Session: 03-02-01: Advanced Manufacturing

Paper Number: 151985

151985 - Enhanced Temperature Prediction in Wire Arc-Directed Energy Deposition Using Physics-Informed Neural Networks 

Wire Arc Directed Energy Deposition (WA-DED) is an advanced metal 3D printing process that enables the production of near-net-shape metal parts by depositing material layer-upon-layer at a high deposition rate. While this technique offers excellent potential for creating complex geometries with both material and time efficiencies, it also poses a significant challenge: maintaining precise temperature control. This is crucial because temperature directly impacts the desired microstructure and mechanical properties of the fabricated parts. Accurate temperature prediction during the process can enhance control over the deposition quality, reducing defects such as warping, residual stresses, or inconsistent microstructure.

To address the need for accurate temperature monitoring during the WA-DED process, we employed Physics-Informed Neural Networks (PINNs). Unlike conventional neural networks, which heavily rely on experimental data, PINNs integrate fundamental physical equations, such as the heat transfer equation, directly into the training process, specifically as a form of loss function. This way, PINNs not only minimize the discrepancy between observation and prediction in the objective function but also satisfy any given governing initial/boundary conditions. PINNs can offer several benefits in predicting temperature in WA-DED, including enhanced accuracy, data efficiency, and the ability to produce physically admissible results, even in scenarios with limited labeled data where conventional neural networks would struggle to generalize. 

In this study, we used PINNs to predict the temperature evolution in a metal substrate subjected to a moving heat source, simulating the heat transfer process in WA-DED. To prepare training data for PINN, we sampled the labeled data from Finite Element Models (FEMs) simulation results, which were calibrated using the temperature measurements obtained from thermocouples at specific off-bead locations as well as infrared (IR) thermal profile data for the melt pool area. The FEMs simulation results served as the ground truth for training and validation of the PINN model. To ensure consistency, the same heat source model, such as the Goldak double-ellipsoid equation, and boundary conditions used in FEMs were also applied to PINN simulations. 

To optimize the number of unlabeled samples required to satisfy the governing equations, also known as collocation points (CPs), we adopted region-specific sampling strategies based on the temperature variation characteristics. In areas away from the heat source center, where temperature changes are gradual, we utilized Latin Hypercube Sampling (LHS). For the center region, where temperature changes more rapidly near the center, we implemented a concentrated sampling scheme based on a time-dependent multivariate Gaussian function, resembling the distribution used in the heat source equation. This targeted sampling approach improved the model’s ability to capture temperature variations accurately, especially the center of the heat source. 

Additionally, we also investigated several neural network architectures and their impact on PINNs' ability to achieve more accurate thermal predictions. Key hyperparameters optimized during this process include the number of layers and neurons, the number of blocks for ResNet, grid points for the Kolmogorov-Arnold Network (KAN), and the number of CPs. The hyperparameters to optimize include the number of layers/neurons and the number of CPs for all architectures, the number of blocks specifically for ResNet, and grid points for KAN. The optimization algorithm used for this study is the Tree-Structured Parzen Estimator (TPE) algorithm. In conjunction with the concentrated sampling scheme, our fully trained PINN models with different architectures achieved accurate predictions of approximately 2% average percent error over the FEM models. 

Presenting Author: Haiye (Justin) Xie Center for Advanced Vehicular Systems (CAVS), Mississippi State University

Presenting Author Biography: Justin is a graduate student in Computational Engineering at Mississippi State University (MSU), studying machine learning. His work involves running machine learning simulations and neural network tuning using high-performance computing (HPC) systems. Justin completed his undergraduate studies in Electrical Engineering at MSU and is now pursuing a master's to further his expertise in computational techniques and machine learning applications.