Joint Optical–Thermal Optimization of Laser Processing Parameters Using Bayesian Optimization and Neural Emulators

Abstract

The precise control of laser material processing, particularly in regimes such as laser welding, cutting, and additive manufacturing, necessitates a rigorous understanding of the complex interplay between optical beam propagation and thermal material response. Traditional methods for parameter optimization, which rely heavily on trial-and-error experimentation or computationally expensive Finite Element Method simulations, often fail to converge efficiently within the high-dimensional parameter space inherent to modern laser systems. This paper proposes a novel framework for the Joint Optical–Thermal Optimization of processing parameters by leveraging a hybrid architecture comprising Bayesian Optimization and deep Neural Emulators. We introduce a data-driven methodology where a high-fidelity physics-based model, coupling ray-tracing optical solvers with transient heat transfer equations, generates a training corpus. This data trains a deep neural network surrogate—the Neural Emulator—capable of predicting melt pool geometries and thermal gradients with near-instantaneous inference times. Subsequently, a Bayesian Optimization engine utilizes this emulator to navigate the parameter space, balancing exploration and exploitation to identify optimal process windows. Our results demonstrate that this approach reduces the computational cost of optimization by three orders of magnitude compared to direct numerical simulation while maintaining predictive accuracy within two percent of experimental baselines. This work bridges the gap between optical setup and thermal history, offering a scalable pathway for autonomous laser process manufacturing.