Multi-layer thermal history prediction framework for directed energy deposition based on extended physics-informed neural networks (XPINN)
File(s)
Published version
Author(s)
Peng, Bohan
Panesar, Ajit
Type
Journal Article
Abstract
This paper presents an eXtended physics-informed neural networks (XPINN)-based framework for predicting the temperature history during a multi-layer Directed Energy Deposition (DED) process. The proposed XPINN-based framework, advancing from its PINN-based counterpart, demonstrates significant accuracy improvement, around
reduction in RMSE and maximum absolute error, and extended capability of temperature history prediction with domain decomposition for more complex configurations such as interpass time, void, and interruption of scan that are prevalent in real-life DED designs. It is validated via a series of 2D benchmark tests against numerical simulations with an increasing degree of complexity. The effect of different domain decompositions is compared and discussed. Strategies that improve the training outcome are also proposed and analysed. With the enhanced capability of working on more complex configurations while retaining the characteristic availability of derivative information, the proposed framework brings process-ware design optimisation based on scientific machine learning (SciML) techniques one step closer to the application to real-life additive manufacturing applications.
reduction in RMSE and maximum absolute error, and extended capability of temperature history prediction with domain decomposition for more complex configurations such as interpass time, void, and interruption of scan that are prevalent in real-life DED designs. It is validated via a series of 2D benchmark tests against numerical simulations with an increasing degree of complexity. The effect of different domain decompositions is compared and discussed. Strategies that improve the training outcome are also proposed and analysed. With the enhanced capability of working on more complex configurations while retaining the characteristic availability of derivative information, the proposed framework brings process-ware design optimisation based on scientific machine learning (SciML) techniques one step closer to the application to real-life additive manufacturing applications.
Date Issued
2025-07-25
Date Acceptance
2025-08-30
Citation
Additive Manufacturing, 2025, 110
ISSN
2214-8604
Publisher
Elsevier BV
Journal / Book Title
Additive Manufacturing
Volume
110
Copyright Statement
© 2025 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by nc-nd/4.0/).
Publication Status
Published
Article Number
104953
Date Publish Online
2025-09-01