A Physics-driven Neural Networks-based Simulation System (PhyNNeSS) for multimodal interactive virtual environments...

Area of Science:

• Computational mechanics
• Real-time simulation
• Machine learning in physics

Background:

• Real-time graphics require 30 Hz updates, but haptics necessitate ~1 kHz.
• Simulating nonlinear deformable objects at high rates with complex material properties is challenging.
• Existing solutions lack generality for arbitrary nonlinearities.

Purpose of the Study:

• To introduce PhyNNeSS, a Physics-driven Neural Networks-based Simulation System.
• To overcome limitations in real-time simulation of nonlinear deformable objects.
• To enable high-fidelity haptic feedback in interactive systems.

Main Methods:

• Off-line pre-computation generates a database from finite element models.
• Data is condensed into Radial Basis Function Network (RBFN) coefficients.
• Neural networks reconstruct deformation fields and interaction forces in real-time.

Main Results:

• Realistic simulations demonstrated for interactive surgical simulation with force feedback.
• Developed models for a deformable human stomach and a Penrose-drain (FLS).
• Scalability and accuracy control shown via error analysis based on neuron count.

Conclusions:

• PhyNNeSS provides a unique system for real-time simulation of nonlinear deformable objects.
• Physics-based pre-computation enables neural network training for real-time use.
• The system is scalable and accurate, integrated into SoFMIS for general applications.