vRBA

Model and experiment

This package reproduces the bubble dynamics operator learning experiment from the paper. The bubble-dynamics operator model uses 200 sensors, width 100, four hidden layers, and a 100-dimensional latent representation with variational residual-based attention.

The package includes model code, config/config.yaml, training, inference and visualization scripts, the recorded results/, and the selected weight/best_model.pth checkpoint.

Data

Download OneScience/vRBA-Bubble with modelscope download --dataset OneScience/vRBA-Bubble --local_dir ./data. res_1000.npz contains del_p and R [1500,1000] float64, and t [1000]. The model uses the first 1,000 trajectories, split 800/100/100.

Original dataset: https://github.com/jdtoscano94/NABLA-SciML/tree/main/vRBA_variational_residual_based_attention_PINNs_Operator_learning Download the dataset before running inference where an external dataset is required. The release configuration uses relative paths, not SCNet personal-directory paths.

Installation and usage

hf download OneScience-Group/vRBA --local-dir ./vRBA
cd vRBA
pip install -r requirements.txt
python scripts/inference.py
python scripts/result.py

For a fresh training run, use an allocated GPU/DCU compute node with a compatible PyTorch build, keep a separate copy of the supplied checkpoint, then run python scripts/train.py from the package root. The full budget is specified in config/config.yaml. Cluster-specific Slurm scripts, if present, need adaptation outside the original cluster.

Training and reported result

The run completed 10,000 epochs and 160,000 optimizer updates without early stopping. The best checkpoint was selected at epoch 9,971. Of the 1,500 public trajectories, the first 1,000 were split 800/100/100; the author did not publish exact paper sample indices.

On 100 test trajectories, mean physical-radius relative L2 was 0.000834989 and normalized-scale relative L2 was 0.0477183. Journal Table 3 reports 0.00516 for exponential vRBA, but the split and aggregation protocol are not fully aligned; no exact accuracy-equivalence claim is made. See results/ for the recorded metrics and visualizations. The supplied checkpoint comes from a completed run; another run may differ.

Source and license

• Paper: https://doi.org/10.1038/s44387-026-00084-4.
• Dataset source: https://github.com/jdtoscano94/NABLA-SciML/tree/main/vRBA_variational_residual_based_attention_PINNs_Operator_learning.
• This model package uses the license indicated in the card metadata. Dataset redistribution terms, where not separately stated, must not be inferred from the code license.