Solvating machine-learning interatomic potentials for solution-phase chemistry

arXiv:2610.17226 · physics.chem-ph, cs.LG · Submitted 2026-10-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Solvating machine-learning interatomic potentials for solution-phase chemistry".

Jane: The gist: AnySolv is a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're talking about "Solvating machine-learning interatomic potentials for solution-phase chemistry" today, which is this new method that lets you add solvent corrections to any gas-phase potential without retraining it at all.

Jane: Right, and the main thing is that this lets you take a gas-phase potential and make it work in liquid or solution environments for chemical reactions and molecular dynamics without having to explicitly sample every single configuration in the solvent.

Lu: The authors establish a general strategy for incorporating these transferable solvation corrections into existing machine learning interatomic potential workflows by providing corrections that can be added to any gas-phase potential and remain valid for charged and open-shell solutes at geometries far from equilibrium.

Tom: It’s about giving you this single point correction, defined as E(R, q, M, s) = E solv - E gas at a given molecular geometry.

Jane: That means you just add that correction term to whatever gas-phase potential you already have without needing extra steps for sampling solvent configurations.

Meng: So the big takeaway is that it takes the complexity of solvent effects and turns it into a fixed, computable correction based only on the molecular coordinates and charge.

Lu: It moves the complexity of handling solvent effects from being an expensive, configuration-sampling problem into something you can just compute based on those inputs.

Jane: And they show that this approach reproduces DFT-SMD solvation energies and forces with a mean absolute error of two to four milli-electron volts on external benchmarks.

Tom: That accuracy level is pretty solid, especially since they claim it matches the accuracy of DFT-SMD against experimental solvation free energies.

Lu: The implications are huge because it enables high-throughput screening of chemical reactions, like those one hundred and twenty nucleophilic aromatic substitution reactions across six different phases.

Jane: And for molecular dynamics, it reproduces experimental solvent effects on things like alanine dipeptide conformational preferences, which is something we actually care about in biology.

Tom: It’s a powerful tool because it lets you screen mechanisms and run these kinds of molecular dynamics simulations much faster than if you had to do the full quantum chemistry work every time.

Meng: For deployment, the fact that it works across different solvent types and doesn't require retraining is a huge plus for reliability at scale.

Lu: The limitation they mention is that the accuracy ultimately depends on their underlying reference data, which includes their DFT-SMD calculations and whatever gas-phase potential you combine it with.

Jane: So, while it’s very transferable, you still need a good starting point from your existing model to get the best results out of AnySolv.

Tom: It’s about establishing this general strategy for incorporating these corrections into existing MLIP workflows with AnySolv that sets the stage for routine, high-throughput investigations of solvent-dependent reactivity and molecular dynamics.

Lu: This work opens up a route toward using these powerful machine learning interatomic potentials in complex solution chemistry environments where before we were severely limited by solvent dependency.

Jane: It really bridges the gap between fast gas-phase simulations and the slow, expensive reality of full solution chemistry by making solvation effects part of the standard workflow.

Meng: I’d still want to know how fast they can get this correction into standard simulation pipelines, because that’s the real test for practical utility.

Lalam: They established a general strategy for incorporating these corrections into existing MLIP workflows with AnySolv.

The paper's summary: Tom: So, we've talked about how AnySolv works right now, and now we need to talk about what they suggest improving it to make it even better for real-world use in simulations.

Jane: Right, so the main thing is that this paper isn't just presenting a finished model; they’re pointing out exactly where the next steps should go if you want to push this further.

Tom: They mention that the accuracy of AnySolv is tied directly to the underlying reference data they used, so improving those DFT-SMD calculations would definitely boost the performance of this whole system.

Lu: The authors suggest that they could use experimental solvation free energies as a way to create even more accurate labels for their model training.

Jane: So, instead of relying on DFT-SMD data to train it, they propose using those real-world experimental values to fine-tune the AnySolv potential.

Tom: If you can get better reference data from experiments, you're essentially improving the quality of the learning signal for the AI model.

