---
title: GeoMPNN
canonical_url: "https://www.modelscope.cn/models/wangqi0928/GeoMPNN"
md_url: "https://www.modelscope.cn/models/wangqi0928/GeoMPNN.md"
repository: wangqi0928/GeoMPNN
chinese_name: GeoMPNN
last_updated: 2026-07-17
library_name:
  - pytorch
frameworks:
  - Pytorch
downloads: 0
stars: 0
---

# GeoMPNN

> GeoMPNN - wangqi0928 在 ModelScope 开源的模型。对论文 INFINITY: Neural Field Modeling for Reynolds-Averaged Navier-Stokes Equations（arXiv:2307.13538）中模型在 AirfRANS 数据集上的复现实现。代码实现完全基于论文正文（Figure 1、Algorithm 1、Table 1）及公开数据构建，未参考任何官方或第三方实现

- **Repository**: wangqi0928/GeoMPNN
- **Downloads**: 0
- **Stars**: 0
- **Last updated**: 2026-07-17

Source: https://www.modelscope.cn/models/wangqi0928/GeoMPNN

---

# GeoMPNN

**Geometry-Aware Message Passing Neural Network for Modeling Aerodynamics over Airfoils**

This model won the Best Student Submission award at the NeurIPS 2024 ML4CFD Competition (4th place overall).

## Model

GeoMPNN is a geometry-aware message passing neural network designed for aerodynamic modeling over airfoils. It uses:

- A latent graph representation on the airfoil surface to encode geometry
- Message passing on a directed, bipartite graph to propagate geometry information to all collocation points
- Hybrid Polar-Cartesian coordinate embeddings with sinusoidal and spherical harmonics bases
- Multiple coordinate systems (leading edge and trailing edge) for distinguishing spatial regimes

## Source Code

The GeoMPNN source code is part of the [AIRS library](https://github.com/divelab/AIRS).

## Structure

- `config/` - Configuration files (`config.yaml`, manifest, stats, evaluation results)
- `model/` - Model package (source code in AIRS library)
- `scripts/` - Training and evaluation scripts (in AIRS library)
- `weight/` - Pre-trained model checkpoint files

## Checkpoints

| File | Description |
|---|---|
| `weight/ux_seed0.pt` | Ux (streamwise velocity) model weights |
| `weight/uy_seed0.pt` | Uy (vertical velocity) model weights |
| `weight/p_seed0.pt` | Pressure field model weights |
| `weight/nut_seed0.pt` | Turbulent viscosity model weights |

## Data

This model is evaluated on the AirfRANS dataset.