Neural network architectures for accelerating computational fluid dynamics simulations.

"Akis" means "flow" in Turkish.

akis applies deep learning to CFD problems: solving Navier-Stokes equations with PINNs, predicting flow fields with CNNs, analyzing unstructured meshes with GNNs, and generating meshes with GANs.

Solve Navier-Stokes equations by embedding physics constraints directly into the loss function. No mesh required — the network learns the continuous solution field.

Predict steady-state flow fields from geometry inputs. Train on simulation data, infer in milliseconds instead of hours.

Operate on unstructured meshes directly using graph neural networks. Predict where mesh refinement is needed without running the full simulation.

Generate high-quality computational meshes from geometry descriptions. Trained on meshes from existing simulations.

akis/
  data/
    cfd_simulations/
    meshes/
    processed/
  notebooks/
    1_pinn_navier_stokes.ipynb
    2_cnn_flow_prediction.ipynb
    3_gnn_mesh_analysis.ipynb
    4_gan_mesh_generation.ipynb
  src/
    models/
      pinn/
      cnn/
      gnn/
      gan/
    data_processing/
    visualization/
    utils/
  tests/
  requirements.txt

git clone https://github.com/selimozten/akis.git
cd akis
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Requires a CUDA-capable GPU for training.

• PyTorch for model training
• NumPy / SciPy for numerical computations
• OpenFOAM for validation against traditional CFD
• NetworkX for graph representations

1. Raissi et al. (2019) — Physics-informed neural networks for solving PDEs
2. Guo et al. (2016) — CNNs for steady flow approximation
3. Pfaff et al. (2021) — Learning mesh-based simulation with graph networks
4. Chen & Jaiman (2022) — MeshingNet for deep learning mesh generation

Apache 2.0