Enrichment-Oriented Virtual Screening of Perovskite Solar Cells through Reinforcement Learning
Authors: Yezeyuan Zheng, Ziyang Hu, Maojun Sun. Ruijian Han, Jian Huang, Guanhua Chen, Jie Liu
Organizations: Hong Kong Quantum AI Lab Limited, Pak Shek Kok, Hong Kong SAR, China. · Department of Chemistry, The University of Hong Kong, Pokfulam, Hong Kong SAR, China. · Department of Data Science and AI, and Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China. · MattVerse Limited, Pak Shek Kok, Hong Kong SAR, China. · QuantumFabless Limited, Pak Shek Kok, Hong Kong SAR, China. · Lixin Technology (Shenzhen) Co. Ltd., Shenzhen, Guangdong, China.
Experimental optimization of perovskite solar cells is constrained by combinatorial diversity in absorber compositions, device layers, and processing, which makes virtual screening attractive for prioritizing recipes before fabrication. We develop a surrogate-guided reinforcement-learning framework that casts recipe generation as virtual screening, enriching candidate shortlists near high-performing experimental neighbors. A frozen Tabular Prior-data Fitted Network (TabPFN) guides recipe generation, while a reference pool of 890 measured devices enables neighborhood-based assessment. Five-fold evaluation on 1,846 devices shows top-ten selection raises mean power conversion efficiency (PCE) from 19.68% to 23.74%. Three learned policies are compared with six baselines under 3,000 surrogate evaluations per run. In top-100 shortlists, Proximal policy optimization (PPO) and recurrent PPO (RPPO) achieve neighborhood-supported hit rates of 23% and 24% (7.9- and 8.3-fold Pure Random's 2.9%), versus 10% (3.4-fold) for the baselines with the highest supported hit rate (L-SHADE and Ax BO, tied) at K=100. Gains coexist with weak-support fractions. Of 9,000 recipes, 65.4% match a composition family at similarity 0.80, falling to 11.7% at 0.90. Results support enrichment-oriented prioritization but not experimental performance, seed robustness, or matched-budget advantage.
Figures & tables
Figure 1: Overview of the configuration space, surrogate-guided policy optimization, virtual-screening workflow, and evaluation framework for perovskite solar-cell design. (a) Device configurations encode layer identities, ABX 3 composition fractions normalized to stoichiometry, and processing features for n–i–p or p–i–n architectures; these representations do not specify complete fabrication protocols. (b) A TabPFN surrogate is fitted on a 1,292-record training partition drawn from a 1,846-record high-PCE subset of the Perovskite Database and large language model (LLM)-extracted literature (PCE ≥18% , publication year ≥2016 ; stratified 70:30 split, seed 50), and is then frozen for policy optimization. Proximal policy optimization (PPO), recurrent PPO (RPPO), or trust region policy optimization (TRPO) policies propose candidates defined by 21 controlled features, together with band gap and a fixed publication year. Policy updates use the immediate change in surrogate score, r=R(candidate)−R(current) (where candidate corresponds to st+1 and current corresponds to st as shown in the methods, with discount factor γ=0 ). (c) Three complementary evaluations assess held-out surrogate ranking through five-fold cross-validation, generated-candidate similarity to 890 reference devices published from 2020 onward using SHapley Additive exPlanations (SHAP)-weighted device neighborhoods, and composition similarity to separately retrieved OpenAlex titles and abstracts of non-open-access works published from 2024 onward.
Model
Full dataset
High-PCE subset
R2
MAE
RMSE
R2
MAE
RMSE
Rumman Rumman et al. (2023)
0.706
1.980
2.802
0.201±0.034
1.401±0.037
2.064±0.043
Kusuma Kusuma et al. (2025)
0.751
1.901
2.868
0.335±0.056
1.274±0.030
1.883±0.085
Ours
0.526±0.012
2.593±0.050
3.485±0.050
0.414±0.077
0.867±0.021
1.118±0.042
Table 1: Reward-model performance. Values with uncertainties are reported as mean ± standard deviation across five cross-validation folds. MAE and RMSE are expressed in percentage points of PCE. Full-dataset baseline results are taken from the original publications and use their respective data preparation and evaluation protocols. High-PCE baseline reproductions retain different record subsets under their respective preprocessing pipelines. These comparisons are therefore indicative rather than strictly controlled.
