Repository containing the charge prediction model and the scripts to reproduce the analyses shown in the paper "Machine-Learned Electrostatic Potentials for Accurate Hydration Free Energy Calculations" (https://arxiv.org/abs/2512.13579).

To use the model it is necessary to have a copy of the MACEOFF23-large model available here: https://github.com/ACEsuit/mace-off .

To reproducee the analyses, the following is necessary:

• FreeSolv dataset: https://github.com/MobleyLab/FreeSolv
• Aquamarine dataset: https://zenodo.org/records/10208010
• MACE descriptors + ESP charges for Aquamarine molecules: https://doi.org/10.5281/zenodo.17790330

The exact environment as used for calculations and analysis can be retrieved from the file environment.yml

The repository is structured as follow:

• ahfe_calculations contains all the necessary to reproduce the parity plots shown in the paper: AM1-BCC, ESP, RESP, 1-shot, and Boltzmann percentile charges, as well as energies and errors computed with each method. The subfolder "molecules" contains the .sdf and .xyz files for the selected molecules.
• boltzmann_percentile contains an example of MD simulations and BP charge assignment for one molecule, and the script to replicate the analysis of charge variability across conformations.
• charge_prediction contains the model, an example of how to use it to predict charges, and scripts to train and validate the model on a subset of the real training data.
• functional_groups_analysis contains the script to replicate the analaysis on the functional groups.
• solar_chgs contains an alternative charge prediction model, not presented in the main text, based on SO3LR descriptors and available for commercial usage.