Leveraging Physics-Informed Neural Networks for Efficient Tolerance Analysis

Kopatsch J, Wartzack S, Götz S (2026)


Publication Type: Journal article

Publication year: 2026

Journal

Original Authors: Jan Kopatsch, Sandro Wartzack, Stefan Goetz

Book Volume: 145

Pages Range: 135-140

DOI: 10.1016/j.procir.2026.03.130

Abstract

The effect of manufacturing-induced deviations on non-geometric Key Characteristics (KC) within the scope of tolerance analysis is determined by the iterative solving of partial or ordinary differential equations. The statistical assurance of the KCs coupled with the computationally expensive evaluation of differential equations results in unacceptably long computation times. Data-driven surrogate models provide a solution to this problem, as they only require the numerical solution of the differential equations for their training and are capable to approximate the solutions after training with low computational effort. However, the accuracy of the predictions of the surrogate model depends on the quantity and quality of information about the system that is considered during its training. Physics-informed neural networks pursue this approach by incorporating physical principles during network training. This additional system knowledge aims to increase prediction accuracy in comparison to purely data-driven approaches. In tolerance analyses, prediction accuracy in the evaluation of non-geometric KCs is crucial. Thus, this paper proposes an approach that incorporates physics-based knowledge for surrogate modeling to increase the efficiency of tolerance analysis with non-geometric KCs, namely the computation time, approximation error and model uncertainty. A review of the approach’s efficiency in comparison to non-physics-informed neural networks is conducted using an academic demonstrator, leading to 72 % less tolerance analysis time with comparable accuracy and uncertainty measures.

Authors with CRIS profile

How to cite

APA:

Kopatsch, J., Wartzack, S., & Götz, S. (2026). Leveraging Physics-Informed Neural Networks for Efficient Tolerance Analysis. Procedia CIRP, 145, 135-140. https://doi.org/10.1016/j.procir.2026.03.130

MLA:

Kopatsch, Jan, Sandro Wartzack, and Stefan Götz. "Leveraging Physics-Informed Neural Networks for Efficient Tolerance Analysis." Procedia CIRP 145 (2026): 135-140.

BibTeX: Download