AeTHERON

Autoregressive Topology-aware Heterogeneous Graph Operator Network for Fluid-Structure Interaction

AeTHERON is a heterogeneous graph neural operator for surrogate modeling of body-driven fluid-structure interaction (FSI) governed by the sharp-interface immersed boundary method. The architecture mirrors the IBM discretization through a dual-graph representation coupled via sparse cross-attention, enabling fast prediction of coupled fluid-membrane dynamics. Full details in the paper.

Continuously developing project — results and code will be updated in subsequent versions.

---

Installation

git clone https://github.com/sushrutkr/AeTHERON.git
cd AeTHERON
pip install torch torch-geometric
pip install -r requirements.txt

---

Usage

# Training
python src/train.py --config input/config.json

# Inference
python src/inference.py --config input/config.json --checkpoint ref_models/

---

Model Architecture

Encoder  : fluid (R^4) + membrane (R^10) -> shared latent space (d_h = 32)
           sinusoidal time embedding (d_t = 16)
Processor: L = 10 heterogeneous message-passing layers
           intra-domain GNO + cross-domain sparse attention (d_A = 32)
           time-conditioning via LayerNorm scale-shift
Decoder  : Euler update  x_f(t+tau) = x_f(t) + tau * psi(xi_f)

---

Citation

@article{kumar2026aetheron,
  title   = {AeTHERON: Autoregressive Topology-aware Heterogeneous Graph
             Operator Network for Fluid-Structure Interaction},
  author  = {Kumar, Sushrut},
  journal = {arXiv preprint arXiv:2604.13369},
  year    = {2026}
}

Model Visual Overview