Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

Accelerating CFD with Modular Surrogates: Insights from MeshGraphNets and UPT Architectures

AMS Seminar Series

Speaker: Viktor Rozsa, Rescale

Presentation

Seminar Slide Deck (PDF-11.2MB)

Abstract

Modeling and Simulation (M&S) remains a cornerstone of high-fidelity aerospace design, yet trade-space studies routinely require millions of CFD evaluations, driving a need for surrogate models that balance accuracy and throughput. Surrogate models offer a path forward, yet most production tools rely on rigid grid or particle structures and aren’t packaged for seamless integration into everyday CAE workflows.

This work describes a modular surrogate-modeling framework built on Nvidia PhysicsNemo and other open-source frameworks. Our primary approach leverages MeshGraphNets (MGNs), a specialized graph-neural-network architecture that encodes arbitrary CFD/FEA meshes into latent vectors via an MLP-based encoder, propagates information through iterative message passing in a processor core, and decodes back to physical field predictions. MGNs naturally respect mesh connectivity, generalize across resolutions, and deliver orders-of-magnitude speedups over traditional solvers while retaining high accuracy.

Complementing MGNs, we integrate Universal Physics Transformers (UPTs) as modular components to transcend grid- or particle-centric limitations. UPTs employ a hierarchical encoder combining message passing, transformer blocks, and a perceiver-style bottleneck to fuse diverse spatio-temporal inputs into a fixed latent representation. A transformer-based approximator performs rapid “latent rollouts,” and a perceiver-inspired decoder reconstructs the state at any space–time query. By using inverse reconstruction losses and feature modulation for conditions such as timestep and boundary parameters, UPTs offer a unified, scalable surrogate modeling paradigm adaptable across fluid, structural, and multiphysics simulations.

This seminar will present this work from MGN and UPT models on representative CFD cases, highlighting trade-offs in accuracy, inference speed, generalization to new geometries, and integration into standard CAE pipelines. By modularizing advanced AI physics models into CAE-ready components, Rescale empowers engineers to accelerate design exploration, reduce computational cost, and maintain high fidelity across a broad range of simulation scenarios.

Biography

Photo of Viktor Rozsa
Viktor Rozsa is a Solutions Architect at Rescale, where he supports the deployment of high-performance computing workflows across disciplines ranging from computational chemistry to aerospace engineering. He holds a Ph.D. in Molecular Engineering from the University of Chicago, where his research focused on first-principles molecular dynamics simulations of materials under extreme conditions, with applications to water technologies, renewable energy materials, and planetary science. As a Department of Energy Stewardship Science Graduate Fellow, he also conducted research at Lawrence Livermore National Laboratory, studying confined aqueous solutions relevant to energy technologies. Prior to joining Rescale, he worked in the manufacturing data science sector at Sight Machine, applying advanced analytics to industrial operations. His current interests include cloud-native scientific computing, ML-enhanced simulation, and the intersection of data-driven and physics-based modeling.