Lu: And they also explore how to make this correction term even more flexible so it can handle a wider variety of chemical environments beyond just solvents represented by those descriptors.

Jane: That means making sure the representation—those one hundred twenty-eight-dimensional vectors—is strong enough to capture every nuance of a different solvent.

Tom: And from an engineering standpoint, they look at how to make this correction term even more efficient so it adds less overhead during the actual molecular dynamics runs.

Meng: I'm interested in that efficiency part because if the correction term itself becomes too slow to compute, then adding it won't help a lot when you’re running simulations on really big systems.

Lu: They suggest looking into more advanced representation learning techniques to see if they can get a smaller vector embedding with the same or better predictive power.

Jane: It’s about finding that sweet spot between having enough detail in your solvent description and keeping the overall computation light.

Tom: Basically, they're suggesting we refine the way we map those physical descriptors into the AI's understanding so it gets even smarter about what makes a solvent behave a certain way.

Lu: It’s about pushing the boundaries of representation learning to see if we can achieve this transferability without needing to retrain from scratch every time.

Jane: So, they are looking at making the model more general and more accurate for a wider range of chemicals, not just the ones in their training set.

Tom: It’s about taking this starting point and systematically improving the data pipeline to ensure that when you use it on something new, it performs reliably.

Meng: If they can make that correction term faster and more robust across different solvent types, then this moves from a neat research tool to something you could actually deploy reliably in a production environment.

Lu: The paper shows this is a general method that can be applied across different MLIP architectures, which suggests it's not tied to just one specific type of potential.

Jane: It really opens up the door to simulating things that are inherently solution-dependent, whether it’s how proteins fold or how molecules react in a liquid.

Tom: So what this means for the people listening? It means we can finally use these powerful machine learning interatomic potentials in complex solution chemistry environments where before we were severely limited by solvent dependency.

Jane: It really bridges the gap between fast gas-phase simulations and the slow, expensive reality of full solution chemistry by making solvation effects part of the standard workflow.

Lu: It's a solid piece of research because it shows how to make solvent effects computable rather than just something we have to guess at during simulations.

Tom: We're talking about a model called "Solvating machine-learning interatomic potentials for solution-phase chemistry" that’s pretty efficient at this.

Lalam: They established a general strategy for incorporating these corrections into existing MLIP workflows with AnySolv.

The paper's improvements: Tom: So, we're talking about improving AnySolv, which is this transferable implicit solvation potential that lets you add solvent corrections to any gas-phase potential without retraining it at all.

Jane: Right, so the main thing is that this paper isn't just presenting a finished model; they’re pointing out exactly where the next steps should go if you want to push this further.

Tom: They mention that the accuracy of AnySolv is tied directly to the underlying reference data they used, so improving those DFT-SMD calculations would definitely boost the performance of this whole system.

Lu: The authors suggest that they could use experimental solvation free energies as a way to create even more accurate labels for their model training.

Jane: So, instead of relying on DFT-SMD data to train it, they propose using those real-world experimental values to fine-tune the AnySolv potential.

Tom: If you can get better reference data from experiments, you're essentially improving the quality of the learning signal for the AI model.

Lu: And they also explore how to make this correction term even more flexible so it can handle a wider variety of chemical environments beyond just solvents represented by those descriptors.

Jane: That means making sure the representation—those one hundred twenty-eight-dimensional vectors—is strong enough to capture every nuance of a different solvent.

Tom: And from an engineering standpoint, they look at how to make this correction term even more efficient so it adds less overhead during the actual molecular dynamics runs.

Meng: I'm interested in that efficiency part because if the correction term itself becomes too slow to compute, then adding it won't help a lot when you’re running simulations on really big systems.

Lu: They suggest looking into more advanced representation learning techniques to see if they can get a smaller vector embedding with the same or better predictive power.

Jane: It’s about finding that sweet spot between having enough detail in your solvent description and keeping the overall computation light.

Tom: Basically, they're suggesting we refine the way we map those physical descriptors into the AI's understanding so it gets even smarter about what makes a solvent behave a certain way.

