Multikernel Gaussian process regression with automatic relevance determination for predictive modeling and sensitivity analysis in laser cutting processes

A novel multioutput Gaussian process regression (GPR) framework is proposed to address multiphysics modeling challenges in laser cutting. Multikernel learning is integrated with automatic relevance determination (ARD) for enhanced interpretability. High-precision joint prediction of temperature, stress, and displacement fields is achieved under limited-sample conditions. An aluminum alloy case study is implemented, where a finite element dataset is constructed. Training and testing samples are generated through orthogonal experiments and Latin hypercube sampling. The GPR-MAK (multi automatic kernel) model is demonstrated to achieve 98.10% average accuracy (MAE = 11.36 and RMSE = 15.51), significantly outperforming baseline methods. Key parameter influences are revealed through ARD, partial dependence plot, and SHapley Additive exPlanations analyses: Temperature is found to be dominated by laser power and gas pressure, while stress is primarily controlled by scanning speed and defocus amount. These findings are shown to be consistent with fundamental thermomechanical principles. The framework is validated to provide both theoretical insights and practical optimization guidance for laser cutting processes.