ConSolv

This repository contains the code and published model accompanying ConSolv: Solvent-Conditional Machine Learning Implicit Solvent Potential by Linying Zhang and Julija Zavadlav.

ConSolv is a solvent-conditional machine-learning implicit-solvent potential. It uses an attention-based solvent embedding and was developed using experimental solvation free energies together with ab initio training data.

Repository structure

.
├── dataset/
│   ├── descriptor.csv       # solvent descriptors used by the model
│   ├── descriptor.py        # descriptor utilities
│   └── solvatum/            # Solv@TUM data interface (Git submodule)
├── examples/SFE/
│   ├── train/               # vacuum and solvent-model training workflows
│   ├── evaluation/          # molecule and solvation free-energy evaluation
│   ├── output/
│   │   ├── vacuum/          # published vacuum parameters 
│   │   └── solvent/         # published ConSolv parameters 
│   └── postprocess/         # postprocessing and benchmark helpers
└── external/
    ├── chemtrain/           # molecular-simulation and training utilities
    └── chemutils/           # datasets and model implementations

Clone and setup

Clone the repository together with the Solv@TUM submodule:

git clone --recurse-submodules <repository-url>
cd <repository-directory>

For an existing clone, initialize the submodule with:

git submodule update --init --recursive

Use Python 3.10 or newer and install the local packages and their dependencies in a compatible JAX environment:

conda env create -f environment.yml
conda activate supramol

Models and workflows

The evaluation scripts resolve model files relative to the repository, so they do not depend on the shell's working directory:

• examples/SFE/output/vacuum/best_params_1.pkl contains the published vacuum model parameters.
• examples/SFE/output/solvent/final_params.pkl and full_init_params.pkl contain the published solvent-model parameters.

The complete large model outputs, saved evaluation results, and trajectory setup/data files are archived on Zenodo.

Training and evaluation entry points are under examples/SFE/train and examples/SFE/evaluation. Inspect an entry point's options before running it, for example:

JAX_PLATFORMS=cpu CUDA_VISIBLE_DEVICES=-1 \
  python examples/SFE/evaluation/evaluate_resolv_con_add.py --help

See examples/SFE/readme.md for the workflow outline.

Citation

If you use this code, please cite:

@article{zhang2026consolv,
  title   = {ConSolv: Solvent-Conditional Machine Learning Implicit Solvent Potential},
  author  = {Zhang, Linying and Zavadlav, Julija},
  journal = {arXiv preprint arXiv:2606.24983},
  year    = {2026}
}