Temporal probabilistic soft sensing and constrained Bayesian recipe screening for semiconductor CMP processes

Delayed metrology and changing process conditions make reliable material removal rate (MRR) prediction and recipe screening difficult in chemical–mechanical planarization (CMP). This study links temporal soft sensing, uncertainty calibration and constrained Bayesian screening within a reproducible manufacturing-data workflow. Light Gradient Boosting Machine (LightGBM) predicts MRR from wafer–stage features; historical residuals provide a process-state uncertainty scale, and a disjoint chronological calibration window determines prediction intervals. A candidate-dependent residual Gaussian process (GP) ranks recipe proxies within a shared empirical feasible region. On 396 time-out observations, the point predictor achieved R2 = 0.9543 and root mean square error (RMSE) = 5.9743; the guarded 95% interval attained 95.96% empirical coverage. In 20 model-in-the-loop scenarios, the proposed method achieved a 95.0% target-window hit rate and generated feasible candidates throughout the evaluation budget. Constrained baselines attained the same final hit rate, identifying the contribution of shared feasibility screening. The framework supports uncertainty-aware CMP recipe pre-screening through temporally separated predictive validation and independent evaluator-based decision analysis.