PBE-assisted machine learning of HSE06+SOC photovoltaic efficiency in halide perovskites

Abstract Hybrid-functional calculations with spin–orbit coupling (HSE06 + SOC) provide more accurate electronic properties of halide perovskites than semilocal PBE calculations but are substantially more expensive. Here, we assess whether low-cost PBE information improves machine-learning (ML) prediction of HSE06 + SOC band gaps and spectroscopic limited maximum efficiency (SLME). Using a public dataset of 243 pure and mixed ABX₃ perovskites, we compare seven ML model families using nested cross-validation with random, chemical-system-grouped, and B-site-grouped splits. PBE inputs substantially improve SLME prediction: the mean absolute error decreases from 0.68 to 0.45 percentage points on chemical-system-grouped splits, with the best model achieving 0.36 percentage points. Δ-learning provides no advantage over direct learning. However, performance deteriorates when the B-site chemistry is absent from training, and conformal intervals become under-calibrated. An audit of the source labels further shows that the HSE06 + SOC SLME is strongly constrained by the HSE06 + SOC band gap and approaches a saturation near 16%, because the reported spectrum was generated by shifting the PBE absorption spectrum. These findings highlight both the value and limitations of PBE-assisted ML for multi-fidelity prediction of perovskite properties.