Reinforcement learning-based toolpath optimisation with 3D U-Net driven rapid thermal prediction

Abstract

Fabricating complex geometries in additive manufacturing induces severe thermal gradients, resulting in significant warpage in the bottom layers. Although optimising toolpath strategies is essential for mitigating these issues, the prohibitive computational costs associated with calculating thermal distributions for complex 3D geometries limit current approaches. To overcome this limitation, we propose a novel framework that accelerates reinforcement learning (RL)-based toolpath optimisation by integrating a 3D U-Net as a surrogate model. The 3D U-Net is trained on data generated via finite difference method (FDM) simulation to enable rapid thermal predictions. This acceleration enables the RL agent to efficiently derive optimal toolpaths to minimise thermal distortion and ensure thermal uniformity. Experimental results indicate that the 3D U-Net accelerates thermal field prediction by reducing computation time by over 99.8% compared to FDM simulations. This rapid inference capability compresses the process from hours to seconds and enables efficient process planning. Based on these fast predictions, the RL-optimised toolpaths achieved a warpage reduction of 95.99% on the pyramidal geometry compared to the conventional toolpath. This study establishes a scalable paradigm for intelligent manufacturing and provides a robust solution to thermal distortion in complex arbitrary geometries.