Lu: It’s about pushing the boundaries of representation learning to see if we can achieve this transferability without needing to retrain from scratch every time.

Jane: So, they are looking at making the model more general and more accurate for a wider range of chemicals, not just the ones in their training set.

Tom: It’s about taking this starting point and systematically improving the data pipeline to ensure that when you use it on something new, it performs reliably.

Meng: If they can make that correction term faster and more robust across different solvent types, then this moves from a neat research tool to something you could actually deploy reliably in a production environment.

Lalam: They established a general strategy for incorporating these corrections into existing MLIP workflows with AnySolv.

Conclusion: Tom: So, to wrap up what we just heard today on "Solvating machine-learning interatomic potentials for solution-phase chemistry," we’re talking about this new method that lets you add solvent corrections to any gas-phase potential without retraining it at all.

Jane: Right, and the main thing is that this lets you take a gas-phase potential and make it work in liquid or solution environments for chemical reactions and molecular dynamics without having to explicitly sample every single configuration in the solvent.

Lu: The paper establishes a general strategy for incorporating these transferable solvation corrections into existing machine learning interatomic potential workflows by providing corrections that can be added to any gas-phase potential and remain valid for charged and open-shell solutes at geometries far from equilibrium.

Tom: It’s about giving you this single point correction, defined as delta E (R, q, M, s) equals Esolv minus Egas at a given molecular geometry.

Jane: That means you just add that correction term to whatever gas-phase potential you already have without needing extra steps for sampling solvent configurations.

Meng: So the big takeaway is that it takes the complexity of solvent effects and turns it into a fixed, computable correction based only on the molecular coordinates and charge.

Lu: It moves the complexity of handling solvent effects from being an expensive, configuration-sampling problem into something you can just compute based on those inputs.

Jane: And they show that this approach reproduces DFT-SMD solvation energies and forces with a mean absolute error of two to four milli-electron volts on external benchmarks.

Tom: That accuracy level is pretty solid, especially since they claim it matches the accuracy of DFT-SMD against experimental solvation free energies.

Lu: The implications are huge because it enables high-throughput screening of chemical reactions, like those one hundred and twenty nucleophilic aromatic substitution reactions across six different phases.

Jane: And for molecular dynamics, it reproduces experimental solvent effects on things like alanine dipeptide conformational preferences, which is something we actually care about in biology.

Tom: It’s a powerful tool because it lets you screen mechanisms and run these kinds of molecular dynamics simulations much faster than if you had to do the full quantum chemistry work every time.

Meng: For deployment, the fact that it works across different solvent types and doesn't require retraining is a huge plus for reliability at scale.

Lu: The limitation they mention is that the accuracy ultimately depends on their underlying reference data, which includes their DFT-SMD calculations and whatever gas-phase potential you combine it with.

Jane: So, while it’s very transferable, you still need a good starting point from your existing model to get the best results out of AnySolv.

Tom: It’s about establishing this general strategy for incorporating these corrections into existing MLIP workflows with AnySolv that sets the stage for routine, high-throughput investigations of solvent-dependent reactivity and molecular dynamics.

Lu: This work opens up a route toward using these powerful machine learning interatomic potentials in complex solution chemistry environments where before we were severely limited by solvent dependency.

Jane: It really bridges the gap between fast gas-phase simulations and the slow, expensive reality of full solution chemistry by making solvation effects part of the standard workflow.

Tom: We're talking about a model called "Solvating machine-learning interatomic potentials for solution-phase chemistry" that’s pretty efficient at this.

Lu: It’s a solid piece of research because it shows how to make solvent effects computable rather than just something we have to guess at during simulations.

Jane: That's what we're talking about today, and it really gives us a lot of hope for how we can simulate more complex chemical processes in the future.

Meng: I’d still want to know how fast they can get this correction into standard simulation pipelines, because that’s the real test for practical utility.

Lalam: They established a general strategy for incorporating these corrections into existing MLIP workflows with AnySolv.

