Deep learning solves perovskite kinetics inverse problem, revealing key parameters

bravo_abad · x · 2026-08-25

Wenning Chen and coauthors published a study in Nature Communications using deep learning to address the inverse problem in metal-halide perovskites. Traditional time-resolved photoluminescence (TrPL) infers carrier-recombination dynamics but suffers from non-uniqueness. The authors combined transient photoluminescence with excitation-dependent photoluminescence quantum yield to add independent physical constraints. They generated a large simulation-based training set from a mechanistic kinetic model and trained a neural network to simultaneously infer six coupled physical parameters, including bimolecular recombination, electron/hole trapping, trap density, and Auger recombination, achieving R² values between 0.905 and 0.996. Validated on experimental perovskite films, the inferred kinetics match independent device-level electrical characterization. The study highlights a key ML lesson: when an inverse problem is non-identifiable, a larger network cannot manufacture missing information; the measurement must be redesigned first to constrain hidden physics.

Original post →

More from Research

Research channel →