Quantifying Perovskite Solar Cell Degradation via Machine Learning from Spatially Resolved Multimodal Luminescence Time Series

Abstract

Perovskite solar cells (PSCs) have experienced a remarkable rise in power conversion efficiency (PCE) over the past 15 years, positioning them as a promising alternative or complement to silicon for large-scale photovoltaic deployment. However, beyond scalable fabrication, operational stability remains a major bottleneck for commercialization. Reliable and rapid methods to assess device health and degradation mechanisms - ideally compatible with field applications - are therefore essential. We present a deep-learning framework to estimate efficiency retention, $R_\mathrm{PCE}=\mathrm{PCE}_t/\mathrm{PCE}_0$, directly from multimodal luminescence imaging acquired during device aging. Each training sample includes electroluminescence (EL), open-circuit photoluminescence (PLoc), and short-circuit photoluminescence (PLsc) images at an aged state, together with device-specific reference images at $t=0$. This design enables the model to learn spatially resolved degradation patterns relative to the pristine condition. The dataset was collected over 5-70 hours using an automated, in-house measurement platform. We introduce LumPerNet, a compact convolutional neural network that regresses $R_\mathrm{PCE}$ from stacked multimodal image tensors, and benchmark it against an intensity-only multilayer perceptron baseline. Using a leakage-aware protocol with device-level hold-out testing and four-fold cross-validation, restricted to $R_\mathrm{PCE}\in[0.8,1.2]$, LumPerNet achieves substantially improved and more robust performance (MAE -23.4%, RMSE -25.6%, $R^2$ +0.417). Ablation studies highlight the importance of complementary physical contrast across modalities for generalization. Overall, this work establishes a reproducible pipeline linking automated luminescence imaging to electrical labels, enabling accelerated stability testing and non-invasive degradation monitoring in PSCs.

Figures (14)

• Figure 1: Experimental monitoring workflow and acquired data modalities. A custom chamber integrates temperature control and source-measure units (SMUs) with two illumination channels (white LED for J--V acquisition, blue LED for PL excitation) and an sCMOS camera for spatially resolved imaging. At user-selected aging times, the automated monitoring cycle yields paired PL images at open circuit ($\mathrm{PL}_{\mathrm{oc}}$) and short circuit ($\mathrm{PL}_{\mathrm{sc}}$), EL images under forward bias (1.5V), and corresponding J--V curves, forming multimodal time series for each device.
• Figure 2: Timing diagram of one automated monitoring cycle executed between two consecutive user-selected aging times, $t_0$ and $t_1$. The cycle comprises (i) white-light soaking (20s) followed by a J--V sweep under white illumination (30s); (ii) dark relaxation at open circuit (60s); (iii) PL imaging under continuous blue illumination at open circuit ($\mathrm{PL}_{\mathrm{oc}}$, 10s) and then immediately at short circuit ($\mathrm{PL}_{\mathrm{sc}}$, 10s; 20s total blue illumination); (iv) dark relaxation at short circuit (15s); and (v) EL imaging during forward-bias current injection (10s). J--V sweeps were performed at every cycle, whereas luminescence images were stored only at a subset of cycles following a progressively sparser schedule at later aging times; the illumination and voltage program of the cycle was otherwise unchanged across all iterations.
• Figure 3: ML framework used in this work. Multimodal luminescence image time series (EL, $\mathrm{PL}_{\mathrm{oc}}$, $\mathrm{PL}_{\mathrm{sc}}$) constitute the model inputs. Two regressors are benchmarked: LumPerNet, a CNN-based model operating on reference--current image pairs, and a baseline MLP trained on space-averaged intensities. Model selection is performed via four-fold cross-validation within the training--validation subset; for each fold, the best checkpoint is selected on the corresponding validation split and evaluated on the held-out test split. Final predictive performance is reported as the mean across folds, with the fold-to-fold standard deviation reported as an estimate of uncertainty.
• Figure 4: LumPerNet ensemble-mean parity plot and absolute error distribution for $R_\mathrm{PCE}$ prediction on the held-out test set. This performance is achieved by exploiting spatial patterns rather than only mean intensities (see baseline comparison in Section \ref{['ssec:comparison']}). Top: hexbin parity plot comparing measured $R_\mathrm{PCE}$ to the ensemble-mean prediction $\hat{y}$, computed by averaging the outputs of the four CV--trained LumPerNet models (one prediction per test sample). Color denotes the number of samples per bin, and the dashed line indicates the ideal $\hat{y}=y$ relationship; inset reports the corresponding ensemble performance metrics. Metrics shown in the inset are computed on the ensemble-mean predictor (mean of the 4 fold-trained models) evaluated on the held-out test set, whereas Table \ref{['tab:LumPerNet_nostack_perfold']} reports mean $\pm$ std across fold-wise test evaluations. The visible banding around the diagonal reflects the longitudinal nature of the test set, with multiple measurements per device that are not statistically independent; device-specific residual structure can therefore emerge as correlated scatter. Bottom: distribution of absolute residuals $|\hat{y}-y|$ over the held-out test samples, summarizing the typical magnitude and spread of prediction errors within the operational range $R_\mathrm{PCE} \in [0.8, 1.2]$.
• Figure 5: Cross-validated test performance comparison across models under identical device-level splits. Bars report mean $\pm$ standard deviation across the four CV runs on the held-out test set for LumPerNet, single-modality (EL-only, PLoc-only, PLsc-only) and two-modality (EL+PLoc, EL+PLsc, PLoc+PLsc) ablations, and the intensity-only baseline. Higher $R^2$ and lower MAE/RMSE indicate better performance.