nvidia-omniverse-blueprints/digital-twins-for-fluid-simulation/rtdt-aero

RTDT Auto-Aero

Read rtdt-core and rtdt-customize first. This skill maps their contracts onto the shipped specialization.

File map

If a path changes, locate the current file before editing; do not guess an abbreviated schema path.

Change a car geometry axis

The active shipped geometry controls are Rims, Mirrors, and Spoiler. Ride_Height is not a shipped control axis. It is a nested CFD payload dimension in CarCFD.usda, selected as High by Base.usda, but has no AppState attribute, manifest control/binding, active matching CarHero variant, or frontend widget. Treat it as payload inventory, not as an already wired control.

Use the active controls as the template for a new axis:

1. Add the typed token/default to the correct AppState schema and regenerate the schema.
2. Add matching variant sets to the hero and CFD compositions as needed.
3. In CarCFD.usda, keep the outer OmniSciDataset with

OmniSciFileFormatArgsPythonAPI and RtwtModelAPI; put each PLY payload and omni:rtwt:model:tag in the correct variant leaf.

1. Add the manifest control and Variant bindings. Keep the control ID equal to the schema short

key.

1. Author the AppState default, bridged state key, and frontend widget. Keep that default aligned in

Python reset_state() and JavaScript resetControlState(). For an inference axis, declare its <key>_options state and add the matching setOpt(...) mapping in trame-app/static/app_streamer.js so lite-mode restrictions reach the UI. If the axis affects the fixed comparison result, also update FIXED_COMBO and its baked data.

1. Validate schema/manifest/stage wiring and run live inference in the standard profile.

For a genuinely new geometry rather than another car variant, follow geometry-recipe.md.

Change an output or visualization

Keep these synchronized:

• AeroNIM tensor name and aeronim/rtwt/config.pbtxt;
• RtwtResultFieldSelectionAPI:<wire-id> on the inference prim;
• paired OmniSciArrayAPI:<Field> and OmniSciFieldAPI:<Field> on /World/InferenceResults;
• [outputs.<wire-id>] in workflow.toml;
• cae:viz:field_selection:<role>:fieldNames on every consuming operator;
• lite cache payloads and sidecars.

Planar slices use Materials/UnlitScalarColor from cae/mdl/basic.mdl. Do not restore the removed texture-slice material pattern.

Cache consequences

Every inference AppState axis contributes to result identity. An inference-role control with cache_values expands the intended pre-cache product, but that manifest product is not itself proof of complete cache coverage.

The shipped lite cache intentionally covers eight tuples: four velocities by two spoiler states, with mirrors fixed at On and rims fixed at Standard. The manifest's inference cache_values describe a 32-tuple Cartesian product. Per-axis options reported by the runtime are unions across available entries; they do not prove that an arbitrary cross-axis tuple exists.

After changing inference geometry, domain, outputs, or cache axes:

1. explicitly choose the tuples the lite release will support;
2. create a new empty staging cache, then run Kit on the host with reachable standard-profile Triton,

--/exts/omni.rtwt.inference/offline_mode=true and --/exts/omni.rtwt.inference/generate_if_missing=true, and an absolute --/exts/omni.rtwt.inference/offline_cache_dir=<staging-path>; the cache key omits the model payload and domain, so reusing the tracked cache can produce stale hits; drive each approved tuple as documented in README deployment modes; the shipped Compose mounts the cache read-only and its CACHE button only warms memory;

1. verify payload and sidecar consistency for every intended tuple;
2. test every supported tuple in lite mode, plus an unsupported tuple to confirm it is rejected

without leaving stale results;

1. do not call the lite integration complete while any supported tuple falls back, freezes, or is

absent.

For comparison result sets, follow tools/comparison/README.md. Its container command uses the isolated RTWT Warp SimData runtime; do not substitute the AeroNIM base interpreter path or a retired in-repository post-processing tool.

Fixed comparison view

[comparison] names pre-baked result sets and their .npz files. The current comparison callback maps the logical selection to resultSetOverride; it does not perform live inference.

Changing comparison files is a coordinated data change:

• keep manifest names, callback behavior, stage delta field mappings, and baked filenames aligned;
• keep the frontend FIXED_COMBO in trame-app/static/app_streamer.js aligned with the baked

comparison tuple, including when a new inference control or default affects result identity;

• preserve the signed-difference convention used by the visualization;
• regenerate with the same native result format consumed by the Kit-CAE 3.0 inference path;
• validate the fixed view in the browser after loading each result set.

Do not generalize this recipe to arbitrary result sets without updating the callback and frontend.

Validate

Static manifest/schema validation:

tools/validate/rtdt-validate.sh \
  --manifest source/specializations/auto-aero/workflow/workflow.toml \
  --schemas RtwtInferenceAppStateAPI RtwtVizAppStateAPI

Add --stage source/specializations/auto-aero/stages/Main.usda when checking targets, variant sets, result wiring, or colormap readouts. Base.usda is the authoring layer for the base scene, domain, inference, and results; it is not the full validation scope. Main.usda composes the visualization layers and is the correct static approximation of the application stage. Interpret any remaining warnings against runtime-authored state.

Resources

• geometry-recipe.md — exact Kit-CAE 3.0 geometry procedure.
• source/specializations/auto-aero/workflow/workflow.toml — active control/output registry.
• skills/evals/ — deterministic Kit-hosted validator smoke fixtures.