Relighting Test for Electrode Potentials: Keeping Field Effects Out of Learned Desolvation Barriers

Before a lithium or zinc ion can plate onto a battery electrode, it has to shed the shell of solvent molecules clinging to it. The energy cost of this step, the desolvation barrier, often limits fast charging and low-temperature operation. Computing it properly requires simulating the electrode–electrolyte interface while a voltage is applied, and the voltage changes during charging. Machine-learned interatomic potentials (neural networks that predict atomic energies and forces at near-quantum accuracy but far lower cost) could make these simulations affordable. The risk is that a network trained at one voltage quietly absorbs that voltage's electric field into its description of local chemistry. It then reproduces the training conditions and gives wrong barriers at any other voltage.

Computer graphics has a name for this failure. In 3D reconstruction, a model fit to photos can paint the room's highlights and shadows directly into the object's surface color ("baked-in lighting"), so the object looks wrong when relit. That field's fixes are an explicit physical split between surface and light, plus observing the same object under several lightings. We carry the second fix over. We propose relabeling identical atomic snapshots with quantum calculations at several electrode potentials. The energy differences between potentials for the same geometry must then be explained entirely by an explicit electrostatic term, which leaves the short-range network no room to absorb the field. The success criterion is a "relighting test": the model must predict desolvation free-energy barriers at an electrode potential it never saw.

Core Claim

Hybrid interfacial potentials already split energy into a local learned part and a long-range electrostatic part. The split is only as clean as the long-range physics is complete, and the training signal does not force it to be clean. Any field-dependent energy the explicit term misses is fit by the local network as a function of geometry. That is baked-in lighting in atomistic form.

Suppose each geometry $R$ is labeled by density functional theory (DFT) at two or more electrode potentials $U$ and $U'$. Then the difference $\Delta E_{\text{DFT}}(R;U,U') = E_{\text{DFT}}(R;U) - E_{\text{DFT}}(R;U')$ contains no short-range chemistry at all. We add the loss term

$$\mathcal{L}_{\text{relight}} = \sum_{R}\sum_{U,U'} \big[\Delta E_{\text{DFT}}(R;U,U') - \Delta E_{\text{lr}}(R;U,U')\big]^2,$$

where $\Delta E_{\text{lr}}$ is the corresponding difference predicted by the long-range electrostatic model. This supervises the field response directly. It also makes any residual field dependence visible as a measurable error instead of letting it hide inside the local network.

Our prediction is as follows. Models trained with this term, which we call "multi-potential relabeling," will transfer desolvation barriers to held-out potentials close to the accuracy of a model trained at the target potential. Models trained without it, even on data from several potentials, will not.

Relation to Prior Work

The closest work, and the method we build on, is ec-MLP. It pairs a local network with a "Deep Wannier" model of electrolyte polarization and a constant-potential polarizable electrode. It reaches $10^3$–$10^5\times$ speedups over quantum molecular dynamics and reproduces a differential capacitance peak near 80 µF/cm² on Pt(111)/KF(aq). It already performs the physical split we want.

Its own stated simplifications are exactly where leakage can enter:

  • charges attached to the Wannier centers are fixed;
  • the dependence of electrode charges on atomic positions is dropped when computing forces.

It also never computes desolvation barriers or tests transfer across potentials. We do not propose a new architecture. We propose a training signal and an evaluation that expose whether ec-MLP-style splits actually hold.

Two papers motivate the need: - Deriving effective electrode–ion interactions shows that classical force fields disagree qualitatively at an Au(111)–water interface. It uses the short-cutoff foundation model UMA as its reference, and UMA has no explicit field. This tells us that a local foundation model is exactly the kind of model that can bake the field in. - Toward a quantitative interfacial description of solvation shows that weakly binding ether solvents and a degree of ion pairing short of saturation control low-temperature performance. This is why barrier differences of a few kcal/mol between electrolytes matter.

On the graphics side, PhySG (an explicit rendering layer with an analytic light basis) and MATLABER (a learned prior constraining material maps) prevent baking through model structure. Cascaded schedules in CasTex prevent it through optimization order. We borrow the multi-illumination idea (same object, several lightings) rather than an analytic basis. In graphics this is the fix that removes the degeneracy between surface and light instead of merely regularizing it.

Finally, 4D-Decoupling supplies a way to run free-energy calculations with neural potentials without changing their architecture. It was validated only for neutral solutes in water. We use it for the bulk-solvation reference leg of charged ions.

Research Program

Testbeds. The first testbed extends the Pt(111)/KF(aq) interface from ec-MLP by adding a single Li⁺ or Zn²⁺ ion. We compute the free energy along two coordinates: the ion's distance from the surface and its first-shell coordination number. The run stops before electron transfer, since fixed oxidation states rule out modeling plating itself. The second testbed places LiFSI in DME versus DEE, with BTFE as a diluent, against the same inert metal electrode. This lets us check whether the model preserves the experimentally important ranking of solvents that bind weakly or strongly.

