Hypersonic CFD — real-gas validation against measured wind-tunnel data

Abstract

Hypersonic flow is one of the harder regimes in computational fluid dynamics. Real-gas chemistry, shock-layer non-equilibrium, surface ablation, and high-temperature thermodynamics all matter at the same time, and the conventional CFD response is per-case re-tuning of switchable subroutines. We characterise Aether ^[3] performance on two standard re-entry benchmarks across the Mach 10–20 range, demonstrate that a single model configuration handles real-gas chemistry without per-case kinetic-mechanism specification, and document where laminar-to-turbulent transition prediction remains the open problem. We achieve mean heat-flux error of 5.8% on the blunt-body benchmark and surface pressure error of 2.1% on the sphere-cone, with no case-specific re-tuning. Coupled ablation is supported through a dedicated solver tool-call; the platform's multiphysics coupling ^[4] handles the connection between ablation and shock-layer chemistry uniformly with other multiphysics workloads.

1. Why hypersonics is hard

Hypersonic flow is one of the harder regimes in computational fluid dynamics. The reasons compound. The shock layer is hot enough (often above 4,000 K) that air dissociates into a chemically-reacting mixture of N₂, O₂, NO, N, O and ionised species; the dissociation kinetics are finite-rate and the local equilibrium can lag the local thermodynamic state by milliseconds. Vibrational and translational temperatures are not equal — two-temperature ^[2] or full multi-temperature treatments are required for accurate prediction. Surface ablation removes wall mass and outgases ablation products that mix with the shock layer; the wall recession changes the boundary geometry within the simulation horizon.

Most CFD codes handle this regime through switchable subroutines and per-case tuning. The user picks a chemistry mechanism (5-species air, 11-species air, plus ablator species), a vibrational-relaxation model (Park, Millikan-White), a transport-property choice (Wilke's rule, Gupta-Yos), and a turbulence model. Each choice has knobs that have to be tuned per case to get reasonable agreement with measurements. The total complexity is daunting and the results are sensitive to choices that are not obvious from physics alone.

We wanted to see whether a model trained on hypersonic trajectories handled the regime natively — without the per-case configuration step. The answer is approximately yes, with documented exceptions.

2. Benchmarks

We use two standard re-entry-class benchmarks ^[1]. The first is a classic blunt-body shape — a hemisphere-cylinder — with extensive measured surface pressure and heat-flux data at multiple Mach numbers and angle-of-attack conditions. The second is a sphere-cone with re-entry-class conditions including wall-temperature variation. Both have public measurements spanning the Mach 10 to Mach 20 range.

We ran Aether at matched conditions and compared against the tunnel data without changing turbulence-model defaults or boundary treatments between cases. The default configuration is what customers get out of the box; tuning is supported but we explicitly do not tune for these benchmarks. The point of the benchmark is to test the default.

Result tabulation was blinded under the same protocol we use across our benchmark papers ^{[5, 6]}. The reviewer comparing predicted heat flux to measured did not know which tool had produced which prediction.

3. Headline numbers

Across the four Mach-condition × geometry combinations, Aether produces mean heat-flux error of 5.8% against measured surface heat flux at multiple stations on the blunt-body benchmark, and mean surface pressure error of 2.1% on the sphere-cone. Both numbers are competitive with published hypersonic-CFD results that used case-specific tuning; we achieve them with zero tuning.

Per-station behaviour: heat-flux error is tightest at the stagnation region (3.8% on the blunt-body) and grows toward the downstream stations (8.5% at the farthest measurement station from the stagnation point). The pattern is consistent with the dominant source of error being downstream of the shock layer rather than at it; we discuss the mechanism in Section 5.

4. Real-gas chemistry inside the model

The interesting qualitative result is that Aether handles the chemical-source terms — finite-rate dissociation of N₂ and O₂, formation and dissociation of NO, vibrational-translational relaxation in the Park-style two-temperature framework ^[2] — without an explicit kinetic-mechanism configuration step. The training corpus included enough hypersonic trajectories that the model learned the relevant relaxation behaviour as part of pretraining ^[3].

This is operationally important. Setting up a hypersonic case in a conventional CFD code routinely takes a senior engineer several days, much of it spent on chemistry-mechanism configuration and convergence-debugging the resulting non-linear system. Aether skips this step; a new geometry-and-conditions case can be specified in a few minutes and converged with default settings. The cost of this convenience is documented in Section 7 (limitations).

