[Paper Review] NVIDIA SimNet^{TM}: an AI-accelerated multi-physics simulation framework

Oliver Hennigh, Susheela Narasimhan+8 more|arXiv (Cornell University)|Dec 14, 2020

Model Reduction and Neural Networks34 references31 citations

TL;DR

SimNet is an AI-driven framework that solves forward, inverse, and data-assimilation PDE problems with parameterized geometries, enabling multi-physics simulations on GPUs.

ABSTRACT

We present SimNet, an AI-driven multi-physics simulation framework, to accelerate simulations across a wide range of disciplines in science and engineering. Compared to traditional numerical solvers, SimNet addresses a wide range of use cases - coupled forward simulations without any training data, inverse and data assimilation problems. SimNet offers fast turnaround time by enabling parameterized system representation that solves for multiple configurations simultaneously, as opposed to the traditional solvers that solve for one configuration at a time. SimNet is integrated with parameterized constructive solid geometry as well as STL modules to generate point clouds. Furthermore, it is customizable with APIs that enable user extensions to geometry, physics and network architecture. It has advanced network architectures that are optimized for high-performance GPU computing, and offers scalable performance for multi-GPU and multi-Node implementation with accelerated linear algebra as well as FP32, FP64 and TF32 computations. In this paper we review the neural network solver methodology, the SimNet architecture, and the various features that are needed for effective solution of the PDEs. We present real-world use cases that range from challenging forward multi-physics simulations with turbulence and complex 3D geometries, to industrial design optimization and inverse problems that are not addressed efficiently by the traditional solvers. Extensive comparisons of SimNet results with open source and commercial solvers show good correlation.

Motivation & Objective

• Motivate the need for fast, flexible PDE solvers that handle complex 3D geometries and multi-physics coupling.
• Introduce an AI-driven framework that can perform forward, inverse, and data-assimilation problems without relying on training data.
• Present architecture, loss weighting, and geometry/PDE modules to tackle industrial-scale problems.
• Demonstrate real-world use cases including turbulent flow, complex geometries, design optimization, and inverse problems.

Proposed method

• Formulate neural network solvers that approximate PDE solutions via a differentiable network trained with physics-based residuals.
• Use a loss function that sums PDE residuals and constraint residuals weighted by spatially varying terms, including Signed Distance Function (SDF) weighting.
• Incorporate automatic differentiation to compute PDE derivatives for residuals and use stochastic gradient-based optimization (e.g., Adam).
• Integrate geometry modules (CSG/TG) and point-cloud generation to handle complex 3D geometries; enable parameterized design spaces.
• Apply advanced architectures (Fourier feature networks, modified Fourier feature networks, SiReNs) to mitigate spectral bias and improve convergence.
• Support PDE modules (Navier–Stokes, diffusion, advection-diffusion, wave, elasticity) with symbolic equations (SymPy) for extensibility.

Experimental results

Research questions

• RQ1 Can SimNet accurately solve high-Reynolds-number, turbulent, multi-physics problems on complex 3D geometries without training data?
• RQ2 How do geometry parameterization and loss weighting (including SDF-based weighting) affect convergence and accuracy?
• RQ3 Can SimNet perform design optimization and inverse problems more efficiently than traditional solvers for multi-parameter geometries?
• RQ4 What is the impact of advanced network architectures on training convergence and solution accuracy for industrial-scale PDEs?
• RQ5 How do forward solutions compare to OpenFOAM and commercial solvers across turbulence, geometry, and inversion tasks?

Key findings

• SimNet achieves good correlation with OpenFOAM and commercial solvers across cases involving turbulence and complex geometries.
• Loss weighting with SDF and advanced architectures improve convergence and accuracy, outperforming standard fully connected models in several tests.
• In a multi-physics design optimization, SimNet parameterizes geometry to evaluate many designs in a single trained model, accelerating the process dramatically.
• For a turbulent conjugate heat transfer problem, SimNet variants (Fourier-based) show favorable pressure drop and peak temperature results compared to some baselines.
• The NVSwitch heat sink optimization demonstrates orders-of-magnitude speedups (up to ~135k×) over OpenFOAM and commercial solvers due to parameterized neural networks.