Learning 3D Signed Distance Functions for Automotive Performance Prediction
Please login to view abstract download link
When predicting performance metrics for complex 3D geometries, surrogate models must first capture each shape’s essential features in a compact form. To meet this need, we propose a two-stage neural architecture in which an autoencoder and a latent-parameter network share a single decoder. Trained on signed distance functions (SDFs), the network converts intricate geometries into smooth distance grids, enabling precise surface localization and robust handling of challenging shapes. In the first stage, the autoencoder compresses dense SDF samples into a low-dimensional latent vector; the shared decoder then reconstructs the full 3D SDF field from this code. Running in parallel, the latent-parameter network predicts the same latent vector directly from simple geometric descriptors ensuring that every inferred code corresponds to a valid SDF [1]. We investigate convolutional and fully connected variants of the architecture and introduce a custom activation function that sharpens surface reconstructions.Numerical tests on spheres, ellipsoids, and simplified automotive panels demonstrate the method’s promise as a foundation for handling truly complex, non-parameterized geometries and for incorporating physics-informed constraints in future work [2, 3].