Neural fields paper for Orion reentry CFD

When a spacecraft comes home, it doesn’t just hit air. At Orion reentry speeds, the air piles up into a shock wave and heats to extremes, so engineers have to predict pressure, temperature, and heat on every part of the capsule before anyone flies it. (arxiv.org, nasa.gov) The standard tool for that job is computational fluid dynamics, which means chopping the air around the vehicle into a huge 3-D grid and solving physics in each tiny cell. NASA notes that mesh generation and adaptation are often the bottleneck in hypersonic runs, especially when strong shocks have to line up cleanly with the grid. (nasa.gov) A neural field tries a different trick. Instead of storing answers on a fixed grid, it learns a continuous function that can answer a question like “what are the pressure and temperature at this exact point in space” the way a map app gives you a location on demand. (nature.com) That idea works nicely for smoother aircraft flows, but hypersonic flow is full of cliffs instead of hills. Shock waves create abrupt jumps, and ordinary neural models tend to blur those jumps the way a low-resolution photo smears a sharp edge. (arxiv.org, nature.com) The new paper, posted on March 21, 2026, builds a 3-D neural field specifically for the Orion reentry capsule. The model takes spatial coordinates plus angle of attack and predicts pressure, temperature, and velocity components around the vehicle. (arxiv.org) The authors add Fourier positional features, which is a way of feeding the network finely spaced wave patterns so it can represent sudden changes instead of only smooth ones. They also force in wall rules from fluid physics, including no-slip and isothermal wall conditions, so the model is not free to invent impossible behavior right at the capsule surface. (arxiv.org) They compare that setup with other surrogate approaches, including graph neural networks, and report better performance on the steep gradients that dominate hypersonic flow. The paper describes the result as a continuous aerothermodynamic surrogate rather than a replacement for every high-fidelity solver. (arxiv.org) That distinction matters because Orion entry design is driven by narrow margins. NASA documents show the capsule’s hypersonic lift-to-drag ratio sits around 0.25 to 0.27, and entry constraints include landing accuracy, maximum heat rate, total heat load, acceleration, and even whether the vehicle stays inside an allowed entry corridor. (nasa.gov) A faster surrogate is useful when engineers want to sweep through many angles of attack or flight conditions without paying for a full giant simulation each time. The Orion paper says its model supports rapid exploration of operating conditions under realistic flight profiles, which is exactly the kind of repetitive trade study that makes conventional hypersonic computation expensive. (arxiv.org) The catch is in the paper’s own scope. It is aimed at steady hypersonic flow around Orion, not the full mess of every reentry regime, and NASA’s hypersonic workflows still rely on specialized solvers like DPLR and US3D because heat-flux prediction near strong shocks is hard enough that even mesh choice can change the answer. (arxiv.org, nasa.gov) So this is less “artificial intelligence replaces spacecraft aerodynamics” than “artificial intelligence learns a fast stand-in for a narrow but costly slice of it.” If the method holds up beyond Orion, it points to a future where engineers use neural fields for rapid screening and save the most expensive physics runs for the cases that really need them. (arxiv.org, nature.com)