UPT

Model Introduction

UPT (Universal Physics Transformer), proposed by Benedikt Alkin and collaborators, is a general-purpose physics Transformer. It encodes irregular meshes or particle data into a compact latent representation and propagates information in latent space for flexible, scalable physical-surrogate modeling.

This repository is an independent OneScience reproduction of the UPT ShapeNetCar experiment without an SDF. The model predicts steady surface pressure from three-dimensional vehicle-surface coordinates and does not use SDF values, normals, mesh connectivity, or flow variables as model inputs.

Paper: Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators

Model Description

For ShapeNetCar, UPT takes vehicle-surface coordinates ((x,y,z)) and predicts scalar pressure at specified surface-query coordinates.

The architecture consists of coordinate encoding, a Perceiver encoder, a latent-space Transformer, a Perceiver decoder, and pressure projection. Continuous sine/cosine positional encodings and an MLP first map scaled coordinates to input features. Perceiver cross-attention compresses a variable number of input points into learnable latent tokens, and Transformer blocks model global geometric relationships in latent space. During decoding, encoded query coordinates attend to the latent representation before normalization and linear projection produce surface-pressure predictions.

Intended Uses

Use case	Description
Vehicle-surface pressure	Predict steady aerodynamic pressure from ShapeNetCar surface coordinates.
Irregular point-set modeling	Regress physical values on 3D surface point sets without a regular grid or explicit mesh connectivity.
Scientific surrogate	Approximate expensive numerical simulation within a fixed data distribution and flow regime.
Variable-point inference	Accept variable-size input sets and query pressure at surface coordinates within the training distribution.

Usage

1. OneCode

Launch the OneCode AI-for-Science environment

2. Manual Setup

Hardware requirements

A GPU or DCU is recommended.
A CPU can run imports and small connectivity checks, but full training and inference will be slow.
DCU users should install DTK 25.04.2 or later, or the OneScience-recommended version for the cluster.

Download the model repository from Hugging Face

pip install -U huggingface_hub
hf download OneScience-Group/UPT --local-dir ./UPT
cd UPT

Install the runtime environment

DCU environment

# Activate DTK first.
conda create -n onescience311 python=3.11 -y
conda activate onescience311
pip install onescience[cfd-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

GPU environment

conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12
conda activate onescience311
pip install onescience[cfd-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

Download the training dataset from Hugging Face

hf download OneScience-Group/ShapeNetCar \
  --repo-type dataset \
  --local-dir ./data

Set data.root in config/config.yaml to the dataset's preprocessed_data directory. The dataset contains 889 vehicle geometries, split with a fixed random seed into 700 training and 189 test samples. Each preprocessed sample includes:

pos.npy: all node coordinates, shape [num_nodes, 3].
surf.npy: surface-node mask, shape [num_nodes].
y.npy: target physical fields, shape [num_nodes, 4]; the first three channels are velocity and the fourth is pressure.
x.npy: node features, shape [num_nodes, 7]. This implementation uses normals only to filter surface nodes; they are not model inputs.
edge_index.npy: graph connectivity, which is not used as a model input.

The loader selects 3,586 valid surface points per mesh, rescales coordinates, and normalizes pressure with the configured mean and standard deviation. During training, 80%–100% of each sample's surface points are randomly selected as encoder inputs while the pressure loss is computed on all surface points. During testing, all valid surface points are used as both inputs and queries.

Train

config/config.yaml contains the no-SDF ShapeNetCar reproduction configuration.

python scripts/train.py --config config/config.yaml

Training uses masked MSE on normalized surface pressure. Per-epoch training and test metrics are written to results/train_metrics.jsonl. The script prints batch loss, learning rate, normalized pressure MSE, normalized pressure MSE × 100, and original-scale pressure RMSE. The checkpoint with the lowest test normalized pressure MSE × 100 is saved to weight/best_model.pth.

Pretrained weights

The repository includes weight/best_model.pth, trained on ShapeNetCar and ready for surface-pressure inference and numerical evaluation.

Inference and evaluation

Run inference on the fixed set of 189 test samples:

python scripts/inference.py \
  --config config/config.yaml \
  --checkpoint weight/best_model.pth \
  --device auto

The script reports normalized pressure MSE, normalized pressure MSE × 100, normalized pressure MAE, original-scale pressure MSE, and original-scale pressure RMSE. It writes:

results/inference_metrics.json
results/predictions/
└── <sample_id>.npz

inference_metrics.json contains aggregate and per-sample metrics plus paper-reference values. Each <sample_id>.npz stores surface coordinates, normalized and original-scale predictions and targets, and absolute errors.

Visualization

python scripts/result.py --config config/config.yaml

Outputs:

results/figures/
├── training_metrics.png
├── <sample_id>_pressure_comparison.png
└── visualization_metadata.json

training_metrics.png shows normalized training MSE and test normalized MSE × 100 with the paper reference.
<sample_id>_pressure_comparison.png compares target surface pressure, UPT prediction, and absolute pressure error in 3D.
visualization_metadata.json records prediction and evaluation sources, output paths, and plotting parameters.

OneScience

Platform	OneScience repository	OneSkills repository
Gitee	https://gitee.com/onescience-ai/onescience	https://gitee.com/onescience-ai/oneskills
GitHub	https://github.com/onescience-ai/OneScience	https://github.com/onescience-ai/oneskills

Citation and License

Paper: Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators.
Official implementation: ml-jku/UPT, released under the MIT License.
This repository uses the Hugging Face-compatible MIT identifier (mit). ShapeNetCar and all other third-party resources remain subject to their original licenses and terms.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

OneScience-Group
/

UPT