CHAPERONE AI
Each model has a nested data page. The next solver drops onto this grid — not into the top nav.
Poseidon-B adapted to bluff-body wakes and Kelvin–Helmholtz mixing layers in one joint run. Same 4-channel interface. Research checkpoint on Hugging Face.
View the model →240 solver-grade trajectories. Holdout sits beyond the training Re and Mach bands — extrapolation, not a shuffled split.
View the dataset →A code model taught the exact schema of Palace, the finite-element electromagnetics solver — writes and repairs configs the solver accepts. Apache 2.0 on Hugging Face.
View the model →Different geometry, Reynolds band, or 3D. We generate the data and adapt the operator under a commercial license.
Request a campaign →Every physics build follows the same four moves — Flow-1.0 and Palace-LoRA are the public proofs, and the same run works for your regime, geometry, or solver.
A neural operator for fields (Poseidon-B, ETH Zurich CAMLab) or a 32B code model for solver tooling (Qwen2.5-Coder) — never an architecture from scratch.
Solver-grade PyFR trajectories for flow; authoring and repair tasks derived from Palace's own schema, docs, and examples. Holdout tests extrapolation, not memorization.
Joint regime training keeps the operator's original families intact; low-rank adapters leave the code model's general skills untouched.
Adapted weights plus the exact data they were trained on — drop-in whether you evaluate Poseidon-B or serve LoRA modules with vLLM.
Fast physics tools are screening instruments, not sign-off authorities. Knowing the boundary is what makes them useful.
CAE and digital-twin teams screening wakes and mixing — and simulation engineers who spend real hours writing or debugging solver configuration files.
Early design loops, parameter sweeps, and rollouts inside the validated window — and schema-exact solver configs authored or repaired in seconds instead of an afternoon.
Certification runs, fine-scale structures, and final sign-off. Surrogate fields get checked against the solver; generated configs get schema-validated and dry-run before anyone trusts a result.
These aren't demo checkpoints — both models run inside NumericalAI, our GPU simulation platform, where engineers launch cloud CFD and electromagnetics runs every day.
NumericalAI's CEM platform runs Palace-powered full-wave 3D electromagnetics on cloud GPUs. Before a run launches, Palace-LoRA validates and diagnoses the input config — wrong types, invalid enums, dangling mesh references. And when a simulation fails anyway, it works inside the AI diagnosis pipeline that traces the failure back to the config and proposes the fix.
See CEM on NumericalAI →SRS runs the same high-order PyFR engine that generated Flow-1.0's training data; CMF covers multiphase compressible CFD. On both, our models validate and diagnose input scripts and geometries before GPU time is spent — and drive the AI diagnosis pipeline that turns a crashed or diverged simulation into a corrected setup instead of a support ticket.
See SRS & CMF on NumericalAI →Zero license fees, pay-per-compute, one account across all three solvers — the fastest way to see these models working on a real problem.
Try it on numericalai.net