Multimodal Pre-Training Enables Time-Dependent Physics Inference from B-Rep
Please login to view abstract download link
Estimating 3D physics fields in structural mechanics is critical during early design phase but often computationally expensive, especially when using high-fidelity Finite Element Analysis. Although parametric surrogate modelling methods such as the one proposed in [1] provide faster inference alternatives, their applications are limited to cases where geometric design parameters are predefined and consistent across training and testing samples. In this research, we introduce a novel self-supervised learning approach that maps geometry data to latent representations by training on a pseudo-task of converting Boundary Representation (B-Rep) to its corresponding Signed Distance Field [2,3,4]. The multimodal pre-training task eliminates the need for parameterization, as it operates directly and depends only on CADnative B-Rep geometry. The bottleneck latent vector produced by the encoder effectively captures the varying geometric features, serving as latent parameters for multimodal downstream tasks involving structural property and 3D physics field estimations. We demonstrate the efficacy of this method on thin-walled crash column simulations in explicit dynamics, accurately predicting time-dependent reaction forces and nodal displacement fields given new geometric designs in B-Reps, showcasing the potential of multimodal pre-training for time-dependent physics inference in engineering design.