Figure 2: Validation of surrogate ranking and the reference-neighborhood signal. (a) Predicted versus measured PCE for all 1,846 devices, using out-of-fold predictions from five-fold stratified cross-validation (each point was predicted by a model that did not see that device). (b) Mean measured PCE of the ten highest-scoring devices in each validation fold compared with the random-selection expectation, defined as the mean measured PCE of that fold. Paired-point annotations show selection gains in percentage points (pp); the mean gain across folds is 4.05 pp. (c) Spearman correlation between measured PCE and neighborhood-mean PCE within the 890-device reference pool for maximum neighborhood sizes of k=3 , 5, and 10. Self-matches are excluded and eligible neighbors have similarity ≥0.50 . Error bars show reported bootstrap 95% confidence intervals; coverage is 96.4% at each setting. Panel (c) assesses the neighborhood signal internally within the reference pool, not the experimental performance of generated candidates. Detailed ranking results, residual analysis, and neighborhood-validation controls are provided in Appendices H , D.2 , and I.8 , respectively.
Method
Mean best score
SD
Evaluations to threshold
Reinforcement-learning agents
PPO
34.50
0.35
–
RPPO
34.89
0.04
–
TRPO
33.30
0.61
–
Optimization baselines
L-SHADE
30.36
0.24
1648.7±514.2
Table 2: Best-found surrogate scores and baseline time-to-threshold statistics. Score means and standard deviations are calculated from the best score in each of three seeded runs, each with 3,000 reward-model evaluations. RL runs use different inference seeds of one frozen trained policy per algorithm; baseline runs are separately initialized optimization runs. RL standard deviations therefore do not measure training-seed variability. Time to threshold is the first evaluation at which the running best reaches 95% of that run’s final within-budget best score, reported as mean ± SD across runs. No comparable time-to-threshold statistic is reported for pretrained RL policies. Pure Random is excluded because it has no surrogate evaluation.
Figure 3: Reference-neighborhood evidence for the top-100 surrogate-ranked candidates. Bars show the proportions classified as Strong, Positive, Weak, or Uncovered for PPO, RPPO, TRPO, and L-SHADE, the non-RL baseline with the highest best-found surrogate score (Table 2 ); L-SHADE is tied with Ax BO on the K = 100 supported hit rate (10% each). All other baselines are shown in Figure 8 . The lower bar shows Pure Random, a background reference of 1,000 unselected recipes that were not surrogate-ranked; it is not a matched-shortlist comparison. Neighborhood PCE is the unweighted mean measured PCE of up to k=5 reference neighbors with similarity ≥0.50 . Strong denotes neighborhood PCE at or above the reference 85th percentile; Positive denotes values at or above the reference median but below the 85th percentile; Weak denotes values below the median. Candidates with no eligible neighbor are Uncovered. Percentages use all candidates in each group as the denominator (100 per method; 1,000 for Pure Random). Segments below 7% are not labeled; for Pure Random, Strong is 2.9%. These categories describe reference-neighborhood evidence, not experimentally verified candidate performance.
