Machine learning based prediction of optical properties in two-dimensional Mo-W-S-Se-Te transition-metal dichalcogenide alloys through physics-informed sampling
Two-dimensional transition-metal dichalcogenide (TMD) alloys provide a compositionally tunable platform for controlling the optical and electronic properties. However, systematic prediction of their dielectric response across multicomponent alloy spaces remains challenging owing to the combinatorial cost of first-principles calculations. In this work, we combined ab initio optical-property calculations with a tabular foundation-model regression to predict the real and imaginary components of the frequency-dependent dielectric function for Mo-W-S-Se-Te TMD alloys. A dataset of 99 alloy structures spanning binary, ternary, quaternary, and quinary compositions was generated using density functional theory (DFT). The resulting polarization-dependent dielectric spectra were used to train a tabular prior-fitted network (TabPFN) and evaluated against the conventional Extra Trees and XGBoost models. To accommodate the in-context capacity limit of TabPFN, we introduced a non-uniform, physics-informed energy subsampling strategy that concentrates sampling in the optically active region above the band gap, where interband absorption is strongest. Trained solely on quaternary alloys, our TabPFN reconstructed the dielectric spectra of held-out quaternary compositions with an R2 > 0.98 and a mean absolute error below 0.10 for all four dielectric components, outperforming both baselines while requiring no gradient-based training or hyperparameter tuning. Our model further predicted derived optical quantities, including refractive index, extinction coefficient, and absorption coefficient. Additionally, our model generalized in a zero-shot manner to binary, ternary, and quinary alloys absent from the training set, with quinary predictions achieving an R2 > 0.97.