NNWarp: Neural Network-Based Nonlinear Deformation.

Area of Science:

• Computational physics
• Machine learning
• Scientific computing

Background:

• Deformable simulation requires reconstructing complex force-displacement relations.
• Directly applying neural networks (NNs) to general deformable simulation is challenging due to variable, high-dimensional input data.
• Existing methods struggle with the parametrization-unfriendly nature of simulation inputs.

Purpose of the Study:

• To introduce NNWarp, a reusable and efficient neural network-based framework for nonlinear deformable simulation.
• To address the limitations of directly using NNs for complex force-displacement reconstruction in simulations.
• To enable real-time simulation of large 3D models with varying geometries.

Main Methods:

• NNWarp partially restores the force-displacement relation by "warping" nodal displacements from a linear elasticity model.
• A compact feature vector (geodesic, potential, digression) is used to sort training data for per-node linear and nonlinear displacements.
• Deformation substructuring assists the NN in handling diverse 3D model shapes and tessellations.

Main Results:

• NNWarp demonstrates robustness across different model shapes and tessellations.
• A single NN training session effectively handles a wide range of 3D models through deformation substructuring.
• The framework achieves real-time simulation speeds for large models due to efficient matrix solves.

Conclusions:

• NNWarp provides an efficient and robust solution for nonlinear deformable simulations using neural networks.
• The approach successfully overcomes the challenges of high-dimensional and variable input data in physics-based simulations.
• NNWarp enables real-time performance, making it suitable for complex 3D modeling and simulation tasks.