Data-driven prediction of the high-dimensional thermal history in directed energy deposition processes via recurrent neural networks

Abstract

Directed Energy Deposition (DED) is a growing additive manufacturing technology due to its superior properties such as build flexibility at multiple scales and limited waste. However, both experimental and physics-based models have limitations in providing accurate and computationally efficient predictions of process outcomes, which is essential for real-time process control and optimization. In this work, a recurrent neural network (RNN) structure with a Gated Recurrent Unit (GRU) formulation is proposed for predicting the high-dimensional thermal history in DED processes with variations in geometry, build dimensions, toolpath strategy, laser power and scan speed. Our results indicate that the model can accurately predict the thermal history of any given point of the DED build on a test-set database with limited training. The model’s general applicability and ability to accurately predict thermal histories has been demonstrated through two overarching tests conducted for long time spans and non-trained geometries.