An Adaptive Thermal Finite Element Simulation of Direct Energy Deposition With Reinforcement Learning: A Conceptual Framework

Abstract

During the last decades, metal additive manufacturing (AM) technology has transitioned from rapid prototyping application to industrial adoption owing to its flexibility in product design, tooling, and process planning. Thus, understanding the behavior, interaction, and influence of the involved processing parameters on the overall AM production system in order to obtain high-quality parts and stabilized manufacturing process is crucial. Despite many advantages of the AM technologies, difficulties arise due to modelling the complex nature of the process-structure-property relations, which prevents its wide utilization in various industrial sectors. It is known that many of the most important defects in direct energy deposition (DED) are associated with the volume and timescales of the evolving melt pool. Thus, the development of methodologies for monitoring, and controlling the melt pool is critical. In this study, an adaptive numerical transient solution is developed, which is fed from the set of experiments for single-track scanning of super-alloy Inconel 625 on the hot-tempered steel type 42CrMo4. An established exponential formula based on the response surface methodology (RSM) that quantifies the influence of process parameters and geometries of deposited layers from experiments are considered to activate the volume fraction of passive elements in the finite element discretization. By resorting to the FORTRAN language framework capabilities, commercial finite element method software ABAQUS has been steered in order to control unfavorable defects induced by localized rapid heating and cooling, and unstable volume of the melt pool. A thermodynamic consistent phase-field model is coupled with a transient thermal simulation to track the material history. A Lagrangian description for the spatial and time discretization is used. The goal is to present a closed-loop approach to track the melt pool morphology and temperature to a reference deposition volume profile which is established based on deep reinforcement learning (RL) architecture aiming to avoid instabilities, defects and anomalies by controlling the laser power density adaptability. Despite the small number of iterations during RL model training, the agent was able to learn the desired behaviour and two different reward functions were evaluated. This approach allows us to show the possibility of using RL with openAI Gym for process control and its interconnection with ABAQUS framework to train a model first in a simulation environment, and thus take advantage of RL capabilities without creating waste or machine time in real-world.