CHAPERONE AI
Have a different geometry, Reynolds band, or 3D setup?
The public checkpoint proves the method. The product is adapting a physics foundation model to your regime — without erasing what it already knows.
Normalized field error on held-out vortex shedding
Holdout cases above the training Mach band
Original Poseidon families preserved or improved
Median field correlation after autoregressive rollout
High-order CFD is the gold standard — and too slow for early design loops. Foundation physics models are fast, but they only know the flows they were pretrained on. Chaperone-Flow-1.0 is the missing piece: we teach a proven foundation model your missing regimes, keep what it already does well, and ship a drop-in surrogate on the same 4-channel state the base model already uses.
Incompressible flow past a circular cylinder with periodic vortex shedding. Trained across Re ≈ 62–179, tested just beyond that band — the kind of screening loop aero and energy teams run every day.
Kelvin–Helmholtz double shear layers across Mach ≈ 0.20–0.32, with holdout at higher Mach. Useful for mixing, shear-driven instability, and early what-if exploration before a full solver run.
No new architecture. No extra input channels. Geometry is encoded in the existing density field. If you already evaluate Poseidon-B, this checkpoint drops in.
Starting from Poseidon-B (~158M parameters, ETH Zurich CAMLab), we generated solver-grade data for two missing industrial regimes and adapted the full model in one joint run — designed so existing capabilities do not collapse.
We start from Poseidon-B, a published physics foundation model, instead of training a new operator network for every geometry.
Cylinder and mixing-layer trajectories generated with PyFR, then resampled to a common 128×128 state the model can consume at inference time.
Both new flows are taught together, in a fixed training budget, so you get one checkpoint instead of a zoo of one-off surrogates.
Original Poseidon families stay in the mix during training. After adaptation they are near — or better than — the untouched base model.
Error is reported as normalized RMSE against the channel range — a 1.8% score is 1.8% of the observable field span, not a raw residual. New regimes beat a no-change baseline. Existing regimes were not sacrificed to get there.
| Capability | What we measured | Chaperone-Flow-1.0 | Why it matters |
|---|---|---|---|
| New industrial regimes | Cylinder wake, lead-2 field error | 1.8% NRMSE | Usable wake-scale prediction on unseen Re |
| Kelvin–Helmholtz holdout error | 3.8% NRMSE | Works slightly outside the training Mach band | |
| Existing physics retained | Families kept in the training mix | 0.63× / 0.69× | Known regimes got more accurate, not worse |
| Family never shown during adaptation | 1.16× | Still within ~16% of the original model | |
| Tracked original families | 0.63× – 1.16× | One model, not a trade-off | |
| Forecast stability | 20-step autoregressive rollout | NRMSE 0.7% → 1.7% | Error stays bounded over multi-step use |
| Field correlation at step 20 | median 0.985 | The predicted field still looks like the flow |
Means over 3 seeds. NRMSE is RMSE divided by each channel’s value range. Cylinder checkpoint selected on 30 held-out trajectories. Retention is error relative to the unchanged Poseidon-B checkpoint (< 1.0 is better).
Adaptation only works if the data is solver-grade. We generated a dedicated PyFR dataset for the two missing regimes, with a holdout split that tests extrapolation — not memorization.
Teams building their own surrogates can start from the same tensors. For proprietary geometries, Reynolds bands, or 3D setups, we generate the data and adapt the model under a commercial license.
CAE and digital-twin teams in aero, auto, and energy who need fast screening of wakes and mixing — and a path to adapt a foundation model to the next geometry, not a chatbot.
Early design loops, parameter sweeps, and multi-step rollouts inside the validated window. Same 4-channel (ρ, u, v, p) interface as Poseidon-B, on a 128×128 field.
This is a 2D, wake-scale surrogate — not a certification CFD replacement. Fine scales are smoothed; cylinder Re is kept in the 2D-valid band; validated rollout is about four shedding periods.
Flow-1.0's lineage runs inside NumericalAI, our GPU simulation platform. The SRS platform runs the same high-order PyFR engine that generated this model's training data; CMF covers multiphase compressible CFD.
Before a run launches on SRS or CMF, our models validate and diagnose the input scripts and geometries — malformed case files, inconsistent parameters, broken geometry references — caught before GPU time is spent.
When a simulation crashes or diverges, our models drive the AI diagnosis pipeline — tracing the failure back to the setup and proposing the fix, so a dead run comes back as a corrected case instead of a support ticket.
SRS ships ready-to-run PyFR launch templates across 7 physics categories — zero license fees, pay only for compute.
See it working on numericalai.netThe public checkpoint proves the method. The product is adapting a physics foundation model to your regime — without erasing what it already knows.
We use cookies
Some cookies are needed for this site to work. We would also like to set analytics cookies to understand how the site is used — but only if you agree. Privacy policy
Strictly necessary
Session and security cookies that keep forms and logins working. These cannot be switched off.
Analytics
Google Analytics, to count visits and see which pages are read. Off unless you turn it on.