Figure 4: Early enrichment and screening-quality trade-offs. Candidates from PPO, RPPO, TRPO, and L-SHADE are ranked by descending saved surrogate-predicted PCE, and statistics are evaluated for top-K prefixes at K = 10, 25, 50, and 100. (a) Enrichment factor (EF): the neighborhood-supported hit rate (the fraction of all K candidates that are covered and have mean neighborhood PCE at or above the reference-pool 85th percentile) divided by the same fraction among the 1,000 unselected Pure Random recipes (2.9% at k=5 ). The dashed horizontal line marks EF=1 , the rate expected from unselected sampling. (b) Reference coverage: the fraction of all K candidates with at least one eligible reference neighbor. (c) Mean candidate-level neighborhood PCE among covered candidates only; it is undefined when no candidate in the prefix is covered. In all panels, the grey band spans the 2.5th–97.5th percentiles of the statistic over 1,000 random draws of K distinct recipes from the Pure Random sample, and the dotted line is the median of those draws. The band describes sampling variability of unselected selection only; it is conditional on the observed Pure Random sample. For each candidate, neighborhood PCE is the unweighted mean measured PCE of up to k=5 reference neighbors with similarity ≥0.50 . K denotes shortlist size and k the maximum number of reference neighbors. All results use the designated seed-42 run for each method. Lines connect evaluated shortlist sizes as visual guides only; they do not represent optimization trajectories or uncertainty across runs. These statistics describe experimental-reference support, not measured performance of generated devices.
Figure 5: External composition-family matching against the OpenAlex reference pool. (a) Fractions of generated recipes whose best-match composition similarity meets the threshold τcomp=0.80 , shown separately for PPO, RPPO, and TRPO; labels report matched and total recipe counts. (b) Matching rates across the evaluated composition-similarity thresholds for the pooled 9,000 RL-generated recipes and 1,000 unselected Pure Random recipes. At τcomp=0.80 , 5,889 RL-generated recipes (65.4%) and 377 Pure Random recipes (37.7%) meet the threshold. At 0.70, the corresponding rates are 92.2% and 77.3%; at 0.90, they are 11.7% and 3.2%. Connecting lines are visual guides only. Matching concerns composition-family similarity and does not establish reproduction of a published recipe, synthesizability, or measured performance. Detailed numerical results are provided in Appendix J .
Feature
Type
Encoding
Perovskite composition
Composition
7-dim ratio vector
Perovskite band gap
Numerical
not transformed
Substrate stack sequence
Sequence
Positional, ≤2 layers
ETL stack sequence
Sequence
Positional, ≤3 layers
HTL stack sequence
Sequence
Positional, ≤2 layers
Backcontact stack sequence
Sequence
Positional, ≤2 layers
Table 3: Selected base features and their representation. Ten agent-controlled features expand into the 21-dimensional recipe vector. Band gap and publication year are additional inputs to the reward model and are not controlled by the agent.
Hyperparameter
PPO
RPPO
TRPO
Policy architecture
MLP
MLP + LSTM
MLP
Requested timesteps
8×105
8×105
8×105
Actual timesteps
819,200
819,200
819,200
Parallel environments
100
100
100
nsteps
256
256
256
Batch size
512
512
512
Table 4: Reinforcement learning hyperparameters. All three algorithms use the same environment, reward function, and total training budget.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Stack
1
2
3
≥4
Substrate
4.5%
95.3%
0.1%
0.2%
ETL
37.2%
52.7%
9.8%
0.3%
HTL
87.2%
12.4%
0.5%
0.0%
Backcontact
95.3%
4.5%
0.2%
0.0%
Appendix
Table 5: Sublayer-count distribution per stack type in the cleaned dataset. Percentages are computed over entries with a non-missing stack sequence for that field. The sublayer counts are used to choose the positional truncation length for each stack in the recipe vector.
Position
Feature
Values
1
Substrate layer 1
Categorical
2
Substrate layer 2
Categorical
3
ETL layer 1
Categorical
4
ETL layer 2
Categorical
5
ETL layer 3
Categorical
6
HTL layer 1
Categorical
Appendix
Table 6: Layout of the 21-dimensional recipe vector (RL action and observation space). Two additional non-agent features (band gap, fixed publication year) are appended at reward-model evaluation time.
