Advancing CFD with AI: Surrogate Modeling Approaches in the NASA Software Ecosystem

Insights from Rescale’s Presentation in NASA’s Advanced Modeling & Simulation Seminar Series

High-fidelity simulation is central to aerospace innovation, but complex computational fluid dynamics (CFD) and finite element analysis (FEA) solvers can often take days to complete. These long runtimes limit the number of design variations that can be explored, slowing down the pace of discovery. AI-powered surrogate models, built from high-fidelity simulation data, offer a breakthrough: rapid simulation predictions at lightning speed–up to 1,000X faster–part of a growing trend of applying “AI physics” methods to open new frontiers in design exploration and innovation.

In a recent seminar of NASA’s Advanced Modeling and Simulation (AMS) Seminar Series, Viktor Rozsa, Solutions Architect at Rescale, showed how Rescale customers like Boom Supersonic are combining AI physics frameworks with NASA simulation software to dramatically transform digital engineering workflows. Rescale supports NASA-developed software used by public and private sector companies in the US, making it easy for engineers to deploy advanced simulation techniques, analyze their data, and develop new AI methods like the ones discussed in this webinar. 

AMS Seminar Series: Where NASA’s Modeling & Simulation Community Connects

Hosted by the Computational Aerosciences Branch at the NASA Advanced Supercomputing (NAS) facility, the AMS series features weekly talks from experts across government, academia, and industry. Each session highlights advances in modeling and simulation that can have broad impact across engineering and scientific disciplines. Since its launch in 2011, the seminar series has centered on aerospace engineering and scientific research, often highlighting NASA’s software ecosystem for advanced aerospace modeling, and tools like FUN3D, OVERFLOW, Cart3D, and CAPE provide the high-fidelity computational foundation for cutting-edge research and design. Speakers typically include leading researchers, engineers, and software developers presenting on topics ranging from mesh adaptation and turbulence modeling to AI-augmented CFD and high performance computing workflows.

Breaking Bottlenecks in Design Space Exploration

Designing aerospace systems means weighing millions of possible combinations of variables like lift, drag, and structural loads to find the best-performing design. This wide range of possibilities and tradeoffs is often referred to as the “design tradespace.” Running each scenario with traditional simulation methods can consume significant high performance computing resources, tying up hardware for days, so only a fraction of potential designs can realistically be tested. This tradespace bottleneck limits the ability to explore new shapes, configurations, and operating conditions. AI-powered surrogates overcome this by learning from simulation data and predicting results in near real time, allowing engineers to explore far more of the tradespace within the same compute budget.

Graph Neural Nets: A Foundation for Mesh-Based Surrogates

Graph neural networks (GNNs) represent physical systems as graphs, where nodes correspond to mesh points and edges encode spatial relationships, often carrying local properties such as pressure or velocity. By iteratively passing messages between nodes, GNNs capture spatial relationships critical to physics-based simulations. MeshGraphNets (MGNs) apply this approach by encoding meshes into a latent space, processing them through multiple message-passing steps, and decoding them back into predicted physical fields, achieving up to 100× speedups over traditional solvers while keeping local errors around 10% in some cases. In contrast, DoMINO (Decomposable Multi-scale Iterative Neural Operator) is a mesh-free neural operator within NVIDIA PhysicsNemo that learns patterns directly from point clouds across multiple scales, without relying on fixed meshes, enabling fast training and adaptation to new geometries without retraining.

Broad Applications Across Aerospace and Beyond

The advantages of these approaches extend beyond any single aircraft study. Surrogate models have potential in aerodynamic optimization, thermal management, structural analysis, and multiphysics design challenges. They are particularly powerful in workflows where rapid iteration is critical–such as exploring unconventional wing configurations, refining heat shield designs for re-entry vehicles, or optimizing fuselage structures for both performance and safety.

In one demonstration shown in the Rescale Seminar, steady-state CFD simulations for a Boeing 787 were used to train an MGN model to predict aerodynamic coefficients and local flow variables. Once trained, the surrogate could deliver results in near real time, enabling up to 50X more design iterations without increasing compute costs. DoMINO has been similarly applied to structural and aerodynamic optimization in supersonic aircraft, revealing opportunities to reduce mass while preserving safety margins.

Strategies for Effective Surrogates

Successful surrogate models start with high-quality ground truth simulations that span the range of geometries, conditions, and scenarios the model will encounter. Clean, consistent data is critical–issues like naming errors, missing tags, and unstructured logs can undermine performance. Rescale’s data lakehouse addresses this by standardizing metadata capture so key parameters are always recorded and linked to results. With a solid data foundation, careful hyperparameter tuning–such as adjusting settings like learning rate, layer size, and batch size–helps balance accuracy, speed, and generalization. NASA case studies show that pairing robust simulation data with well-optimized hyperparameters enables surrogates to deliver rapid, reliable predictions that support confident engineering decisions.

Rescale’s Advantages for NASA Software Users

For organizations using NASA flow solvers, Rescale’s invocation of NASA CAPE (Computational Aerosciences Productivity & Execution) integrates directly into existing workflows for FUN3D, OVERFLOW, or Cart3D. After each simulation run, Rescale automatically extracts solver inputs and outputs, such as Mach, angle of attack, geometry identifiers, lift, drag, and residual histories, and formats them into a standardized metadata record. This record is then ingested into Rescale’s data lakehouse to ensure full data traceability. This process operates transparently without disrupting current CAE pipelines. By automating dataset assembly, it eliminates the need for manual file wrangling and accelerates the creation of training sets for Rescale AI, enabling surrogate model development directly from operational CFD runs.

Deploying and Scaling Surrogates Across Industries

If you’re an engineer working in aerospace, automotive, energy, manufacturing, or any field where complex simulation is a bottleneck, surrogate models are worth serious consideration. Whether you’re trying to explore more design parameters, cut simulation time from weeks to hours, or make faster, data-driven decisions, AI-powered surrogates can give you a major edge. Mesh-based methods like MGNs can deliver high accuracy for consistent workflows, while mesh-free approaches like DoMINO offer more flexibility when designs or conditions change. The key is pairing the right approach with quality simulation data, smart workflow integration, and tuned hyperparameters so you get both speed and reliability.

With these building blocks in place, AI-augmented simulation can transform your design process and unlock faster iterations, deeper tradespace exploration, and better engineering outcomes. Watch the full AMS seminar recording to dive deeper into these methods and learn from demonstrations and performance comparisons you can apply to your own work. Watch the seminar and contact Rescale to learn how to build high-performance AI surrogates from your CFD data and start accelerating your simulation workflows today.