Anton Z. Ni, Joonho Lee

Harvard University

physics.chem-ph, cs.LG

Submitted: 2026-10-12

Updated: 2026-10-12

Comments: 19 pages, 7 figures, 11 tables

Code: https://github.com/Ant-on-knee/anysolv

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: The gist: AnySolv is a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them, enabling

Key concepts

AnySolv
AnySolv is a method that adds solvent-dependent energy and force corrections to standard gas-phase potentials without retraining them. It functions as an equivariant message-passing graph neural network that learns solvation effects independently of the base potential, making simulations faster and more scalable.
Equivariant Message-Passing Graph Neural Network
This is the AI architecture used by AnySolv. It is a type of neural network designed to understand how molecules interact by passing messages between neighboring atoms in a structure. 'Equivariant' means the model correctly handles rotations and translations, which is crucial for accurately modeling molecular systems.
DFT-SMD Solvation Energy
This refers to the solvation energy and forces calculated using Density Functional Theory (DFT) combined with the SMD (Solvation Model Density) approach. AnySolv learns to reproduce these DFT-SMD results, meaning it can accurately predict how a molecule behaves when dissolved in a solvent based on quantum chemistry calculations.

Summary

The gist: AnySolv is a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them, enabling scalable simulations of solvent-dependent chemical reactivity and molecular dynamics.

How it works

AnySolv is an equivariant message-passing graph neural network that learns solvation energy and force corrections independently of the gas-phase potential. It predicts the difference between solvated and gas-phase DFT energies at a given molecular geometry, together with the corresponding forces obtained as the negative energy gradient. This single-point correction, defined as ∆E(R, q, M, s) = Esolv(R, q, M, s) − Egas(R, q, M), is designed to be added to a standard computational workflow using MLIPs or quantum chemistry calculations without explicit sampling of solvent configurations.

Solvent Representation and Training

The model represents solvents through continuous physical descriptors from the Minnesota Solvent Descriptor Database. The solvent descriptor vector s is defined as s = (n, α, β, γ, ε, ϕ, ψ), where these descriptors are transformed and normalized to create a zero solvent descriptor vector s˜k using a transformation Tk applied to the original descriptors. This normalized vector ˜s ∈ R7 is then mapped to a C = 128-dimensional solvent embedding by a two-layer perceptron. The model is trained on approximately 1.4 million DFT-SMD calculations in 36 solvents at the ωB97M-V/def2-TZVPD level of theory, using geometries drawn primarily from OMol25 that span charged and open-shell species as well as nonequilibrium structures along reaction pathways.

Model Architectures and Training

Two versions of AnySolv are trained: the MoE model, which is fine-tuned from the UMA-S-1.2 checkpoint, a 291.3 M-parameter mixture-of-experts (MoE) model, and the compact model, which is a dense 6.4 M-parameter model trained from scratch. The MoE model utilizes an expert mixture where the solvent embedding modulates the message weights, while the compact model is conditioned on the solvent only through node features. The loss function is a weighted sum of the per-atom mean absolute error of ∆E and the L2 norm of the force error. The training data includes 1,355,281 labels from DFT-SMD energy and force calculations with the ωB97M-V range-separated hybrid exchange–correlation functional, plus other datasets to improve coverage of systems with widely separated charges and experimental single atom labels.

Performance and Transferability

AnySolv reproduces DFT-SMD solvation energies and forces, including for solvents absent from training, while retaining the accuracy of the arXiv:2610.17226v1 [physics.chem-ph] 12 Oct 2026. On external benchmarks, AnySolv reproduces DFT-SMD solvation energies to 2–4 meV mean absolute error at more than three orders of magnitude lower cost, and it matches the accuracy of DFT-SMD against experimental solvation free energies. The MoE model achieves solvation-energy mean absolute errors (MAEs) of 0.016 and 0.020 eV, force-magnitude MAEs of 0.026 and 0.030 eV/Å, and mean force cosine similarities of 0.97 and 1 in the test set.