Actual PCE range (%)
Count
Mean bias (%)
% over
% under
18–20
1,184
+0.553
77.4
22.5
20–22
552
−0.868
9.0
90.9
22–24
75
−1.296
20
80.0
>24
35
−2.551
0
100
Appendix
Table 7: Prediction bias by actual PCE range, aggregated across the five cross-validation folds. Mean bias is the average of y^−y over held-out records in each bin, in percentage points. A positive mean bias indicates overestimation. “% over” and “% under” are the fractions of records in the bin for which y^>y and y^<y . Bins are pooled across folds; the pooled mean bias across all 1,846 records is the same as the count-weighted mean of the per-bin values. The <18% bin is absent because the cleaned dataset is restricted to PCE ≥18% .
Figure 6: Residual analysis across the five cross-validation folds. Left: residuals r=y^−y as a function of actual PCE. Right: distribution of residuals. The dashed line indicates zero bias. The model overpredicts below 20% and underpredicts above 20% , with the underprediction growing at the high-PCE end.
Figure 7: Learning dynamics under different discount factors. (a) Mean episode reward for PPO, RPPO, and TRPO with γ=0 (solid) and γ=0.99 (dashed). (b) Entropy loss for PPO and RPPO; TRPO does not log this quantity in the implementation used here. Within each algorithm, the two discount-factor settings use otherwise identical hyperparameters and a requested training budget of 8×105 environment steps, resulting in 819,200 actual steps per run. Each trajectory represents one training run with 100 parallel environments. Each rollout comprises 100×256=25,600 environment steps. Curves are smoothed using a three-point moving average. With γ=0.99 , rewards remain near the initial random-policy level, whereas γ=0 produces substantial improvement. Increasing entropy loss in panel (b) corresponds to decreasing policy entropy.
Algorithm
Discount
Mean episode reward
Entropy loss
Explained variance
PPO
γ=0
−15.7→−3.9
−71.8→−55.1
0.02→0.50
γ=0.99
−15.2→−15.4
−71.8→−71.6
0.01→0.92
RPPO
γ=0
−15.6→−3.1
−71.8→−51.8
0.00→0.48
γ=0.99
−15.5→−15.4
−71.8→−71.6
0.01→0.94
TRPO
γ=0
−15.5→−7.1
—
0.01→0.36
γ=0.99
−15.7→−14.9
—
0.01→0.92
Appendix
Table 8: Discount-factor ablation. Values are reported as initial → final for each training run, with a requested budget of 8×105 environment steps and an actual total of 819,200 steps. Mean episode reward is based on the differential predicted PCE, expressed in percentage points. Because the reset state is not evaluated by the surrogate, its reference value is set to the constant offset R(s0)=35 (Section 4.3 ), so the return of an episode is R(sT)−35 and the reported rewards are relative to this offset. For example, a final mean reward of −3.1 corresponds to an end-state surrogate score of about 31.9 . Entropy loss is the logged negative policy entropy, not a PCE quantity; increasing entropy loss corresponds to decreasing policy entropy. Explained variance is dimensionless. TRPO does not log entropy loss in the implementation used here, so the corresponding entries are marked “—”. Each row summarizes one training run.
Method
Tuned parameters
Values
Random Search
—
—
L-SHADE
μF , μCR
0.766 , 0.270
PSO
c1 , c2 , w
2.197 , 1.312 , 0.225
RUN
—
default
CMA-ES
σ0
1.237
Ax BO
—
default
Appendix
Table 9: Hyperparameters for the six optimization baselines. Values marked “default” use the library’s recommended settings.
