Answers are grounded in this paper and cite their sources.
Demonstrates improved stability of solar cell predictions using multi-task learning, indicating better model performance.
Conventional machine learning models for perovskite solar cells predict photovoltaic parameters independently, disregarding the physical constraint PCE=Voc×Jsc×FF/100. This approach can yield mutually incompatible predictions for the four parameters, a failure mode that has not been hitherto quantified in the perovskite solar cell literature. This paper proposes a multi-head neural network with a shared backbone, physics-guided feature construction, and task-specific prediction heads, and validates it on 7176 SCAPS-1D simulations across 12 perovskite compositions. When benchmarked against architecturally matched single-task baselines, the multi-task model, optimized via 5-fold cross-validation, achieves R2 values of at least 0.994 for all four targets, with cross-fold standard deviations of 0.001. In particular, fill factor prediction improves from R2=0.617±0.254 (single-task) to 0.994±0.001 (multi-task), a 233-fold reduction in cross-fold standard deviation. Application of a physical consistency metric developed in this work reveals that 36.5% of single-task predictions exceed a 2 PCE-unit implausibility threshold, compared to only 0.01% for the multi-task model. The multi-task model outperforms the single-task baseline in all 20-fold target comparisons, with large effect sizes (Cohen’s d=1.33–8.93). These results confirm multi-task learning as an effective approach for achieving robust, stable, and internally consistent predictions in simulation-based photovoltaic virtual screening.
No takes yet. Share an insight, caveat, or question.
Chahine et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: