Experimental and Numerical Study of Millisecond Laser Micro-Processing Based on a Dual Neural Network System

Abstract To address the technical challenge of precisely controlling the depth of nonpenetrating structures in millisecond laser microprocessing, this study proposes a closed-loop prediction and simulation system based on dual neural network models. A feedforward backpropagation neural network is used to construct a nonlinear mapping model of “target features–process parameters–simulation results” to achieve high-precision control of blind-hole features in a nickel-based high-temperature alloy. The prediction model outputs the process parameters (peak power, duty ratio, and pulse number) by inputting the target depth and diameter. In contrast, the simulation model simulates the processing results to form a closed-loop calibration. The experiment was conducted with a millisecond quasi-continuous wave fiber laser, and datasets for training the network were obtained by a confocal microscope. Through model structure optimization, employing a prediction model with four hidden layers of 30 neurons and a simulation model with three hidden layers of 30 neurons, the nonlinear error caused by the thermal accumulation effect was effectively suppressed. The results show that the closed-loop simulation depth average error of the two models is 3.96%, while the actual processing verification depth average error is 4.52%, and the diameter average error is 3.77% and 4.59%, respectively. The study reveals the potential influence of the dynamic instability of the high-power molten pool on the model error, demonstrates the engineering applicability of the depth closed-loop control at ± 5% and provides an efficient and intelligent solution for complex microstructure processing in aerospace, energy equipment, and related fields.