Skip to content

NAVIER-CFD logo

NAVIER-CFD

Neural and Agentic Verification, Integration, Evaluation, and Recommendation for Computational Fluid Dynamics

NAVIER-CFD is a CFD-first Python platform for neural PDE solvers, hybrid acceleration, benchmark data, task-aware model recommendation, and agentic experiment planning.

Open the project website Launch the interactive recommender View the repository

What makes it different

  • CFD-first taxonomy spanning acceleration, surrogates, general PDE solvers, geometry models, specialized tasks, foundation models, inverse methods, uncertainty, particle/multiphase, and generative models.
  • Uniform access to PDEBench, CFDBench, RealPDEBench, AirfRANS, DrivAer datasets, The Well, APEBench, ScalarFlow, and arbitrary Hugging Face datasets.
  • Explainable model ranking by dimension, mesh, geometry, temporal regime, physics, memory, conservation, uncertainty, and numerical role.
  • Deterministic and LLM-ready experiment planning.
  • CFD-aware field, spectral, conservation, rollout, OOD, quantity-of-interest, and computational-cost metrics.
  • Metadata-first external adapters that never execute untrusted repositories automatically.

navier-cfd

Install

pip install navier-cfd

Five-minute tour

navier models list --category geometry
navier datasets list
navier datasets discover "computational fluid dynamics"
navier datasets download cfdbench --local-dir ./data/cfdbench --pattern "cylinder/**"
navier recommend --problem cylinder_wake --task surrogate --dimension 2 --mesh structured --temporal autoregressive --geometry varying --top-k 8
navier agent plan "Benchmark RealPDEBench cylinder sim-to-real forecasting with 24 GB VRAM"

Recommender status

The recommender is operational as an explainable rule-based engine. It applies hard compatibility filters and then ranks candidates by task role, physics, geometry, mesh, temporal mode, rollout requirements, conservation, uncertainty, hardware memory, and integration maturity. See the recommender documentation.