SimNet: A Neural Framework for Physics Simulations

Simulations are pervasive in every domain of science and engineering. In comparison with the

traditional solvers, physics-driven neural network solvers can not only do parametrized

simulations in a single run, but also address problems not solvable using traditional solvers.

We present SimNet, an end-to-end physics based neural framework offering several novel

features, including signed distance functions for loss weighting, integral continuity planes for

flow simulation, advanced neural network architectures that are optimized for high-

performance GPU computing, point cloud generation for real world geometries using

constructive geometry and STL modules, and parameterization of both geometry and physics.

Additionally, for the first time to our knowledge, we solve high Reynolds number turbulent

flow in industrial applications without using any training data.

Due to several performance enhancements for both single and multiple GPU/nodes, SimNet

offers fast turnaround time by enabling parameterized system representation that solves for

multiple configurations simultaneously, and offers scalable performance for multi-GPU/multi-

node implementation with accelerated linear algebra as well as FP32, FP64 and TF32

computations.

Through several numerical examples, we show that SimNet addresses use cases across four

major areas in computational science and engineering, that are, inverse and data assimilation

problems, real time simulations, improved physics and predictions, and digital design and

manufacturing. The examples include data assimilation in patient-specific intracranial

aneurysm, digital twin of damage accumulation in an aircraft fuselage, automated detection

of signal integrity hotspots, full waveform inversion, and design space exploration of

NVSwitch heat sinks for Nvidia RTX & DGX servers.

We also present extensive accuracy and efficiency comparisons of SimNet with open source code, documentations and

examples are available at: https://developer.nvidia.com/simnet