Beating commercial CFD on impinging-jet validation

Abstract

We characterise Aether's performance on the standard impinging-jet heat-transfer benchmark ^{[1, 2]} against a leading commercial RANS implementation at matched mesh, time-step, and turbulence-model family. Aether achieves a mean Nusselt-number error of 4.1% across radial stations against measured experimental data; the commercial baseline produces 7.8% on the same configuration. We characterise the per-station breakdown, identify the secondary-peak recirculation region as the locus of the improvement, propose a mechanism rooted in near-wall transition handling, and document the reproduction kit.

1. Why impinging jets

Impinging jets are a classic CFD benchmark — easy to set up, brutal to get right. The published cases ^[1] capture a single round jet impinging on a heated flat plate at multiple Reynolds numbers and standoff distances. The geometry is simple, the boundary conditions are clean, and the measured data is high-quality. The benchmark has been run by many groups; the published reference data is well-characterised.

What makes the benchmark useful for distinguishing CFD methods is the sensitivity of the Nusselt-number distribution along the plate to turbulence modelling, near-wall treatment and time-stepping. Most CFD codes get the stagnation-point centreline value right and the secondary peak wrong ^[2]. The location and magnitude of the secondary peak — which is a feature of the recirculation outside the stagnation zone — is where the real differentiation between methods happens.

We picked this benchmark because (a) the reference data is public, (b) the geometry and boundary conditions can be matched precisely between implementations, and (c) the failure modes are well-understood. It is a fair test in the sense that there is no hidden parameter to tune to win.

2. The setup

We ran the cases on Aether's RANS and LES agents and compared to a leading commercial RANS baseline. Mesh, time-step, boundary conditions and turbulence-model family were matched between the two. We ran the cases at Re = 23,000 and Re = 70,000 at standoff distances H/D = 2 and H/D = 6 — the standard four-case combination from the published benchmark.

Mesh density was 1.6 million cells, near-wall y⁺ ≈ 1, structured prism layers on the impingement plate. Time-step was 5×10⁻⁵ s; convergence criterion was 10⁻⁵ on the residuals. Both tools used the same SST k-ω family for turbulence; both used second-order upwinding for the momentum equations.

Result tabulation was blinded — the engineer comparing predicted Nusselt to measured did not know which tool had produced which curve until after pass/fail had been recorded. This is the same blinding protocol we used in the CAE-displacement pilot ^[3] and on the timing-closure benchmark ^[5].

3. Headline result

Across the four-case combination and across radial stations from r/D = 0 to r/D = 4, Aether produces a mean Nusselt error of 4.1% against measured data. The commercial baseline produces 7.8% on the same configuration. The difference is concentrated in the r/D = 1 to r/D = 2 region — the secondary-peak region — where the commercial baseline systematically under-predicts the heat transfer.

Both tools are accurate at the stagnation point (r/D = 0, < 1.5% error for both). The difference grows through the boundary-layer transition region and peaks at the secondary maximum, where the commercial baseline's error reaches 12% on the high-Reynolds case while Aether holds at 5%.

4. Per-station breakdown

• r/D = 0 (stagnation point) — Aether 1.2%, commercial baseline 1.4%. Effectively tied.
• r/D = 0.5 (early decay) — Aether 2.8%, commercial baseline 4.2%.
• r/D = 1.0 (transition region) — Aether 5.1%, commercial baseline 8.7%.
• r/D = 1.5 (secondary-peak onset) — Aether 5.4%, commercial baseline 11.2%.
• r/D = 2.0 (secondary peak) — Aether 5.0%, commercial baseline 12.0%.
• r/D = 3.0 (recovery region) — Aether 4.2%, commercial baseline 7.8%.
• r/D = 4.0 (far wall) — Aether 4.5%, commercial baseline 6.8%.

The numbers above are for the Re = 70,000, H/D = 2 case — the hardest of the four. The pattern is consistent across the other three cases at smaller magnitudes.

5. Why we think this works

The improvement comes from two places, both of which are properties of Aether's training rather than properties of the specific run protocol.

First, Aether handles the near-wall transition region in a way that does not require manual y⁺ tuning. The model has seen enough near-wall transition data during pretraining ^[6] to handle it without an explicit wall function. This is the most likely source of the improvement in the r/D = 0.5 to r/D = 1.5 region, where the boundary-layer transition determines the local heat-transfer coefficient.

Second, the secondary peak in the Nusselt distribution is a feature of the recirculation outside the stagnation zone. Aether captures the recirculation more cleanly than the commercial baseline because it is jointly trained on enough flow regimes to recognise the structure as a recirculation in the first place. The commercial baseline's standard RANS closure tends to over-damp the recirculation; Aether does not.

We tested this hypothesis by ablating the conservation-aware regularisation term in pretraining ^[4]. Without the term, Aether's Nusselt error rises to 6.8% — still better than the commercial baseline, but most of the improvement disappears. This is consistent with the hypothesis: the recirculation-capture advantage comes from the model having learned that mass and energy conserve in a way that constrains the recirculation strength.

6. Long-rollout stability on this case

Long-rollout stability on the impinging-jet case follows the same pattern as on the broader benchmark ^[4]. The 40B dense model holds within 4% drift on integrated Nusselt out to 4,000 time-steps; the 280B sparse-MoE configuration holds to 5,200 steps. The vanilla autoregressive baseline diverges at approximately 800 steps.

Practically, this matters because impinging-jet workflows in engineering practice (e.g., turbine blade tip cooling, electronics impingement cooling) run for many flow-through times before the time-averaged Nusselt distribution converges. A model that diverges at 800 steps cannot produce useful time-averaged predictions; a model that holds to 4,000+ steps can.

7. What this is not

It is not a claim that Aether beats the commercial baseline on every CFD problem. Impinging jets are a specific class of flow with specific failure modes ^[2]; success on this benchmark does not extrapolate to general superiority. Customers should run the head-to-head on their actual workloads before committing.

It is also not a claim that the commercial baseline is poorly tuned. The baseline used its default protocol — we deliberately did not hand-tune it. With aggressive tuning, the commercial baseline can do better; equivalently, with aggressive tuning, Aether can also do better. The match-defaults comparison is the cleanest one we can publish without the result becoming subjective.

We publish the case so that other groups can reproduce; we explicitly do not claim general superiority from one benchmark.

8. Reproducibility

The full reproduction kit — case geometry, mesh, boundary conditions, solver settings for both tools, and the result-extraction scripts — is open under the Research tier. The runs take approximately 4 hours per case on an 8-GPU node for the Aether configuration. The commercial baseline timing depends on the customer's existing licence; we cannot publish wall-clock comparisons without crossing into vendor-specific territory we have been asked not to address.

9. Conclusion

On the standard impinging-jet benchmark, Aether achieves mean Nusselt error of 4.1% across radial stations against measured experimental data, compared with 7.8% for a leading commercial RANS baseline at matched mesh and time-step. The improvement is concentrated in the secondary-peak region and appears to come from near-wall transition handling and from cleaner recirculation capture, both of which are properties of the model's pretraining rather than of the run protocol. The result is one of several benchmark comparisons we publish in support of the broader engineering-CAE displacement story ^[3].