NOTS-DarcyFlowRate

Model Introduction

This package reproduces the Darcy flow-rates experiment in Thompson Sampling in Function Spaces via Neural Operators. A neural operator predicts a 16x16 pressure field from a permeability field, then supports sequential selection using a negative total-flow objective.

Model Description

The run completed ten independent trials with 300 queries per trial on one DCU. The included weight/best_model.pt contains the selected trained model. The public checkpoint changes only the configuration fingerprint for portable data/output paths; its model parameter tensors are identical to the original run.

The published checkpoint SHA256 is dac0b15574242acabd25d5bc52b7a415ec72e1d55411433402affe61876131f7; the original training checkpoint SHA256 was 2cf139f9cf75a01064fe54b5e8ba0d4e58a4b76a39aa7cf924320acbee4336b2. The archived results/inference_metrics.json records the original file hash; rerunning inference writes a fresh metric file for the published checkpoint.

Intended Use

Task	Description
Darcy surrogate inference	Predict pressure fields on 16x16 permeability grids
Sequential optimization	Study neural-operator Thompson sampling for flow-rate objectives

Usage

Download the Model Package

hf download OneScience-Group/NOTS-DarcyFlowRate --local-dir ./NOTS-DarcyFlowRate
cd NOTS-DarcyFlowRate

Install Dependencies

Use an allocated GPU/DCU compute node for the provided scripts. The tested formal run used a PyTorch DCU environment; CPU execution of the full pipeline is not claimed.

pip install -r requirements.txt
pip install 'modelscope[datasets]'

Training Data

Hugging Face hosts only this model package, not a dataset mirror. Download the author's original 16x16 Darcy train/test tensors from OneScience/NOTS-Darcy-16 on ModelScope:

modelscope download --dataset OneScience/NOTS-Darcy-16 --local_dir ./data

The package expects data/darcy_train_16.pt with 1,000 pairs and data/darcy_test_16.pt with 50 pairs. The original files are from the NOTS authors' repository.

Train

From the package root on an allocated GPU/DCU compute node:

python scripts/train.py

This runs ten configured seeds and 300 queries per seed. Save a copy of the bundled checkpoint and results before retraining.

Inference and Results

python scripts/inference.py
python scripts/result.py

The reproduced final average regret was 0.289031 +/- 0.016963 across ten trials, versus 0.012 +/- 0.001 in paper Table 1. The measured test pressure-field mean relative L2 was 0.339917. The regret is far from the paper result; this package is a documented reproduction attempt, not a paper-accuracy claim.

OneScience Repositories

Platform	Main repository	Skills repository
Gitee	OneScience	OneSkills
GitHub	OneScience	OneSkills

Citation and License

• Paper: Thompson Sampling in Function Spaces via Neural Operators.
• Dataset: NOTS authors' data directory, originally from NeuralOperator and redistributed there under the accompanying MIT license. The data is hosted separately on ModelScope.