Physics-guided transfer learning for Bayesian optimization of chemical port-Hamiltonian systems

Bayesian optimization (BO) has emerged as a powerful black-box optimization approach for complex systems, making sequential decisions through Gaussian process (GP) models to explore complex search spaces. However, conventional BO faces certain challenges when applies to optimizations of chemical systems, particularly with limited measurement data and physical constraints. This paper proposes an adaptive framework combining transfer learning with physics-informed GP to enhance BO performance for chemical process optimization. By incorporating physics-based priors through Gaussian Process Port-Hamiltonian Systems (GP-PHS) in the point-by-point transfer learning methodology, the proposed approach dynamically leverages knowledge from related source domains while satisfying physical constrains. The framework’s effectiveness is demonstrated across three chemical systems including a water tank, an electrochemical cell, and an isothermal continuous stirred tank reactor (CSTR). Results show improvements in both optimization accuracy and convergence speed compared to traditional BO methods. This proposed approach bridges the gap between data-driven optimization and physical principles, offering a robust solution for complex chemical system optimization under data scarcity.