Interpretable machine learning diagnosis of degradation mechanisms in lead-free RbSn0.5Ge0.5I3 solar cells from temperature- and illumination-dependent electrical measurements
Authors: Nabanita Sarker, Hafiz Imtiaz
Organizations: Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology, Dhaka 1205, Bangladesh
Machine learning coupled with device simulation is widely used to optimize thin-film solar cells, but almost exclusively in the forward direction, from device parameters to predicted performance. The inverse problem encountered after fabrication, namely identifying the mechanism that limits a given device, is still addressed by manual inspection of characterization data. In this work, both directions are examined for lead-free RbSn0.5Ge0.5I3 solar cells using approximately 42,000 SCAPS-1D simulations. Replacing the Spiro-OMeTAD hole transport layer with CuI increases the simulated power conversion efficiency from 19.6% to 21.6%, almost entirely through the fill factor. Surrogate models trained and tested on a factorial parameter grid substantially overstate their predictive accuracy. An XGBoost regressor attains R2=0.9997 on held-out grid points but only 0.58 on independently sampled devices, whereas training on a Latin hypercube design of comparable size restores R2=0.995. For the inverse problem, a multi-label classifier identifies which of six degradation mechanisms are present from simulated temperature- and illumination-dependent open-circuit-voltage measurements. It achieves a macro F1 score of 0.938, compared to 0.687 for a calibrated physics-rule baseline. The classifier does not, however, generalize to combinations of mechanisms absent from the training set.
Figures & tables
Figure 1: Schematic of the simulated FTO/TiO 2 /RbSn 0.5 Ge 0.5 I 3 /CuI/Au device. Light enters through the FTO front contact; electrons are collected at the FTO contact and holes at the Au contact. Layer thicknesses are not to scale. In the simulation, the p-side is placed at the left of the SCAPS domain, as recommended in the SCAPS manual for numerical stability, and the Au contact is represented as a boundary condition rather than as a layer of finite thickness.
Parameter
FTO
TiO 2
RbSn 0.5 Ge 0.5 I 3
CuI
Thickness ( \SIUnitSymbolMicrom )
0.4
0.02
0.35
0.1
Band gap Eg ( eV )
3.3
3.0
1.2
3.1
Electron affinity χ ( eV )
4.1
4.0
4.0
2.1
Relative permittivity εr
9
9
6.5
6.5
NC ( cm−3 )
1.2\times1018
1.2\times1019
2\times1018
2.8\times1019
NV ( cm−3 )
1.8\times1019
1.8\times1019
1\times1019
1\times1019
Table 1: Baseline layer parameters of the simulated FTO/TiO 2 /RbSn 0.5 Ge 0.5 I 3 /CuI/Au device, read from the SCAPS definition file used for all simulations. A value of zero indicates that the mechanism is disabled for that layer, not a missing value. Several parameters are overridden in individual runs, as described in Section 2.3 . Absorber, TiO 2 and FTO values follow Gomathi et al. Super , and CuI values follow Al Amin et al. Super ; the absorber permittivity, which Gomathi et al. Super do not report, is taken from Laouid et al. Super , Al Amin et al. Super .
Figure 2: Baseline FTO/TiO 2 /RbSn 0.5 Ge 0.5 I 3 /CuI/Au device, including the interface defects described in Section 2.1 . Top: simulated J – V curves and external quantum efficiency of the baseline device (21.6%) and the surrogate-optimized device (22.9%). Bottom: equilibrium band diagrams of the baseline device (left) and the surrogate-optimized device (right).
Model
Held-out grid (1,024)
Off-grid (300)
XGBoost
0.9997
0.579
Random forest
0.9987
0.811
MLP
0.9979
0.994
Quadratic
0.9689
0.955
Linear
0.8024
0.725
Table 2: Coefficient of determination ( R2 ) for PCE prediction by five models trained on the same 3,072 grid devices and scored on two test sets. The two columns differ only in the location of the test devices.
Figure 3: Predicted against simulated PCE for the same 300 independent off-grid devices. XGBoost was trained on (a) 3,072 devices of the factorial grid, which form the 75% split of Table 2 , and (b) the 3,000 devices of the space-filling design. The R2 values shown in the panels (0.579 and 0.995) are those reported in Table 2 and Section 3.1 . The horizontal banding in (a), which is absent in (b), reflects the piecewise-constant predictions that a tree ensemble produces between grid levels.
Figure 4: SHAP attributions of the space-filling-trained XGBoost surrogate, computed on held-out devices, for (a) PCE, (b) VOC , (c) JSC and (d) FF (left to right, top to bottom). Bulk defect density dominates PCE and VOC , and absorber thickness dominates JSC . FF is governed jointly by the defect density and the transport-layer doping.
Figure 5: Confusion matrices for the single-degradation devices. Each method was required to name one class instead of giving its native multi-label output: (a) calibrated physics rules, with a forced-choice accuracy of 61%; (b) XGBoost classifier, with 99%. Row labels give the number of test devices in each class.
