materials-chemistry/ai produced the result/ACS Nano 2025 · v2
Simulated water in graphene gaps shows acid's charge hugging the surface
Researchers trained a machine learning model to imitate quantum chemistry calculations, then used it to simulate water squeezed between graphene sheets. The simulations tracked where a single positive or negative charge in the water prefers to sit.
spectrum · one line per step, placed by what the step does · bright lines used AI
Protons Accumulate at the Graphene–Water Interface
ACS Nano, 2025
doi:10.1021/acsnano.5c02053 · record aix-00078 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Graph neural network
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Water is not a passive liquid. It is a shifting network of molecules linked by hydrogen bonds, weak attractions between a hydrogen on one molecule and an oxygen on another. In acid or alkali, that network also carries charge: a hydronium ion is a water molecule with an extra proton, a hydroxide ion is one missing a proton. These charges do not travel as fixed lumps. They hop along the network as bonds break and re-form, which makes them awkward to model. Any simulation that follows them must describe chemical bonds as they come apart, not just molecules bouncing off one another.
The question here concerned what happens when such water is trapped in a very narrow gap. Graphene is a sheet of carbon one atom thick, and two sheets can be held a short distance apart, leaving room for only a handful of water layers. The researchers set out to find where a hydronium or a hydroxide ion sits in such a slit: pressed up against the carbon surface, or out in the water between the sheets. They varied the gap width and the temperature.
Where AI came in
The accurate way to describe bond-breaking in water is density functional theory, a quantum mechanical calculation of how electrons arrange themselves. It is accurate but slow, which limits how many atoms can be followed and for how long. The researchers trained a model called MACE, a neural network that treats atoms as points in a graph exchanging messages with their neighbours, on energies and forces computed by density functional theory. The training set was built up in stages, adding configurations where the model was uncertain, a method known as active learning. Validation errors against held-out data were 0.7 meV per atom for energies and 17.2 meV per ångström for forces.
The trained model then stood in for the quantum calculation inside the molecular dynamics, the step-by-step tracking of every atom's motion. That substitution is what allowed the simulations to run long enough and over enough gap widths to gather statistics: five slit widths, both ion types, five independent runs each, more than 200 nanoseconds of simulated time in total. Every density profile, free energy curve and hydrogen-bond count in the paper comes from trajectories the model drove. Setting up the reference calculations, counting hydrogen bonds, fitting the temperature series and the separate electron-density analysis of individual snapshots were done without a learned model.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors trained a MACE machine learning potential on revPBE-D3 density functional theory energies and forces and used it to run reactive molecular dynamics of water containing a single hydronium or hydroxide ion confined between free-standing graphene sheets. Simulations covered five slit widths with average values of about 6.5, 9.2, 12.2, 14.7 and 19.7 Å, with five independent 4 ns runs per slit width and species, giving over 200 ns of trajectory. Density and free energy profiles show the hydronium ion residing mainly in the first contact layer of water, while the hydroxide ion is found both near the surface and in layers farther from it. A temperature series from 300 to 400 K for the three-layer system gives ΔH = −2.4 ± 0.2 kcal/mol and ΔS = −3.0 ± 0.6 cal/mol/K for hydronium adsorption and ΔH = 1.2 ± 0.3 kcal/mol and ΔS = 3.1 ± 0.8 cal/mol/K for hydroxide, and density functional theory analysis of snapshots shows charge rearrangement limited to the contact layers.
How AI was used
A MACE equivariant message-passing potential was developed as a surrogate for the density functional theory potential energy surface, so that reactive molecular dynamics covering bond breaking and formation could be run at length and time scales beyond ab initio molecular dynamics. The model used two layers, a 6 Å cutoff per layer, 128 equivariant messages and maximal message equivariance L = 1, giving a 12 Å receptive field, and was fitted to revPBE-D3 energies and atomic forces computed in CP2K/Quickstep with GTH pseudopotentials, TZV2P basis sets for oxygen and hydrogen, DZVP for carbon and a 1200 Ry density cutoff. The training set was built over five generations, starting from structures from previous work and extended by two rounds of active learning, path integral molecular dynamics configurations, additional slit widths and neutral frames containing a hydronium–hydroxide pair, giving a final set of 3378 structures of which 1303 involve graphene interfaces and 2075 bulk conditions. The trained potential then drove NVT molecular dynamics in ASE with a 0.5 fs time step and a Langevin thermostat, for five slit widths and both defect species at 300 K, for a 3L temperature series from 300 to 400 K, and for umbrella sampling runs used to compare the free energy profiles. Trajectory post-processing, hydrogen-bond counting, the linear fit of ΔF = ΔH − TΔS, and the separate density functional theory electron density difference and Bader charge analyses of selected snapshots did not involve a learned model.
The shape of the work
Structural · the record, drawn
no AI
Generate DFT reference energies and forces
Numerical or physics simulation, including where a learned surrogate replaces it.
we train our MLP model using energies and atomic forces obtained from DFT calculations using the CP2K/Quickstep codewhere the paper describes this · verbatim
AI
Train MACE potential with active learning
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation. Its result feeds back into an earlier step.
In the fourth generation, we conducted an additional round of active learning to refine the model based on the conditions sampled thus far.where the paper describes this · verbatim
AI
Run production MD for five slit widths and both defects
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
For each of the five slit widths and two species, we conducted five independent simulations.where the paper describes this · verbatim
AI
Run temperature series for the 3L system
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
we observe how these ions behave within the 3L system featuring an intermediate region, across temperatures from 300 to 400 Kwhere the paper describes this · verbatim
no AI
Derive density, free energy and hydrogen-bond statistics
Extracting understanding from model behaviour.
Using the simulated density profiles, we quantified the surface affinity of hydronium and hydroxide ions by examining their free energy profiles.where the paper describes this · verbatim
AI
Check free energy profiles with umbrella sampling
Testing outputs against ground truth. The AI stood in for simulation.
we conducted additional MLP-based biased simulations using umbrella sampling to compare the free energy profiles reportedwhere the paper describes this · verbatim
no AI
Analyse interfacial electronic structure with DFT
Numerical or physics simulation, including where a learned surrogate replaces it.
we used DFT to analyze their electron density difference (Δρ)where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
All reported density profiles, free energy profiles, hydrogen-bonding statistics and thermodynamic decompositions come from molecular dynamics driven by the machine learning potential the authors trained; the conclusions have no non-AI source in this paper.
we use the MACE architecture, which allows for fast and highly data-efficient training with high-order equivariant message passingwhere the paper describes this · verbatim
The final energy and force validation root-mean-square errors were 0.7 meV/atom and 17.2 meV/Å, respectively.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of MACE machine learning potential (revPBE-D3 reference)Which version of the model was used is not stated.
About this article
Record aix-00078, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error