Treschau M, Witzgall C, Wartzack S (2026)
Publication Language: English
Publication Type: Conference contribution, Conference Contribution
Publication year: 2026
Publisher: Springer Nature Switzerland
City/Town: Cham
Pages Range: 1-15
Conference Proceedings Title: Proceedings of the Munich Symposium on Lightweight Design 2025
DOI: 10.1007/978-3-032-43935-2_4
Finite element (FE) simulations are widely used in mechanical design but are computationally expensive due to large, high-dimensional equation systems. Recent advances in artificial intelligence enable physics-informed
neural networks (PINNs) to predict physical behavior with reduced computational effort. This paper applies PINNs to structural mechanical simulations using a statically loaded truss as an example. The PINN is intended to be trained using a hybrid strategy that combines data generated from FE simulations with embedded physical knowledge. A key challenge arises from the inhomogeneity of input data required for automatic differentiation and the need for uniformly sized training samples, currently resulting in a restricted PINN formulation. A parameter study of the elastostatic PINN shows that substantially fewer training data is required compared to conventional neural networks. Moreover, the PINN benefits from a scheduled learning rate and a progressively increasing network width tailored to the structure of the input and output data. Even this restricted approach achieves sufficient accuracy for a simplified elastostatic problem while requiring significantly less computation time than conventional FE simulations.
APA:
Treschau, M., Witzgall, C., & Wartzack, S. (2026). An Approach for Structural Mechanical Simulations with Physics-Informed Neural Networks Using the Example of a Truss. In Jasper Rieser, Klemens Rother, Philipp Höfer, Tobias Dickhut, Markus Zimmermann (Eds.), Proceedings of the Munich Symposium on Lightweight Design 2025 (pp. 1-15). München, DE: Cham: Springer Nature Switzerland.
MLA:
Treschau, Moritz, Christian Witzgall, and Sandro Wartzack. "An Approach for Structural Mechanical Simulations with Physics-Informed Neural Networks Using the Example of a Truss." Proceedings of the MSLD 2025, München Ed. Jasper Rieser, Klemens Rother, Philipp Höfer, Tobias Dickhut, Markus Zimmermann, Cham: Springer Nature Switzerland, 2026. 1-15.
BibTeX: Download