Fold
n
MAE
RMSE
R2
Spearman ρ
Top-ten PCE (%)
Random PCE (%)
Gain (pp)
1
370
0.864
1.146
0.504
0.457
24.680
19.721
4.959
2
369
0.895
1.160
0.343
0.410
23.588
19.671
3.917
3
369
0.865
1.095
0.306
0.382
22.264
19.639
2.626
4
369
0.831
1.047
0.444
0.487
23.674
19.669
4.005
5
369
0.879
1.144
0.475
0.441
24.479
19.721
4.758
Mean ± SD
0.867±0.021
1.118±0.042
0.414±0.077
0.436±0.037
23.737±0.853
19.684±0.032
4.053±0.822
Appendix
Table 10: Held-out regression and ranking performance across the five cross-validation folds. MAE, RMSE, and R2 are computed on the validation fold. Spearman ρ is the rank correlation between predicted and measured PCE. Top-ten PCE is the mean measured PCE of the ten recipes with the highest predicted scores in each validation fold; Random PCE is the expected mean under uniform random selection from that fold; Gain is the difference in percentage points (pp). The final row reports the mean and, on the second line, the standard deviation across folds.
Table 11: Components of the similarity metric. Each component groups one or more features of the reward model and is scored by a similarity function appropriate to its data type.
Table 12: Component weights for the similarity metric, derived by summing the SHAP feature importances of the separate random forest model within each component. The Pub_Year feature is excluded as a metadata attribute rather than a recipe property. Weights are not normalized to sum to unity; the metric normalizes internally.
Figure 8: Reference-neighborhood evidence across all compared optimization methods. Bars show the proportions classified as Strong, Positive, Weak, or Uncovered within each method’s top-100 surrogate-ranked shortlist, extending Figure 3 to all methods included in the candidate-level analysis: the three reinforcement-learning policies (PPO, RPPO, and TRPO) and six optimization baselines (L-SHADE, Ax BO, CMA-ES, Random Search, RUN, and PSO). The lower bar shows Pure Random, a background reference of 1,000 unselected recipes that were not surrogate-ranked; it is not a matched-shortlist comparison and should not be confused with the Random Search baseline. All groups use the same reference pool, a maximum of k=5 neighbors, and a minimum similarity of 0.50 . Neighborhood PCE is the unweighted mean measured PCE of the eligible neighbors. Strong denotes neighborhood PCE at or above the reference 85th percentile; Positive denotes values at or above the reference median but below the 85th percentile; Weak denotes values below the median. Candidates without an eligible neighbor are Uncovered and are not classified as Weak. Percentages use all candidates in each group as the denominator (100 per method; 1,000 for Pure Random). Segments below 7% are not labeled. The PSO and RUN shortlists are not diverse in reference support: all 88 covered PSO candidates matched the same single reference device (PCE 17.90%), and RUN neighborhoods were drawn from a few reference devices, so their category fractions reflect a narrow region of the design space rather than independent evidence from many candidates. These categories describe reference-neighborhood evidence, not experimentally verified candidate performance.
k = 3
Method
Strong
Pos.
Weak
Avg. PCE
Cov.
Reinforcement learning agents
PPO
31.5%
21.9%
46.6%
18.64 ± 4.36%
73.0%
RPPO
26.7%
25.6%
47.8%
18.23 ± 4.39%
90.0%
TRPO
45.7%
34.8%
19.6%
20.65 ± 3.73%
46.0%
Optimization baselines and sampling reference
Appendix
Table 13: Signal distribution of generated recipes for each method. For each recipe, up to k nearest reference recipes with similarity ≥τ=0.50 are selected from the deduplicated Perovskite Database and LLM-extracted reference pool ( n=890 ; see Appendix A.1 ); their measured PCE values define the neighborhood PCE. Recipes without an eligible neighbor are uncovered . Covered recipes are labeled Strong ( ≥ Q85), Pos. ( ≥ Q50 and < Q85), or Weak ( < Q50) by neighborhood PCE. Avg. PCE is the mean ± std of recipe-level neighborhood PCE (%) among covered recipes. Cov. is the fraction of analyzed candidates with at least one eligible neighbor (shared across k ). Strong/Pos./Weak percentages use covered recipes as denominator; Cov. uses the full analyzed set (100 shortlisted candidates per RL policy or baseline; 1,000 unselected samples for Pure Random). Percentages may not sum to 100 due to rounding.