Data generation. For each system we generate active-learning snapshots at a training set of potentials. Every snapshot is relabeled by DFT single-point calculations at every potential in that set. Single points are far cheaper than new molecular dynamics, and we vary how many snapshots receive multi-potential labels to map cost against transfer accuracy.

Training. We train three variants:

  • ec-MLP-style models with the standard loss;
  • the same models with $\mathcal{L}_{\text{relight}}$ added;
  • the same models with a cascade: fit the long-range term on difference data first, freeze it, then fit the local network.

Evaluation. We run umbrella sampling to obtain desolvation profiles at training potentials and at one interpolated and one extrapolated held-out potential. We also track total cost in GPU-hours and DFT core-hours, so the accuracy-efficiency trade-off is reported explicitly rather than assumed.

Baselines

Naive reference. A purely local equivariant potential (MACE-class, or DeePMD) trained on the pooled data from all training potentials, with no field information. It should fail at held-out potentials. How badly it fails calibrates the size of the effect.

Field-conditioned local model. The same local network with the electrode potential, or the surface charge density, supplied as an extra input. It is trained on identical data, including the multi-potential relabels, with the same hyperparameter-search budget. This is the cheapest serious competitor. To beat it, our model must be clearly better at the extrapolated potential; matching it only when interpolating is not enough.

Strongest existing method. ec-MLP trained on the same snapshots and potentials, without $\mathcal{L}_{\text{relight}}$, using its published protocol and the same compute. Its reported fitting fidelity on Pt(111)/KF(aq) is about 1.3 meV/atom for energies and about 75 meV/Å for forces (ec-MLP). We found no reported result on Lacuna for desolvation barriers from ec-MLP, or from any potential, at held-out electrode potentials.

Ablation. The key ingredient is the relabeling itself, so we isolate it in two ways. One arm uses the same number of extra DFT calculations spent on new geometries instead of on relabels. Another arm uses relabels but no difference loss, with the extra labels simply added as ordinary energy targets.

Oracle and success threshold. The oracle is the same architecture trained directly on data at the held-out potential. It is backed by DFT force spot-checks on snapshots from the barrier region. We count a result as beating the baselines if our barrier error at the extrapolated potential is within roughly 1 kcal/mol of the oracle while each baseline's error is at least twice ours. The 1 kcal/mol figure is chosen to match the sub-kcal/mol precision that neural free-energy methods reach in bulk (4D-Decoupling).

Why not just…?

Why not just condition a local network on the electrode potential and train on many potentials? This is a real threat. For interpolation it may work as well as our method. It has no reason to be correct beyond the training range, and a single scalar potential does not describe how charge is distributed across a particular surface configuration. The extrapolated-potential test, together with a test on a different electrode facet at matched potential, settles the question. If the conditioned model passes both, we will report that the explicit long-range term is unnecessary for barrier transfer.

Why not just use ec-MLP as published? Its field term may already be complete enough, in which case there is no leakage to remove. We test this directly with a leakage measure: the spread of short-range energies on identical snapshots across models trained at different potentials. If this spread is below DFT noise and ec-MLP transfers barriers well, the contribution reduces to a validated benchmark, and we will say so.

Why not tune a classical or polarizable force field against the learned model, as in the UMA study? That study had to hand-tune Lennard-Jones parameters for each ion and surface just to recover adsorption profiles such as a −8.40 kcal/mol minimum for Cl⁻ (Deriving effective electrode–ion interactions). Such parameters carry no response to the field. We include a tuned classical model as a cheap, fast reference point. We expect it to preserve barriers at the tuning potential and drift away from them at other potentials.

Why not run quantum molecular dynamics directly? Its roughly 20 ps horizon (ec-MLP) is too short to converge desolvation free energies. We use it only for spot checks.

Decisive Checks

Three results decide the proposal.

The first is barrier transfer. For each method we report the barrier error at the extrapolated held-out potential, relative to the oracle and to DFT spot-checks. The relabeling loss must close most of the gap that separates ec-MLP from the oracle.

The second is the leakage measure. It should fall with relabeling and track barrier error across all variants. If leakage and barrier error do not move together, the baked-in-field story is wrong.

The third is ordering. The model must preserve the DEE-below-DME barrier ordering at every potential. That ordering is the experimentally meaningful signal behind the 98.9% Coulombic efficiency at −60 °C reported in Toward a quantitative interfacial description of solvation.

Across all three, we plot DFT relabel budget against transfer error. This shows how much quantum computation the relighting guarantee actually costs.