核心概念
Geom-DeepONet combines SDF and SIREN to predict field solutions on variable 3D shapes efficiently.
要約
The article introduces Geom-DeepONet, a novel deep operator network variant that encodes parameterized 3D geometries and predicts full-field solutions. It leverages SDF and SIREN for spatial geometric awareness, outperforming PointNet and vanilla DeepONet in accuracy and efficiency. The model's generalizability to dissimilar shapes is highlighted, showcasing superior performance in design similarity-based data splits.
Introduction:
Modern engineering relies on computational models.
Neural networks offer efficient design insights.
Challenges with varying 3D geometries in simulations.
Methods:
Comparison of PointNet, vanilla DeepONet, and Geom-DeepONet.
Results and Discussion:
Geom-DeepONet excels in accuracy and efficiency.
Generalizability tested with similarity-based data splits.
Extension to vector predictions:
Cuboid dataset used for von Mises stress and displacement vector predictions.
Effect of Geom-DeepONet model size:
Investigating the impact of increasing trainable parameters.
統計
結果は、我々のアーキテクチャがPointNetおよびバニラDeepONetを凌駕し、精度と効率性で優れていることを示しています。
引用
"Nevertheless, few available NNs can handle solution prediction on varying 3D shapes."
"Our architecture trains fast with a small memory footprint and yields the most accurate results among the three."