Method
k=3
k=5
k=10
PPO
23.0%
23.0%
23.0%
RPPO
24.0%
24.0%
24.0%
TRPO
21.0%
20.0%
20.0%
L-SHADE
10.0%
10.0%
9.0%
Ax BO
10.0%
10.0%
9.0%
Random Search
3.0%
4.0%
3.0%
Appendix
Table 14: Percentages of analyzed candidates that are both covered and classified as Strong, calculated as the product of the coverage fraction and the conditional Strong fraction from Table 13 . The denominator is 100 shortlisted candidates for each RL policy and optimization baseline, and 1,000 unselected samples for Pure Random. Pure Random characterizes background support under sampling without surrogate-based selection and is not a matched-shortlist comparison. The parameter k denotes the maximum neighborhood size. Uncovered candidates are not classified as Weak. PSO has a zero supported hit rate because all 88 of its covered candidates were classified as Weak.
Figure 9: Early enrichment and screening-quality trade-offs across all evaluated optimization methods. This figure extends Figure 4 to the three reinforcement-learning policies (PPO, RPPO, and TRPO; colored solid lines) and six optimization baselines (L-SHADE, Ax BO, CMA-ES, Random Search, RUN, and PSO; grey dashed lines, distinguished by marker). Candidates are ranked by descending saved surrogate-predicted PCE, and statistics are evaluated for top- K prefixes at K=10 , 25, 50, and 100. (a) Enrichment factor (EF): the neighborhood-supported hit rate (the fraction of all K candidates that are covered and have mean neighborhood PCE at or above the reference-pool 85th percentile) divided by the same fraction among the 1,000 unselected Pure Random recipes (2.9% at k=5 ). The dashed horizontal line marks EF=1 . Candidates without a reference neighbor count as non-hits. (b) Reference coverage: the fraction of all K candidates with at least one eligible reference neighbor. (c) Mean candidate-level neighborhood PCE among covered candidates only; it is undefined when no candidate in the prefix is covered. In all panels, the grey band spans the 2.5th–97.5th percentiles of the statistic over 1,000 random draws of K distinct recipes from the Pure Random sample, and the dotted line is the median of those draws. Pure Random is not plotted as a method because its samples were not ranked by surrogate score. For each candidate, neighborhood PCE is the unweighted mean measured PCE of up to k=5 reference neighbors with similarity ≥0.50 . K denotes shortlist size and k the maximum number of reference neighbors. All methods use the same reference pool and classification thresholds, and results use the designated seed-42 run for each method. Lines connect evaluated shortlist sizes as visual guides only; they do not represent optimization trajectories or uncertainty across runs, and the band reflects sampling variability of unselected selection only. These statistics describe experimental-reference support, not measured performance of generated devices.
k
Cov.
ρ
95% CI
ZPCE-shuf
ρˉnbr-shuf
Δρ
3
96.4%
+0.682
[ +0.636 , +0.722 ]
+17.55
−0.001
+0.683
5
96.4%
+0.699
[ +0.655 , +0.740 ]
+17.38
−0.004
+0.702
10
96.4%
+0.707
[ +0.664 , +0.742 ]
+16.81
−0.004
+0.711
Appendix
Table 15: Validation of the SHAP-weighted similarity metric on the reference pool ( n=890 ). For each reference recipe, we compute the mean PCE of its top- k most similar neighbors and correlate this neighborhood score with the recipe’s own measured PCE. Neighbors are filtered by a similarity floor τ=0.50 ; self-matches are excluded. Cov. is the fraction of recipes with ≥1 eligible neighbor. ρ is Spearman’s rank correlation with a bootstrap 95% CI ( Nboot=1000 ). ZPCE-shuf is the number of standard deviations by which the observed ρ exceeds a PCE-shuffled null that holds the similarity graph fixed and permutes PCE labels ( Nperm=500 ; pperm=0.002 in every case). ρˉnbr-shuf is the mean ρ under a neighbor-shuffle null that replaces the top- k neighbors with k recipes drawn uniformly at random. Δρ is the increment of the observed ρ over this null ( Δρ=ρ−ρˉnbr-shuf ), isolating the contribution of similarity-based selection from pool-wide PCE autocorrelation. All entries are rounded to three decimals; Δρ is computed from the unrounded ρ and ρˉnbr-shuf , so in some rows the displayed values may differ in the last digit from ρ−ρˉnbr-shuf computed on the rounded entries.
