Temporal feature enhancement and flow matching microenvironment modeling for PCE prediction and process optimization of perovskite solar cells

Résumé fourni par la source

Metal halide perovskite solar cells (PSCs) are regarded as a highly promising next-generation photovoltaic technology owing to their tunable bandgap, high absorption coefficient, and long carrier diffusion length. Notably, after 2022, the field entered an efficiency explosion period driven by the precision of anti-solvent engineering, standardization of annealing processes, rational design of additive molecules, and optimization of solvent systems, with single-junction certified efficiencies rapidly surpassing 25%. Nevertheless, non-radiative recombination arising from grain boundaries, point defects, and interfacial states in solution-processed polycrystalline films remains the central bottleneck constraining efficiency, while additive engineering—an effective strategy for regulating crystallization kinetics and passivating defects—is also becoming increasingly refined. This study focuses on experimental data of perovskite solar cells from 2022–2025 and constructs a temporal feature concatenation framework: 128-dimensional annual temporal embeddings extracted via Transformer self-attention are concatenated with 106-dimensional raw features (55 basic features + 51 functional group fingerprints) to form 234-dimensional augmented features, which are fed into LightGBM, achieving R² = 0.906 ± 0.008 in five-fold cross-validation. Further concatenation of LSTM, GRU, and Transformer triple embeddings to 490 dimensions raises R² to 0.926 ± 0.007, an absolute improvement of 0.221 in R² over the 106-dimensional static baseline (from 0.705 to 0.926). Additionally, Flow Matching microenvironment modeling is introduced to learn continuous probability paths of formulation distributions in a 50-dimensional latent space. Perturbation analysis reveals that the main solvent backbone and solvent components contribute most significantly to PCE enhancement; the optimal transport path requires only an average of 3.9 steps to guide low-efficiency formulations (<15%) toward the high-efficiency region (≥22%). Virtual formulation screening validated against the literature database achieves an average PCE improvement of +0.65%. Functional group commonality analysis demonstrates perfect overlap in the top-10 high-frequency functional groups between real and generated formulations, confirming that the model captures the scientific principles underlying high-efficiency PSCs. This framework provides an interpretable computational paradigm for data-driven optimization of high-efficiency perovskite formulations.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.