuing the rotation for a total of 30 s, the film was transferred to the hotplate with a transport time of 6 s and annealed at 75 °C for 5 min.

5ANI was dissolved in IPA:CB:DMSO (250:50:1 v/v) at 2.2 mg ml −1 . The solution was dynamically spin-coated by accelerating at 2,000 rpm s − 1 to 5,000 rpm and dispensing the liquid from a height of 8 mm at the 3 s mark. The film was transferred to the hotplate with a mechanical action lasting 6 s and annealed at 110 °C for 10 min. Subsequently, the film surface was rinsed with pure IPA using the identical protocol in which the substrate was accelerated at 2,000 rpm s − 1 to 5,000 rpm and 150 μl of IPA was dispensed from a height of 12 mm after 3 s of spinning. The film was then transferred within a 6-s duration and annealed at 75 °C for 5 min.

The devices were then automatically transferred by means of a conveyor system into the thermal evaporation equipment to complete the electron transport layer and electrode fabrication. After device fabrication, they were automatically transferred to the testing and feedback platform. The entire experimental set-up, including solution preparation, thin-film fabrication, thermal evaporation and the J–V characterization system, is integrated within nitrogen-filled gloveboxes. The internal temperature is maintained at 18–22 °C by an integrated air-conditioning system, while the oxygen and water levels are strictly controlled at <5 ppm and <0.5 ppm, respectively. Consequently, the samples were kept isolated from the ambient atmosphere throughout the entire process, from device fabrication to performance testing.

To evaluate the efficiency of the automated fabrication line, the cycle times for each distinct module were analysed. The film preparation module (including spin-coating and annealing) is capable of processing 12 substrates per 100 min. The thermal evaporation module (for transport layers and metal electrodes) completes a cycle for 12 substrates in 55 min. Finally, the automated J–V characterization module tests 12 substrates in 20 min. Therefore, the film preparation process serves as the rate-determining step for the entire line. On the basis of this fabrication speed, the integrated platform achieves an overall continuous production rate of approximately 7.2 devices per hour.

Perovskite solar mini-module fabrication

The perovskite solar mini-module was composed of seven sub-cells connected in series on a 6.1 × 6.1-cm substrate. The deposition process of the hole transport layer was identical to that used for PSCs. The perovskite precursor solution (1.4 M) was spin-coated in two steps: 1,000 rpm for 10 s, followed by 5,000 rpm for 45 s. With 25 s remaining in the spin-coating process, a hot air blower positioned 55 cm above the film was activated to direct hot air onto its surface, accelerating solvent evaporation, and the perovskite films were annealed at 110 °C for 30 min and 125 °C for 5 min.

Subsequently, electron transport layer and electrode deposition were carried out following the same procedures as those used for PSCs. During mini-module fabrication, three laser scribing processes were performed using a laser marking machine (Han’s Laser PV-HL0303, wavelength 532 nm). The P2 scribing was conducted with a laser power of 2.5 W at 100 kHz, whereas the P3 scribing was performed at 5 W and 100 kHz. The optimized scribing widths were determined to be P1 = 50 µm, P2 = 75 µm and P3 = 50 µm, respectively. With an optimized safety margin between lines, this configuration resulted in a total dead width of 245 µm and a high geometrical FF of 96.9% (Supplementary Fig. 64).

Characterizations

QFLS values were calculated by the data extracted from photoluminescence quantum yield by exciting the corresponding samples using a 375-nm laser (EPL-375, Edinburgh Instruments) inside an integrated sphere (FLS980, Edinburgh Instruments). Electroluminescence quantum efficiency (EQE EL ) measurement was conducted by applying an external voltage/current source through the instrument (ELCT-3010, Enlitech). KPFM was conducted through Bruker Dimension Icon and Ti/Ir-coated silicon tips (ASYELELC-01-R2) with a resonance frequency at approximately 58–97 KHz were used in scanning Kelvin probe microscopy and electrostatic force microscopy imaging. A Photon etc. wide-field microscope was used for voltage-dependent PL mapping to obtain the optical PCE maps. The devices were illuminated with three fibre-coupled light-emitting diodes (LEDs; red M455L4, green M530L4 and blue M617L3 from ThorLabs). The power of each LED was adjusted so that the short-circuit current was the same as 1 sun under a calibrated LED solar simulator. An approximation of a white-light solar illumination profile was produced by ensuring that a third of the current came from each LED. Individual pixels were biased using a 3D-printed custom device holder and biased using a Keithley 2450 source meter. Sweeps from 0 V to open circuit and back were performed in 80 steps. A long-pass filter (Semrock BLP01-664R) was added in the PL emission path to remove scattered excitation and emission was collected using a Hamamatsu ORCA Flash 4.0 V3 sCMOS camera. To process the data, the data were denoised using principal component analysis using Hyperspy v1.6.5 ( ). Then the first image at open circuit was used for normalization for the reverse scan and the last for the forward scan. The optical maximum power was found by calculating the value of the product Φ PL (V) × V at the MPP, for which Φ PL (V) is defined as below, in which I is the intensity of PL at either V OC or voltage V. This quantity was calculated at each point across the measured area.

The current–voltage (J–V) characteristics of photovoltaic devices were conducted in a N 2 -filled glovebox at room temperature by using a xenon lamp solar simulator (SS-F5, Enlitech) and a LED solar simulator (PR-SCCS-IV2R, PURI Materials). The power of the light was calibrated to 100 mW cm −2 by a silicon reference cell (with a KG2 filter). All of the devices were measured using a Keithley 2400 source meter under a sweep mode of reverse scan (from 1.30 to −0.1 V) and forward scan (from −0.1 to 1.30 V) with the scan rate of 0.01 V s −1 and the delay time was 10 ms. For the mini-module, the test range was −0.1 to 9.0 V with a scan rate of 0.05 V s −1 and a delay time of 10 ms. No pre-conditioning was needed before measurement. The stabilized power output was conducted by monitoring the stabilized current density output at the MPP bias (extracted from the reverse scan J–V curves). EQE and internal quantum efficiency measurements were carried out using a QE-R EQE system (Enlitech). A PSC was used to adjust the laser intensity to about 0.1–1.0 sun equivalent intensity by obtaining a current density at 0 V approximately equal to its J SC under a standard solar simulator. Highly sensitive EQE was measured by an integrated system (PECT-600, Enlitech), in which the photocurrent was amplified and modulated by a lock-in instrument. EQE EL measurement was conducted by applying an external voltage/current source through the instrument (ELCT-3010, Enlitech).

PSCs MPP stability test (ISOS-L-1I protocol)

The long-term operational stability was conducted by following the standardized ISOS-L-1I (Laboratory, Level 1, indoor) procedure, which entails continuous light soaking under MPP tracking in an ambient laboratory environment. Before the tests, the PSCs were encapsulated inside a nitrogen-filled glovebox with cover glass (0.7 mm) and ultraviolet-cured epoxy resin (LT-U001, Luminescence Technology Corp.). The encapsulated PSCs were operated under continuous 1-sun-equivalent LED lamp illumination in the air atmosphere. To maintain standardized conditions, the shelf temperature of the devices was monitored throughout the test, with the ambient temperature kept at 23.5 ± 3.1 °C and the humidity at 51.7 ± 10.0% relative humidity. The PSCs were continuously biased at the MPP voltage, which was dynamically updated using a perturb-and-observe algorithm and the power output was tracked by using a multipotentiostat (CHI1040C, CH Instruments). During the MPP test, J–V curves of the devices were automatically obtained every 12 h through a dual-direction scan to obtain the proper loads for the MPP.