A synthetic-data-trained deep learning model for predicting residual stress and melt pool characteristics in laser melting with arbitrary beam profiles

• A DL model is developed to predict melt pool and residual stress from laser beam profiles. • The model is trained on synthetic data and generalizes to realistic profiles like Gaussian and Bessel. • A three-stage training strategy enables efficient learning and robust prediction. • The model provides rapid, simulation-free feedback for laser process design and optimization. • Parametric studies demonstrate its power as a design tool for tailoring beam shapes and intensities. This study presents a deep learning-based surrogate framework that emulates CFD outputs for laser melting processes under diverse laser beam intensity profiles. Arbitrary input beams were constructed using combinations of randomly selected piecewise hat functions, enabling the model to learn generalizable mappings for any laser intensity distribution. The model was first pretrained using analytically generated synthetic fields and subsequently trained on CFD-generated data to capture the coupled thermal, fluid, and mechanical responses of the melt pool. Its emulation capability was evaluated on realistic beam shapes—including Gaussian, top-hat, ring, and Bessel profiles—achieving high quantitative accuracy with a mean absolute error of 5.96 °C and R 2 of 0.996 for temperature, and 8.41 MPa and R 2 of 0.972 for von Mises residual stress. The surrogate further demonstrated strong robustness to experimentally measured, irregular, and noisy beam profiles. It employs a conditional generative adversarial network optimized using a hybrid loss function that combines L1 and masked L2 terms. Parametric studies further showed that the surrogate enables rapid evaluation of process outcomes, completing each emulation in approximately 0.013 s compared to the 22 h required for a single CFD simulation, thereby offering a practical and computationally efficient tool for beam-shaping design and optimization.