BLIPs: Bayesian Learned Interatomic Potentials - Official Repository

Dario Coscia , Pim de Haan , Max Welling

[Paper]

📔 Description

BLIP is a scalable, architecture-agnostic variational Bayesian framework for training or fine-tuning machine learning interatomic potentials (MLIPs) - or message passing neural networks more in general. BLIP delivers well-calibrated uncertainty estimates and minimal computational overhead for mean prediction at inference time, while integrating seamlessly with both standard and equivariant message-passing architecture

By adopting BLIP, you can:

• Easily convert existing MLIPs into Bayesian models.
• Maintain the core architecture.
• Achieve more reliable uncertainty estimates, making your models more robust and interpretable.

This repository contains the source code to reproduce the experiment, as well as a simple jupyter notebook to start playing around with BLIP wrapper.

📝 Step-by-Step Guide: Build BLIP from MLIP

If you're interested in applying BLIP to your favourite MLIP (or more in general GNN), we’ve created a Jupyter notebook that can be run on Google Colab. In this notebook, we walk through all the essential steps to build BLIP and give you a hands-on introduction to experimenting with our BayesianModelWrapper!

🛠️ Installation and set up

Clone the git repository, create a virtual conda environment and install the requirements.

# clone project   
git clone https://github.com/dario-coscia/blip
cd blip

# create virtual environment
conda create -n venv python=3.11
conda activate venv

# install project
python -m pip install .   

💾 Data to reproduce the experiments

To download the data used in the experiments run:

sh experiments/{nbody/ammonia/silica}/generate/generate.sh

Important note: The data used in this study were either generated or sourced from previously cited works (see our article), to which full credits are given.

💻 Run Experiments

In order to run the experiments we provide two ways.

The first way is directly running the main file:

python experiments/{nbody/ammonia/silica}_main.py --{extra_arguments}

The extra_arguments are found in experiments/{nbody/ammonia/silica}/args.py.

The second option is to run the shell script on a SLURM based-cluster. First, configure the sbatch file, then run:

sh shell/train_{nbody/ammonia/silica}.sh

You will run the same training (seed and hyperparamters) as in the main article.

💻 Analyse Experiments

Once the model is trained it saves the logs and checkpoints into specific directories, namely logs and ckpt, followed by the experiment type (nbody/ammonia/silica). The experiments/{nbody/ammonia/silica}_main.py file can be called using run_type=test extra_argument, this will save a .pt file with the experiment test output and data.

Citation

@misc{coscia2025blipsbayesianlearnedinteratomic,
      title={BLIPs: Bayesian Learned Interatomic Potentials}, 
      author={Dario Coscia and Pim de Haan and Max Welling},
      year={2025},
      eprint={2508.14022},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.14022}, 
}