[Paper Review] Hybrid thermal modeling of additive manufacturing processes using physics-informed neural networks for temperature prediction and parameter identification

Shuheng Liao, Tianju Xue+4 more|arXiv (Cornell University)|Jun 15, 2022

Thermography and Photoacoustic Techniques4 citations

TL;DR

This paper proposes a physics-informed neural network (PINN)-based hybrid thermal modeling framework for metal additive manufacturing that fuses infrared temperature measurements with heat transfer physics to predict full-field temperature histories and identify unknown material/process parameters. The method achieves high accuracy in both numerical and experimental cases, with RMSE errors below 50 K in real-world data, demonstrating its potential for real-time process control and inverse parameter identification.

ABSTRACT

Understanding the thermal behavior of additive manufacturing (AM) processes is crucial for enhancing the quality control and enabling customized process design. Most purely physics-based computational models suffer from intensive computational costs and the need of calibrating unknown parameters, thus not suitable for online control and iterative design application. Data-driven models taking advantage of the latest developed computational tools can serve as a more efficient surrogate, but they are usually trained over a large amount of simulation data and often fail to effectively use small but high-quality experimental data. In this work, we developed a hybrid physics-based data-driven thermal modeling approach of AM processes using physics-informed neural networks. Specifically, partially observed temperature data measured from an infrared camera is combined with the physics laws to predict full-field temperature history and to discover unknown material and process parameters. In the numerical and experimental examples, the effectiveness of adding auxiliary training data and using the pretrained model on training efficiency and prediction accuracy, as well as the ability to identify unknown parameters with partially observed data, are demonstrated. The results show that the hybrid thermal model can effectively identify unknown parameters and capture the full-field temperature accurately, and thus it has the potential to be used in iterative process design and real-time process control of AM.

Motivation & Objective

• To address the high computational cost and parameter calibration challenges of purely physics-based models in additive manufacturing.
• To overcome the data-hungry and generalization-limited nature of purely data-driven models by integrating sparse, high-quality experimental data with physical laws.
• To develop a surrogate model that enables fast, accurate full-field temperature prediction and inverse parameter identification in metal AM processes.
• To enable online process control and iterative design by combining physics-informed neural networks with experimental measurements.

Proposed method

• A physics-informed neural network (PINN) is trained to solve the heat transfer partial differential equation (PDE) governing thermal behavior in directed energy deposition (DED) processes.
• The PINN loss function incorporates both the PDE residual and boundary/initial conditions, ensuring physical consistency.
• Partially observed temperature data from infrared (IR) camera measurements are incorporated as supervised data points in the loss function to guide the model.
• Auxiliary training data—generated from finite element method (FEM) simulations—are used to improve convergence and accuracy.
• A pretrained PINN model is fine-tuned using experimental data, accelerating training and enhancing generalization.
• The framework enables inverse parameter identification by treating unknown thermal properties (e.g., laser absorptivity, thermal conductivity) as trainable variables.

Experimental results

Research questions

• RQ1 Can a PINN-based model accurately predict full-field temperature histories in metal AM using only partially observed IR temperature data?
• RQ2 How does incorporating auxiliary FEM-generated data affect the convergence and accuracy of the PINN model?
• RQ3 To what extent can pretraining accelerate the training of PINNs for repeated simulations with varying material or process parameters?
• RQ4 Can the hybrid PINN model effectively identify unknown material and process parameters from sparse experimental measurements?
• RQ5 How well does the framework generalize to real experimental data compared to synthetic or simulated data?

Key findings

• The PINN model achieved an RMSE of 14.07 K compared to FEM simulation results in the numerical forward problem, demonstrating its effectiveness as a surrogate model.
• Adding auxiliary FEM data reduced the number of training epochs by two-thirds to reach the same accuracy, significantly improving training efficiency.
• Pretraining reduced the number of epochs required to reach target accuracy by more than 80%, enabling rapid adaptation to new material or process conditions.
• In the inverse problem, the model identified laser absorptivity, heat capacity, and thermal conductivity with less than 5% error using synthetic noisy IR data.
• In the experimental case, the model predicted full-field temperature with an RMSE of 47.28 K compared to measured IR data, proving its capability to fuse real experimental data with physics.
• The hybrid framework enables flexible, arbitrary fusion of experimental data into the PINN model, offering a robust solution for modeling complex thermal behavior in AM.