Prediction of Viscosity and Film Thickness for Laser Cutting Protective Fluid Based on Physics-Enhanced Machine Learning

Traditional data-driven models are frequently restricted by small-sample limitations due to the high acquisition costs of film thickness data and the complex nonlinear coupling between formulation rheological properties and processing dynamics in polymer spin-coating processes. In this study, we developed a physics-enhanced machine learning framework to predict the viscosity and final film thickness of laser-cutting protective fluids. Utilizing the classical Meyerhofer film-forming kinetics equation as a bridge, we experimentally calibrated the characteristic constant. Consequently, we mathematically deduced and expanded the limited experimental viscosity data into a theoretical film thickness data set comprising multiple gradient spin speeds. Through this approach, model training is guided under the constraints of physical mechanisms, and the risk of overfitting is effectively reduced. Based on the evaluation of six regression algorithms using this augmented data set, the highest accuracy in viscosity prediction is achieved by the GBDT algorithm, yielding a test set coefficient of determination R 2 of 0.9916 and an Mean Squared Error (MSE) of 0.0010. Furthermore, the optimal performance in film thickness prediction is demonstrated by the XGBoost algorithm, reaching an R 2 of 0.9995 and an MSE of 2.3403 × 10 –5 . The mapping relationship between formulation processing and film thickness is quantitatively established by this physics-enhanced framework, providing a reliable predictive tool for polymer thin-film processing that eliminates the necessity for extensive trial-and-error experiments.