Interpretable multi-head attention autoencoder based transformer dopant selection model for photoelectrochemical water splitting: The case study of WO3

Tungsten oxide (WO₃) is a key photoanode for the photoelectrochemical water splitting (PEC) application; however, it comes with the challenges of lower efficiency and larger photogenerated charge recombination. Dopant selection is a major PEC performance issue. This study innovatively guide the selection of photoanode dopant utilizing multi-head attention based autoencoder transformer (MAT) model, which can reveal correlations between extensive dopant features and doped photoanode for high-performing PEC systems. The model was trained using 25 different metal dopants to predict suitable dopants that can enhance PEC photoanode performance. Furthermore, an attention-based framework is proposed to describe the relationship between the intrinsic features and doped element. The MAT model achieved the highest R² scores of 0.94 for estimation photocurrent density. SHAP (SHapley Additive exPlanations) analysis indicates that critical features, including electronegativity, electron affinity, and molecular bond energy, play a significant role in dopant selection and PEC performance. The attention weights effectively assessed the significance of dopant features. This integrated strategy not only improves the ability to anticipate dopants for enhanced PEC photoanodes, but also lays the groundwork for enhancing performance in related PEC systems as well as broaden the applicability of this framework to additional energy conversion technologies.