NIM inference (DoMINO) + DrivAerML benchmarking

This workflow holds tutorial notebooks that call the DoMINO Automotive Aerodynamics NVIDIA Inference Microservice (NIM) from Python, map predictions onto DrivAerML VTK meshes, and compute metrics with physicsnemo.cfd.postprocessing_tools.

It complements the Hydra benchmarking workflow at workflows/benchmarking/, which runs packaged checkpoints (for example from Hugging Face) via python main.py. Use this folder when you already have the NIM API and want a notebook walkthrough; use benchmarking for matrix evaluation and CI-style configs.

Contents

Path	Description
notebooks/surface_benchmarking.ipynb	Surface VTP: NIM → interpolate to cell centers → L2 / area-weighted L2, drag–lift, plots, streamlines.
notebooks/volume_benchmarking.ipynb	Volume VTU: NIM → interpolate to points → L2, error vs. SDF, slices, optional physics checks.
notebooks/benchmarking_in_absence_of_gt.ipynb	No ground-truth setting: STL resolution variants, NIM ensemble / variance, Chamfer–Hausdorff geometry distances.

Prerequisites

1. PhysicsNeMo-CFD installed from the repository root (see the main README).
2. DoMINO NIM running and reachable (default in the notebooks: http://localhost:8000/v1/infer). Follow the NIM quickstart.
3. GPU recommended for interpolation and large VTU loads; volume notebooks assume multi‑GB meshes.

If the notebook runs in Docker alongside the NIM, start both with --network host so localhost inference works.

Headless servers may need xvfb (see setup.sh for an example apt line — optional).

API used in the notebooks

• physicsnemo.cfd.evaluation.nims.call_domino_nim — HTTP client for the NIM (see domino_nim.py).
• physicsnemo.cfd.postprocessing_tools — interpolation onto meshes, L2 and area-weighted metrics, forces, visuals, streamlines, etc.

Data

Examples download sample run_202 assets from the DrivAerML Hugging Face dataset. Volume VTUs may be split into .part files on the Hub; the notebook concatenates them before reading.

Legacy location

These notebooks were previously under workflows/deprecated/bench_example/notebooks/. They are maintained here so NIM-based tutorials live next to workflows/benchmarking/ without mixing into the deprecated layout.