Key Findings
AnySolv enables the integration of solvent effects into existing gas-phase MLIPs by predicting the difference between solvated and gas-phase DFT energies and forces. In external benchmarks, the model reproduces DFT-SMD solvation energies to a mean absolute error of 2–4 meV. The tool matches the accuracy of DFT-SMD against experimental solvation free energies while operating at a computational cost more than three orders of magnitude lower. These capabilities allowed the researchers to screen the mechanisms of 120 nucleophilic aromatic substitution reactions in six phases and reproduce experimentally observed conformational preferences of alanine dipeptide in molecular dynamics simulations.
Technical Details
AnySolv is an equivariant message-passing graph neural network. Key technical specifications include:
- Training Data: Trained on 1.4 million solvation energies and forces computed with DFT and SMD in 36 solvents.
- Theory Level: Calculations were performed at the ωB97M-V/def2-TZVPD level of theory, using geometries primarily from OMol25.
- Scope: The training set spans charged, open-shell, and transition-state structures.
- Solvent Support: By using continuous physical descriptors, the model supports 179 solvents, including those not present in the training set.
Background & Context
Most MLIPs are trained on gas-phase data, limiting their use in solution-phase chemistry where solvents can shift reaction barriers and stabilize charged intermediates. While explicit solvent representation is possible, it requires adding hundreds or thousands of atoms. Existing implicit models like PCM and COSMO require self-consistent electron density, which standard MLIPs do not provide. Previous ML-based attempts were often limited to neutral solutes, lacked force predictions, or required retraining for every new solvent or base potential.
Significance & Outlook
AnySolv provides a general strategy for incorporating transferable solvation corrections into existing MLIP workflows. By removing the need for retraining and significantly reducing computational overhead compared to DFT-SMD, it makes large-scale simulations of solvent-dependent reactivity and molecular dynamics feasible.
Source: https://arxiv.org/html/2610.17226v1
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
