The Deep Operator Finite Element Method: A Stable and Efficient Neural Operator for Elasticity Problems with Variable Loading
Authors:
- Wei XiongWXWei Xiong
- Xiangyun LongXiangyun Long
- Jinwu LiJLJinwu Li
- Stéphane P. A. BordasSBStéphane P. A. Bordas
- Chao JiangCJChao Jiang
Preprint
- 9 February 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Learning solution operators for parametric partial differential equations is crucial for many-query engineering tasks. However, existing neural operators, especially physics-informed variants, are often plagued by training instability and efficiency. To address this challenge, this paper introduces a novel deep operator finite element method (DOFEM) for stable training and real-time prediction of elasticity problems with variable loading. The DOFEM uniquely integrates the DeepONet architecture with a variational energy functional constructed directly from the finite element stiffness matrix. This design allows its physics-informed loss function to circumvent high-order derivative computations and numerical integration errors, ensuring exceptional training stability. Numerical examples demonstrate that the DOFEM significantly outperforms the energy-based Physics-Informed DeepONet in both stability and accuracy, remaining robust in three-dimensional problems where the latter fails to converge. Furthermore, the model exhibits strong generalization to unseen loads and provides smoother, physically regularized solutions for singular loads compared to the standard finite element method. DOFEM thus presents an efficient and reliable paradigm for the rapid analysis of complex engineering systems under diverse operating conditions.Keywords
This publication has 27 references indexed in Scilit:
- DCEM: A deep complementary energy method for linear elasticityInternational Journal for Numerical Methods in Engineering, 2024
- CENN: Conservative energy method based on neural networks with subdomains for solving variational problems involving heterogeneous and complex geometriesComputer Methods in Applied Mechanics and Engineering, 2022
- Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s NextJournal of Scientific Computing, 2022
- A physics-informed variational DeepONet for predicting crack path in quasi-brittle materialsComputer Methods in Applied Mechanics and Engineering, 2022
- Physics-informed neural network for modelling the thermochemical curing process of composite-tool systems during manufactureComputer Methods in Applied Mechanics and Engineering, 2021
- hp-VPINNs: Variational physics-informed neural networks with domain decompositionComputer Methods in Applied Mechanics and Engineering, 2021
- Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problemsComputer Methods in Applied Mechanics and Engineering, 2020
- An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applicationsComputer Methods in Applied Mechanics and Engineering, 2020
- Neural algorithm for solving differential equationsJournal of Computational Physics, 1990
- On the limited memory BFGS method for large scale optimizationMathematical Programming, 1989