Physics-guided Gaussian process framework for data-efficient optimization of APCVD growth of 2D semiconductor WS2

Layer-controlled growth of two-dimensional transition-metal dichalcogenides (2D TMDs), such as Tungsten Disulfide (WS 2 ), remains challenging due to the nonlinear interplay among multiple parameters associated with the atmospheric pressure chemical vapor deposition (APCVD) growth technique. In particular, achieving large-area few-layer flakes of TMDs is significantly more challenging than growing bulk using conventional APCVD method. Here, we present a physics-guided Gaussian Process Regression (GPR) framework for data-efficient exploration of the APCVD growth parameter space of WS 2 . An initial database of 12 experimentally measured growths is used to construct a GPR surrogate model, with WS 2 areal flake coverage quantified by optical-image analysis. Virtual candidate conditions are subsequently generated by Monte Carlo sampling and ranked using a composite acquisition score incorporating a domain-guided density target, predictive uncertainty, and standardized parameter-space novelty. Machine Learning (ML) generated candidates are maintained separately from the experimental database and are not used as pseudo-labels for model training. Leave-one-out cross-validation (LOOCV) of the 12 experimental observations gives an R 2 of 0.719, mean absolute error (MAE) of 8.23 percentage points, root mean square error (RMSE) of 11.96 percentage points, and Pearson correlation coefficient of 0.854. The framework identifies promising regions of the APCVD parameter space while explicitly quantifying predictive uncertainty. Further, the optical microscopy, Raman spectroscopy, and photoluminescence measurements are used to characterize experimentally obtained WS 2 crystals.