Applications

The results enable screening of the mechanisms of 120 nucleophilic aromatic substitution reactions in six phases as well as molecular dynamics simulations that reproduce the experimentally observed solvent-dependent conformational preferences of alanine dipeptide. For example, the MoE model reproduces every solvated barrier to within 0.9 kcal/mol and the reaction free energies to within 1.4 kcal/mol for the Menshutkin reaction in six solvents. The model also successfully reproduced trends in SNAr mechanism screening across many substrates and solvents. Finally, the conformational distribution of alanine dipeptide was reproduced by well-tempered metadynamics on ϕ and ψ in chloroform, methanol, and water. AnySolv can be used as an implicit solvent term in molecular dynamics wherever ab initio molecular dynamics with SMD solvation is appropriate.

Conclusion

AnySolv establishes a general strategy for incorporating transferable solvation corrections into existing MLIP workflows by providing corrections that can be added to any gas-phase potential and remain valid for charged and open-shell solutes at geometries far from equilibrium. This work opens a route toward routine, high-throughput investigations of solvent-dependent reactivity and molecular dynamics. The accuracy of AnySolv remains limited by the underlying SMD reference and the gas-phase potential with which it is combined.

--- Page 1 ---

The gist: AnySolv is a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them, enabling scalable simulations of solvent-dependent chemical reactivity and molecular dynamics.

Performance and Transferability

AnySolv reproduces DFT-SMD solvation energies and forces, including for solvents absent from training, while retaining the accuracy of the arXiv:2610.17226v1 [physics.chem-ph] 12 Oct 2026. On external benchmarks, AnySolv reproduces DFT-SMD solvation energies to 2–4 meV mean absolute error at more than three orders of magnitude lower cost, and it matches the accuracy of DFT-SMD against experimental solvation free energies. The MoE model achieves solvation-energy mean absolute errors (MAEs) of 0.016 and 0.020 eV, force-magnitude MAEs of 0.026 and 0.030 eV/Å, and mean force cosine similarities of 97% and 96% in the test set.

Applications

The results enable screening of the mechanisms of 120 nucleophilic aromatic substitution reactions in six phases as well as molecular dynamics simulations that reproduce the experimentally observed solvent-dependent conformational preferences of alanine dipeptide. For example, the MoE model reproduces every solvated barrier to within 0.

Improvements for AI systems

  1. A transferable implicit solvation potential can be added to existing gas-phase machine-learning interatomic potentials without retraining them, enabling scalable simulations of solvent-dependent chemical reactivity and molecular dynamics.

  2. The improved system can reproduce DFT-SMD solvation energies and forces with a cost more than three orders of magnitude lower, making previously impractical simulations feasible.

  3. AnySolv supports 179 solvents, including those absent from training, by representing them through continuous physical descriptors, which allows for the generalization of the model to unseen solvents.

  4. The system can perform high-throughput screening of chemical reactions, as demonstrated by reproducing trends in a mechanistic screen of 120 nucleophilic aromatic substitution reactions at nearDFT quality.

  5. The system can reproduce experimental solvent effects on molecular dynamics, such as the conformational preferences of alanine dipeptide, by providing accurate solvation free energies that match DFT-SMD against experimental values.

Abstract

Most machine-learning interatomic potentials (MLIPs) are trained on gas-phase data, which limits their use for chemistry in solution. Here, we introduce AnySolv, a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them. It is trained on 1.4 million solvation energies and forces computed with density functional theory (DFT) and SMD in 36 solvents, spanning charged, open-shell, and transition-state structures. Because solvents are represented by continuous physical descriptors, AnySolv supports 179 solvents, including those absent from training. On external benchmarks, AnySolv reproduces DFT-SMD solvation energies to 2-4 meV mean absolute error at more than three orders of magnitude lower cost, and it matches the accuracy of DFT-SMD against experimental solvation free energies. This enables screening of the mechanisms of 120 nucleophilic aromatic substitution reactions in six phases as well as molecular dynamics simulations that reproduce the experimentally observed solvent-dependent conformational preferences of alanine dipeptide. By providing transferable solvation corrections to existing MLIPs, AnySolv enables scalable simulations of solvent-dependent chemical reactivity and molecular dynamics.

Sources

Related papers