Customers with non-standard chemistry — ablator outgassing species, hypothetical atmospheric compositions, additive injection scenarios — can override the model's implicit chemistry with their own kinetic mechanism. The mechanism is passed as a tool-callable specification; the model's downstream prediction uses the override for chemical source terms while retaining its learned handling of shock structure and thermal non-equilibrium.

5. Where it still struggles — laminar-to-turbulent transition

Laminar-to-turbulent transition on hypersonic surfaces remains the open research question for our methodology and for the field. Cone-flare transition onset is predicted within roughly 12% of measured onset location in our benchmark cases; this is competitive with the published empirical correlations ^[1] but not yet at the level engineers need for design — typical design margins require 3–5% accuracy on transition onset.

The mechanism we currently rely on is hybrid. The model predicts the flow upstream and downstream of transition correctly; the transition location itself is bias-corrected with Mack-mode and second-mode empirical correlations called as tool functions when the customer's geometry is in the regime where transition matters. The hand-off is documented in the agent traces; the customer can see exactly where the empirical correction was applied.

We are extending pretraining with additional transition-event trajectories. Preview results are encouraging — 8% error on the same cone-flare benchmark in current 1.1 candidates — but the variance is high enough that we are not yet ready to remove the empirical-correlation fallback. Customers running transition-sensitive designs continue to use the hybrid approach.

6. Coupled ablation

Coupled surface ablation — where the wall recedes and outgases under heating — is supported as a tool-call into a dedicated ablation solver, with the wall material database open to customer configuration. The agent's coupling between ablation and shock-layer chemistry uses the same multiphysics-coupling primitives ^[4] that handle thermal-structural coupling on turbine blades and fluid-structure coupling on aerodynamic surfaces.

The hand-off is at the surface mass-flux level. The ablation solver computes the local mass loss per unit area; the CFD agent ingests it as a boundary condition and updates the shock-layer chemistry with the outgassing species. Recession is integrated in time; the boundary geometry evolves; the model accommodates the moving boundary without per-step reconfiguration.

Material databases are customer-configurable. The default set covers carbon-phenolic, silicone-impregnated reusable ceramic ablator, and several polymer-matrix composites. Customers running classified or proprietary materials supply their own database; the database is treated as customer data and is not shared across tenants.

7. Limitations

Three meaningful limitations. First, the published transition-onset accuracy (12% on cone-flare) is not yet at design-grade precision and requires empirical-correlation fallback. Second, the chemistry handling is implicit; customers with chemistry far outside the pretraining distribution (exotic gas mixtures, unusual ablator species) should run validation cases before treating the model's prediction as reliable. Third, the benchmark cases are at wind-tunnel scale; full-vehicle re-entry simulations introduce additional complications (rarefaction at the upper end of the trajectory, full ablation coupling, complex geometries) that we handle but have not benchmarked in this paper.

Reproducibility: the public benchmark cases ^[1] are open under the Research tier. Customer-specific cases involving proprietary geometries or chemistry are not shareable; we run them on customer hardware in customer-controlled environments and the results stay with the customer.

8. Implications for the workflow

For aerospace customers working on re-entry vehicles, hypersonic test vehicles, or atmospheric-capture systems, the operational headline is that one model handles the regime competently across conditions, and the same model handles the subsonic regime that the same vehicle traverses during the rest of its mission. The wins are not just accuracy — they are workflow.

Customers do not maintain a parallel hypersonic configuration of solvers, meshes and chemistry alongside their normal CFD workflow. The same project file ^[5] that handles subsonic external aerodynamics handles hypersonic re-entry on the same agent stack. Engineers who are competent in the subsonic flow can run hypersonic cases with the same tooling; specialist time is concentrated on the cases where it adds the most value.

9. Conclusion

On the standard hypersonic benchmarks ^[1], Aether achieves mean heat-flux error of 5.8% on the blunt-body and surface pressure error of 2.1% on the sphere-cone with zero case-specific tuning. Real-gas chemistry is handled implicitly by the trained model; coupled ablation works through the platform's multiphysics primitives ^[4]. Laminar-to-turbulent transition prediction remains the dominant open problem and continues to rely on empirical-correlation fallback. The result is operationally useful for aerospace customers running vehicles that traverse the hypersonic regime — one model configuration handles the regime, the same configuration handles the surrounding flight envelope, and the workflow does not bifurcate.