Figure 6: Per-class SHAP attributions of the diagnostic classifier. The ideality factor dominates the resistance class, and the linearity of VOC(T) dominates the contact-barrier class. dVOC/dT , together with VOC , dominates the bulk class.
Figure 7: Predicted against simulated bulk defect density for the test devices that contain a bulk defect. The dashed line indicates the identity.
Figure 8: Simulated response of each feature when a single mechanism is swept in isolation. Each cell gives the peak-to-peak variation of the feature over the sweep (its maximum minus its minimum across the sweep levels), divided by the standard deviation of the feature across the training set so that features with different units can be compared. Orange outlines mark the feature on which the classifier relies most for each mechanism (highest mean ∣ SHAP ∣ ). The back-barrier and resistance signatures are sharp and exclusive, whereas the two interface columns are uniformly weak, which explains why these classes are the most difficult to diagnose.
Figure 9: Generalization of the diagnostic model. (a) Exact-match accuracy on the same evaluation devices when the pair of mechanisms they contain is present in training (blue) and when every device containing that pair is removed from training (orange). (b) Per-class recall and exact-match accuracy as a function of how far the degraded parameter lies within its degraded range, obtained from out-of-fold predictions within the training set.
Figure 10: (a) Macro F1 as a function of measurement noise. Shaded bands indicate ±1 standard deviation over three independent measurement realizations per level. (b) Feature ablation. The three lowest bars isolate each measurement series. The standard-condition rows quantify the contribution of each series relative to a single J – V curve.
Input
Model
macro F1
15 physics features (Section 2.4 )
XGBoost
0.935 ± 0.001
15 physics features
MLP
0.915 ± 0.007
raw VOC curves only (13 values)
XGBoost
0.811 ± 0.003
raw VOC curves only (13 values)
1-D CNN
0.857 ± 0.007
information-matched raw curves (33)
XGBoost
0.855 ± 0.005
information-matched raw curves (33)
MLP
0.863 ± 0.012
Table 3: Macro F1 for combinations of input representation and model architecture, all evaluated on the same devices. Values are means over three seeds ± one standard deviation.
Figure 11: Learning curves of the feature-based model and the convolutional network on raw curves.
Figure 12: Per-class ROC curves of the diagnostic classifier. The operating points of the calibrated physics rules are marked.
Figure 13: Application to two published FASnI 3 cells 45 . Left: digitized VOC(Φ) points and their interpolation onto the three intensities used for training. Right: resulting class scores. Bulk recombination scores highest for both cells, but an interface term scores almost as highly, so the diagnosis does not identify a single mechanism.
Pair
Pub.
Pred.
Dev.
Range
FASnI 3 / TG 45
5.34
94.6
17.7×
55×
FASnI 3 / SP 46
2.38
10.8
4.6×
5.6×
Table 4: Sensitivity of the predicted bulk trap-density ratio for two untreated/treated pairs. Pub.: ratio reported by the original authors. Pred.: ratio predicted from their published VOC(Φ) data. Dev.: factor by which the prediction deviates from the published value. Range: factor spanned by the prediction when the digitized points are perturbed by 4mV . TG denotes the templated-growth treatment of Liu et al. Super , and SP denotes the combined solvent and surface engineering of Faheem et al. Super . Both treatments are reported by their authors to reduce the trap density.
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
Using large language models to interpret experimental evidence, design discriminating experiments and formulate well-supported mechanistic explanations remains a central challenge in materials research. Here we establish PSM-Bench, an 800-question benchmark developed by perovskite researchers to evaluate scientific reasoning in perovskite photovoltaics. We also introduce OpenPerov, an expert-informed language-model framework that learns from expert-calibrated evidence records reconstructed from 4,035 training articles; these records link experimental strategies, controls and observations to mechanistic interpretations. Structured-evidence adaptation outperformed conventional question-answer training, and OpenPerov Flash ranked first among 13 closed-book systems. OpenPerov Pro raised Scientific Mastery further, from 87.6 to 89.1, by revising selected passages with retrieved evidence. On questions derived from subsequent studies, both models also outscored expert-authored answers. In an experimental test of self-assembled monolayer design, OpenPerov selected a K analogue that outperformed the Na analogue preferred by experts. Devices incorporating the selected contact reached a power conversion efficiency of 27.05% and retained 96.74% of their initial efficiency after 1,000 h of dark thermal ageing at 85 ∘C. These results show that reconstructed scientific reasoning provides an effective basis for building domain models that interpret experiments and inform materials design.
YanKai Chen, Zhi Wan, ZiHao Li +1
School of Materials Science and Engineering, Tsinghua University, Beijing 100084, P. R. China. · College of Bioresources Chemical and Materials Engineering, Shaanxi University of Science & Technology, Xi’an 710021, P. R. China. · Department of Chemical Engineering, Tsinghua University, Beijing 100084, P. R. China.
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