Veröffentlichungen und Präsentationen

This work presents the development and evaluation of surrogate models based on deep neural networks (DNNs) for accelerating numerical simulations performed in COMSOL Multiphysics. The main objective of the study is to investigate how numerical simulation parameters and neural network architecture influence the predictive accuracy, robustness, and computational efficiency of data-driven surrogate models. The research addresses the challenge of reducing the computational cost associated with repeated finite element method (FEM) simulations while preserving acceptable accuracy for engineering analysis and optimization tasks. The physical model used in this study is the grading ring optimization model obtained from the Application Library of COMSOL Multiphysics. The model was configured and solved within the COMSOL environment using FEM-based stationary simulations. Parametric sweeps were performed to generate datasets for training and validation of the neural networks. Particular attention was given to the influence of numerical settings on data quality, including mesh density, finite element discretization quality and convergence behaviour (Figure 1). The study analyses the trade-off between simulation accuracy and computational cost, emphasizing how these factors affect the reliability of the generated training data and, consequently, the performance of the surrogate models (Figure 2). The work also investigates different DNN architectures available in COMSOL, focusing on the influence of network depth, number of neurons, and activation functions on the generalization capability of the models. Surrogate models created within COMSOL were systematically compared using validation datasets generated from the FEM simulations. The obtained results demonstrate that careful preparation of simulation datasets and appropriate selection of DNN architecture are essential for achieving accurate and stable surrogate models. The study confirms that surrogate modelling can significantly reduce computation time compared to repeated FEM simulations while maintaining good agreement with reference COMSOL results. The comparison between FEM solutions and ANN predictions shows that properly trained models can effectively approximate the behaviour of the original physical system. The presented methodology highlights the importance of combining numerical simulation expertise with machine learning techniques in order to improve engineering workflows. The results indicate that integrating physics-based simulations with neural-network-based surrogate models can provide an efficient framework for optimization, sensitivity analysis, and rapid prediction tasks in computational engineering applications.