ConSolv Attention Model Delivers Transferable Implicit Solvation Across 66 Organic Solvents
ConSolv is a solvent-conditional implicit-solvent MLP trained on 66 organic solvents that outperforms classical and selected ab initio methods on solvation free-energy benchmarks and matches experimental NMR data. The attention-based design supports generalization and interpretability but is currently restricted to common solvents. Further validation on diverse chemical spaces is required.
The preprint introduces a single solvent-conditional MLP that embeds solvent identity via an attention block, allowing one model to handle both polar and non-polar environments. Training combined experimental solvation free energies with quantum-mechanical reference calculations, producing a transferable potential that the authors benchmark against multiple datasets. Performance exceeds selected continuum models and explicit-solvent force fields on held-out solutes, with particular gains in chloroform and other organic media relevant to synthesis and battery electrolytes.
Attention weights provide interpretable solvent-solute interaction maps that align with known NMR shifts for gamma-fluorohydrin in chloroform, offering a route to explainable predictions beyond scalar energies. The architecture's extensibility is emphasized, yet the current training set remains limited to common organic solvents, leaving ionic liquids and extreme conditions unaddressed.
Next steps include expanding the solvent corpus and coupling ConSolv to reactive ML potentials for condensed-phase reaction networks. External validation on blind solvation challenges would strengthen claims of broad transferability.
Zhang et al.: Blind external test sets will show mean absolute error below 0.6 kcal/mol on at least 20 new solvents within 12 months.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2606.24983)
- [2]Supporting Source(https://pubs.acs.org/doi/10.1021/acs.jctc.3c01245)