Similarity interval
Recipes
Fraction
[0.0,0.3)
0
0.0%
[0.3,0.5)
48
0.5%
[0.5,0.7)
654
7.3%
[0.7,0.8)
2,409
26.8%
[0.8,0.9)
4,835
53.7%
[0.9,1.0]
1,054
11.7%
Appendix
Table 16: Distribution of best-match composition similarity between N=9,000 generated recipes and the retrieved OpenAlex reference pool. A total of 1,054 recipes (11.7%) attained a best-match composition similarity of 0.90 or above, and no generated recipe exceeded a similarity of 1.00. At the threshold of τcomp=0.80 , 5,889 recipes (65.4%) were classified as composition-family matches. These results describe similarity to the retrieved reference compositions, not experimental validation or novelty relative to the complete literature.
τcomp
RPPO
PPO
TRPO
Pooled RL
Pure Random
0.50
2,995 (99.8)
2,991 (99.7)
2,966 (98.9)
8,952 (99.5)
979 (97.9)
0.60
2,981 (99.4)
2,970 (99.0)
2,898 (96.6)
8,849 (98.3)
933 (93.3)
0.70
2,866 (95.5)
2,825 (94.2)
2,607 (86.9)
8,298 (92.2)
773 (77.3)
0.80
2,246 (74.9)
2,085 (69.5)
1,558 (51.9)
5,889 (65.4)
377 (37.7)
0.85
1,309 (43.6)
1,163 (38.8)
750 (25.0)
3,222 (35.8)
159 (15.9)
0.90
472 (15.7)
402 (13.4)
180 (6.0)
1,054 (11.7)
32 (3.2)
Appendix
Table 17: Composition-family matching rate as a function of the similarity threshold τcomp , shown as matched recipes (rate in %) for each generation algorithm ( 3,000 recipes each), the pooled RL-generated recipes ( 9,000 ), and the unselected Pure Random background ( 1,000 ). In the designated seed-42 runs, the ordering RPPO > PPO > TRPO holds at every evaluated threshold, and all three algorithms exceed the Pure Random background. The pooled decline is steepest between 0.80 and 0.90 , and the pooled difference from Pure Random is largest at the threshold of 0.80 ( 27.7 pp). These results characterize matching against the retrieved reference pool, not experimental validation.
Algorithm
Recipes
Matched
Rate
RPPO
3,000
2,246
74.9%
PPO
3,000
2,085
69.5%
TRPO
3,000
1,558
51.9%
All RL (pooled)
9,000
5,889
65.4%
Pure Random
1,000
377
37.7%
Appendix
Table 18: Composition-family matching rates by generation algorithm at the threshold τcomp=0.80 . RPPO has the highest matching rate ( 74.9% ), followed by PPO ( 69.5% ) and TRPO ( 51.9% ); TRPO is 17.6 pp below PPO and 23.0 pp below RPPO. All three exceed the unselected Pure Random background ( 37.7% ). Each policy is represented by its designated seed-42 candidate pool, so these differences are not an estimate of training-seed variability. These rates do not establish differences in exploration breadth or experimental validity.
