Physics-informed neural networks for data-free surrogate modelling and engineering optimization – An example from composite manufacturing

• DOI: 10.1016/j.matdes.2023.112034
• Publication Date: 2023-05-26T01:02:22Z
• Authors: Tobias Würth; Constantin Krauß; Clemens Zimmerling; Luise Kärger
• Keywords: ddc:620; Artificial intelligence; Physics-based modeling; 0211 other engineering and technologies; 600; 02 engineering and technology; 620; Composite process optimization; Data-free surrogates; Machine learning; TA401-492; Mesh-free surrogate modeling; Physics-Informed Neural Networks; Materials of engineering and construction. Mechanics of materials; Engineering & allied operations; info:eu-repo/classification/ddc/620

Engineering components require an optimization of design and manufacturing parameters to achieve maximum performance – usually involving numerous physics-based simulations. Optimizing these parameters is a resource-intensive endeavor, though, especially in high-dimensional scenarios or for complex materials like fiber reinforced plastics. Surrogate models are able to reduce the computational effort, however, data generation still proves to be resource-intensive. Additionally, their data-driven nature may lead to physically implausible results in limit cases. As a remedy, physics-informed neural networks (PINNs) include known physics into the training for enhanced surrogate reliability. This allows to cast a physically consistent, data- and mesh-free manufacturing surrogate for variable process conditions and material parameters. The paper demonstrates how PINNs can be embedded in a design-framework to enhance process understanding, to devise engineering-interpretable processing windows and to support time-efficient process optimization at the example of a thermochemical manufacturing process with fiber-reinforced composite materials. In this work, an over 500-fold speed up of the process optimization is achieved compared to conventional approaches.