Uncertainty-aware model-enhanced soft actor-critic for conformal coating parameter optimization

Abstract

Abstract Conformal coating optimization requires control of multiple application parameters to achieve target thickness and coverage uniformity, with spray atomization settings (atomization valve control) playing a crucial role in droplet formation and material deposition. Traditional manual tuning yields a low success rate and wastes significant materials through trial-and-error. We introduce Model-Enhanced Soft Actor-Critic (ME-SAC), a reinforcement learning approach that combines a hybrid dynamics model with uncertainty-aware planning. Our dynamics model integrates polynomial ridge regression with a residual neural ensemble, achieving R $$^{2}$$ = 0.70 using 5-fold cross-validation on the experimental dataset. Unlike conventional model-based RL that trains on synthetic data, ME-SAC uses the learned dynamics model for uncertainty-aware candidate-action evaluation without generating additional synthetic rollouts for replay-buffer augmentation. The RL agents are trained and evaluated through interactions with a surrogate coating environment constructed from 158 physical coating experiments. The planning module evaluates candidate actions using a conservative objective that penalizes high-uncertainty predictions, ensuring safe exploration. Experimental results demonstrate that ME-SAC achieves 95.0% success rate compared to 74.0% for model-free SAC baseline, reaching 90% performance using approximately 600 surrogate-environment interactions while baseline requires >1600 samples without achieving this threshold (>2.67 $$\times$$ sample efficiency). In the evaluated training runs, ME-SAC maintained performance without the mid-training degradation observed for the baseline.