Interaktives 3D Design und Simulation - Basierend auf Neuralen Netzwerken

The presentation by Victor Oancea, Jing Bi, and Thomas Emmel focused on "Interactive 3D Design and Simulation Based on Neural Networks." This presentation highlighted the integration of machine learning (ML) and artificial intelligence (AI) in 3D design and simulation, emphasizing the potential of these technologies to enhance traditional Finite Element (FE) methods. The team demonstrated the efficiency of ML-trained model predictions compared to traditional FE methods. They showcased the creation of faithful 3D surrogate models using machine learning and neural networks. The presentation detailed six structure-centered challenge models, each with specific parameters and key performance indicators (KPIs), covering dynamics, statics, multiphysics, welding, battery safety, and buckling. A significant part of the presentation was dedicated to explaining how parametrized models (in terms of geometry, loads, materials) were prepared using Design of Experiments (DOE) methods. The training model's accuracy and loss were evaluated, with a focus on quasi-interactive design explorations, emphasizing that visualization took longer than computation. A case study on "Tube Crush" was presented, where an aluminum alloy crushing column with four design variables was simulated. The simulation compared the duration and computing resources required by Abaqus reference (traditional FE method) and the ML approach, highlighting the efficiency of ML in reducing simulation time. The presentation also addressed the uncertainties and safety aspects associated with neural networks in design and simulation. It emphasized the importance of training data, parametrization, and the challenge of achieving 100% safety. In conclusion, the presentation provided insights into the early applications of ML and AI in simulation, discussing various models and potential applications. It also pointed out the likelihood of these technologies being used in the near future for subsystems and components, with non-parametric methods currently under development. The results were encouraging but also raised new questions, necessitating further investigations in collaboration with Dassault Systèmes' clients and consideration of upcoming research.