Perovskite models stumble on new alloys: encoding solid-solution continuity cuts unseen-alloy error by 15.7–19.0%

Machine-learning prediction for perovskite solar cells now targets unrecorded alloys and unfamiliar cell architectures, rather than interpolating only among material labels already seen. Earlier models mostly treated device layers as arbitrary categorical labels and represented absorbers either as featureless composition vectors or as rigid crystal graphs requiring a pre-solved unit cell, so they could not generalize to unrecorded alloys or unfamiliar cell architectures. Under an institution-grouped protocol where test devices come exclusively from unseen laboratories, an untuned default configuration of V-PaSS achieved the lowest error among fourteen tested models, 2.861% MAE, a 5.2% relative reduction over the gradient-boosted baseline at 3.017%, and better than the best tuned benchmark at 2.912%.

The model treats mixed-cation, mixed-halide absorbers as a continuous solid solution, so a new alloy lying between known endpoints still receives a prediction. It interpolates relaxed crystal graphs of 23 pure $ABX_3$ endpoint compounds via Vegard's law with learned bandgap bowing, and on the device side couples layer sequence and cell polarity (n-i-p or p-i-n) with interfacial band offsets, encoding chemical descriptions with a scientific language model. The lead is widest where test devices differ most from training: error falls by 15.7–19.0% on unseen absorber alloys, by 16.1% on mixed absorbers from new laboratories, and the margin over the baseline widens by 3.6 times on devices with unobserved transport layers.

This report covers the summary: V-PaSS targets perovskite solar cells, represents mixed-cation, mixed-halide absorbers by interpolating relaxed crystal graphs of 23 pure $ABX_3$ endpoint compounds, and encodes layer sequence, cell polarity, interfacial band offsets and chemical descriptions together. Evaluation uses an institution-grouped protocol where test devices come exclusively from unseen laboratories, compared against fourteen tested models. It is aimed at researchers and engineers interested in materials machine-learning extrapolation and perovskite device screening.

The summary reports results for an untuned default configuration against fourteen tested models, plus breakdowns on unseen alloys, mixed absorbers from new laboratories, and unobserved transport layers; these numbers come from that grouped protocol, and the summary does not state performance under other data splits or device types.