Point-Deeponet: Predicting nonlinear fields on non-Parametric geometries under variable load conditions.

Area of Science:

• Computational Engineering
• Artificial Intelligence
• Structural Mechanics

Background:

• Nonlinear structural analyses rely on computationally intensive finite element simulations.
• Current deep learning surrogates face challenges with complex 3D geometries and variable loads.

Purpose of the Study:

• To develop a high-fidelity, operator-learning-based surrogate model for nonlinear structural analysis.
• To enable rapid prediction of physical response fields for complex, non-parametric engineering problems.

Main Methods:

• Integration of PointNet for geometric representation learning from point clouds into the DeepONet framework.
• Synergistic fusion of geometric embeddings and load conditions to predict 3D displacement and von Mises stress fields.
• Training on a large-scale dataset and validation on unseen load conditions.

Main Results:

• Point-DeepONet achieved high accuracy, with R² values up to 0.987 for displacement and 0.923 for von Mises stress.
• The model demonstrated excellent generalization to randomly oriented, unseen load directions.
• Predictions were achieved in seconds, approximately 400 times faster than traditional finite element analyses.

Conclusions:

• Point-DeepONet offers a significant advancement in computational efficiency and accuracy for nonlinear structural analysis.
• The model's ability to handle complex geometries and variable loads opens new possibilities for design optimization and real-time control.
• This approach shows great potential for accelerating complex engineering workflows.