AdditiveGDL: generative deep learning for predicting local thermal distributions in metal 3D-printed layers

In metal additive manufacturing, understanding the intricate thermal dynamics during the printing process is crucial. These dynamics significantly impact key factors such as microstructure, mechanical properties, fatigue life, residual stresses, dimensional accuracy, shape integrity, and surface quality. The laser toolpath governs heat accumulation, with pronounced effects at particular geometric features such as sharp corners and thin walls. Therefore, the ability to accurately predict heat distribution based on the laser toolpath is vital for optimizing the printing process and ensuring the desired material and structural performance. While numerical simulations using finite element methods are common, they can be computationally prohibitive for complex parts. In this work, we propose a method that leverages Constrained Generative Adversarial Networks to predict the local thermal distribution given the laser toolpath. This data-driven method stands out by reducing computational costs while maintaining high accuracy in thermal predictions. By identifying critical regions of heat accumulation, we can optimize geometry and scan paths, ultimately enhancing the quality and reliability of 3D-printed metal components. The results indicate that generative deep learning effectively predicts heat accumulation in geometries that were not part of the training dataset, achieving a mean L2 norm error of less than 0.0036 in normalized images. This approach offers a computationally efficient alternative to finite element methods, significantly reducing calculation time.