Generative Models Training based on FEM Simulations towards improved Defect Detection

Anastasios Tzelepakis, Nikolaos Dimitriou, Paschalis Charalampous, Konstantinos Tsongas, Dimitrios Tzetzis, Dimitris Tzovaras

Keywords: Defect prediction, Finite element method, Generative AI, Simulation, Physics-informed dataset generation

Abstract: Defect prediction and quality control are critical in manufacturing, where Convolutional Neural Networks (CNNs) demonstrate significant potential. However, traditional Finite Element Method (FEM) simulations, despite their accuracy, are hindered by high computational demands. This paper introduces a novel framework that integrates FEM simulations with generative CNNs to predict strain distributions during antenna manufacturing. By employing physics-informed dataset generation from FEM simulations, the proposed method trains a generative CNN to predict strain distributions during antenna manufacturing, enabling physically consistent strain predictions. Validated against FEM-calculated data, the framework demonstrates its efficacy in defect prevention, while addressing the limitations of traditional offline FEM capabilities. Furthermore, a comprehensive analysis of weight initialization and cost function choices, along with experimental validation, highlights the method’s efficiency, establishing a cost-effective and practical approach to integrating numerical simulations with CNN-based deep learning in manufacturing.

Submission Number: 69