Efficient discovery of precursor additives is essential for improving the performance of perovskite solar cells, yet the large chemical space makes conventional trial-and-error screening inefficient. We develop LEAP(LLM-driven Exploration via Active Learning for Perovskites), an expert-in-the-loop closed framework that couples a domain-specialized large language model(LLM) with active learning for iterative additive prioritization. The LLM is trained to extract mechanism-relevant knowledge from the perovskite additive literature and to represent candidate molecules through interpretable descriptors, which are further integrated into a Bayesian optimization workflow for uncertainty-aware prioritization under low-data conditions. Benchmark results on unseen literature show that the domain-specialized model outperforms general-purpose models in mechanism-consistent reasoning. Experimental validation in an expert-in-the-loop proof-of-concept study suggests improved additive prioritization across three screening rounds, leading to average device PCEs of 20.13% and 20.87% for the later-round 6-CDQ- and 2-CNA-treated devices, respectively, compared with 19.25% for the control, with a champion PCE of 21.32%. These results provide preliminary evidence that literature-grounded mechanistic descriptors, when coupled with Bayesian optimization and expert feasibility review, can support mechanism-aware additive prioritization in perovskite photovoltaics.
Xin-De Wang, Zhi-Rui Chen, Ze-Feng Gao +3
School of Physics, Renmin University of China, Beijing, China · School of Chemistry and Life Resource, Renmin University of China, Beijing, China
Existing machine-learning models for perovskite solar cells struggle to predict devices made from new materials. Most treat device layers as arbitrary categorical labels and represent perovskite absorbers either through featureless composition vectors or through rigid crystal graphs that require a pre-solved unit cell. As a result, they cannot generalize to unrecorded alloys or unfamiliar cell architectures. We present the Virtual-Alloy Physics-Aware Semantic System (V-PaSS), an architecture built around the physical continuity of perovskite solid solutions. V-PaSS represents arbitrary mixed-cation, mixed-halide absorbers by interpolating relaxed crystal graphs of 23 pure ABX3 endpoint compounds via Vegard's law with learned bandgap bowing. For the device stack, it couples layer sequence and cell polarity (n-i-p or p-i-n) with interfacial band offsets and encodes chemical descriptions with a scientific language model. Evaluated under an institution-grouped protocol where test devices come exclusively from unseen laboratories, an untuned default configuration of V-PaSS achieves the lowest error among fourteen tested models (2.861% MAE, a 5.2% relative reduction over a gradient-boosted baseline at 3.017%, and outperforming the best tuned benchmark at 2.912%). V-PaSS pulls furthest ahead where test devices differ most from training: reducing error by 15.7-19.0% on unseen absorber alloys, by 16.1% on mixed absorbers from new laboratories, and widening its lead over the baseline by 3.6× on devices with unobserved transport layers. Combined with tree ensembles, V-PaSS sets a new state of the art across all five benchmark splits.
Qiwei Han, Chi Zhou
Department of Chemistry, Duke University, Durham, NC 27708, USA · Department of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332, USA
Machine-learning models for perovskite solar cells are typically evaluated using random data splits, which we show systematically leak information through massive internal duplication. Because nearly two-thirds of curated devices share their exact recipe with another, random splits primarily reward the mere recall of historical data rather than testing true physical generalisation. To solve this, we mapped the institutional provenance of the Perovskite Database and established a rigorous hierarchy of grouped splits that progressively isolate duplicate devices, publications, and institutions. Benchmarking thirteen models establishes new state-of-the-art performance across all splits and reveals a critical regime shift: while tree ensembles dominate random splits by recognising familiar recipes, their ranking collapses on genuinely unseen institutions and future discoveries, where composition-aware neural architectures excel. Furthermore, feature importance demonstrates that the device architecture and fabrication methodology dictate final efficiency just as strongly as the absorber chemistry. This standardised protocol empowers the community to separate memorisation from true physical discovery.
Qiwei Han, Chi Zhou
Department of Chemistry, Duke University, Durham, NC 27708, USA · Department of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332, USA