WorksPhysics and Astronomy
Bias Conditioning and Long-Range Electrostatics for Consistent Machine-Learning Potentials of Electrified Interfaces
Abstract
Machine-learning interatomic potentials (MLIPs) of electrified interfaces need the bias as input, but it is unclear whether a scalar electrode charge suffices and whether short-range models capture the non-local response. We add charge and field conditioning to short- and long-range MLIPs. State-of-the-art field-conditioned short-range models overpolarize thick water films nearly twofold and miss work-function shifts at stepped Pt. Our long-range model captures both, needs only the electrode charge at metals as input, and reproduces ab initio simulations of the Pt(111)–water interface.
A key requirement for modeling electrochemical interfaces is the ability to apply an electrical bias. The applied bias controls the structure and dynamics of the interfacial electrolyte [1], the charge distribution at the electrode [2], and the rate of electron transfer [3]. Ab initio simulations impose this bias either through an excess electrode charge balanced by a countercharge (jellium, a charged plate, or a layer of inert atoms) [4, 5, 6, 7] or through an applied field with a dipole correction [8]. Both routes fix the same boundary condition, a displacement field normal to the surface [9, 10], and thus sample the constant-charge ensemble. To simulate at constant potential instead of constant charge, a potentiostat adjusts the charge over time so that the electrode potential fluctuates around a target value [11, 12, 13, 14]. In either ensemble, however, ab initio molecular dynamics (AIMD) is too expensive to converge even the potential of zero charge [15]. Machine-learning interatomic potentials (MLIPs) remove this cost barrier and are now being extended to biased interfaces. To serve both ensembles, a biased MLIP must take the electrode charge, i.e., , as input and return the electrode potential, the charge derivative of its energy, as output.
Existing MLIP designs differ in how the bias enters the model [16]. A common approach is to embed the electrode charge or electron number as a scalar [17, 18, 19, 20], following the total-charge inputs established for molecules [21, 22, 23, 24]. Field-aware models instead couple an applied vector field to equivariant features [25, 26, 27, 28, 29]. Whether a scalar input suffices or the field direction is required has not been established. Independent of the input, the response to a bias is non-local. The electrode potential or work function depends on the orientation of every solvent molecule in the cell, not only of those close to the surface. The solvent orientation, in turn, is limited by its own depolarizing field. Short-range MLIPs, including most of the models above, miss such long-range contributions: they overpolarize water [25, 30, 31], but the resulting errors remain hidden in energy and force metrics and appear only in sampled observables [32, 33, 34]. The few long-range MLIPs for biased interfaces [35, 36] learn only the electrolyte charges and treat the metal with classical Gaussian charges [37], which miss the electron spillover into the interfacial water found in ab initio simulations [2]; learning the electronic response of the metal instead predicts the polarization more accurately [38]. Two questions therefore remain open: in which form should the bias enter an MLIP, and which interaction range does a consistent response require?
In this Letter, we address both questions by implementing charge and field conditioning in the short-range graph atomic cluster expansion (GRACE) [39] backbone and in our architecture for learning long-range representations with equivariant messages (LOREM) [40]. LOREM learns latent charges and feeds their Ewald-summed Coulomb potentials back into the descriptor features as long-range messages, so that the predicted properties depend on the electrostatic potential of the entire cell. To show that long-range interactions are required for a correct bias response, we first consider water between two inert electrodes, a system that we can validate against theoretical predictions. We then turn to metal–water interfaces and compare scalar and vector conditioning in short- and long-range models for a single water molecule above flat and stepped Pt surfaces. Finally, constant-charge molecular dynamics (MD) of the Pt(111)–water interface yields the bias-dependent water structure and electrode potential. Together, these systems show that long-range electrostatics is required to describe the bias response beyond the cutoff radius, and that a long-range model needs only a scalar input at metal–water interfaces.
Long-range interactions under bias.
A capacitor consisting of water between two inert electrodes is a simple model system to probe the effect of long-range electrostatics under bias. The bias can be introduced by charging the electrodes with opposite charges , which fixes the displacement field at , where is the lateral area and the surface normal. Here, we realize the same boundary condition by applying an electric field together with a dipole correction, so that . Note that is the field in the empty cell; the macroscopic field inside the water, , is screened by the water polarization . As shown in Sec. S1A of the Supplemental Material, the energy of this system can be written as
where is the energy at zero displacement field and the cell dipole moment. The bias response of the energy is thus given by the cell dipole in , . The capacitor is symmetric under , so its geometry does not reveal the field direction. An MLIP must therefore receive this direction explicitly to learn the bias-dependent energy. We couple the applied field to the equivariant features of the model, similar to Ref. 28 (see Sec. S1B of the Supplemental Material for the implementation). With obtained by automatic differentiation of , the model represents the cell dipole as a sum of atomic dipoles, . To isolate the effect of long-range electrostatics, we implement field conditioning in LOREM and compare
its long-range and short-range variants. The long-range model uses an Ewald sum with latent charges; the short-range model has a 5 Å cutoff with one message-passing step, leaving it with an effective cutoff of at most 10 Å. We train these models on the energy , forces , and cell dipole as learning targets. The training set consists of Ne–water–Ne structures with a 22 Å water film (Figure 1a). We sample frames from a zero-field AIMD trajectory from Ref. 33, label them using single-point calculations at V/Å, and add two rounds of active learning. Each round adds a few hundred structures sampled from MD trajectories driven by the models of the previous round (see Sec. S1C of the Supplemental Material for dataset and training details).
Figure 1. MD of water between Ne electrodes at fixed displacement field. (a) Cells with the labeled 22 Å and the unseen 31 Å water film. (b,c) Distribution of at for the short-range (sr) and long-range (lr) models over three active-learning rounds, against AIMD (gray) and a Gaussian predicted by continuum electrostatics (black, dashed). (d,e) from molecular dynamics at V/Å; dashed line: expected response based on the AIMD fluctuations, Eq. (2). (f,g) The round-2 models for the 31 Å film, which is not in the training set: (f) water orientation with the angle between the water bisector and the -axis, at a field of V/Å; the shading marks the water at least 10 Å from both Ne layers; (g) against ; dashed line: the continuum value . Error bars and bands are block estimates.
In zero-field MD, the short-range models show substantially larger fluctuations in than the long-range models (Figure 1b,Figure 1c), as reported previously [25, 30, 31]. Active learning improves the agreement of the short-range models with AIMD, but they never reach full agreement for the training set sizes tested here. In contrast, the long-range models reproduce the AIMD result with little additional data.
The field response of is related to the variance of its fluctuations (see Sec. S1D of the Supplemental Material for the derivation),
where is the electronic response at fixed nuclear positions. Evaluating this first term by finite differences on a few snapshots labeled at V/Å and obtaining the variance from the zero-field AIMD trajectory yields a reference for . Figure 1d,Figure 1e compares this slope (dashed black line) with finite-field MD. The short-range models deviate from the reference even after two active-learning rounds, consistent with their larger variance in Figure 1b,Figure 1c. The long-range model, in contrast, matches the reference after a single round.
The necessity of long-range interactions becomes even more apparent when extrapolating to a thicker, unseen 31 Å water film (Figure 1a, right). Figure 1f shows the water orientation at 0.1 V/Å in the trajectories obtained with the short-range and long-range round-2 models. While the short-range model behaves similarly to the long-range model near the electrodes, it strongly overpolarizes the water in the center of the film, where the interfaces are outside the model’s effective cutoff. The long-range model extrapolates well: in Figure 1g, its response lies close to the reference (black dashed line), whereas the short-range model overestimates the response nearly twofold. In the absence of an AIMD trajectory at this film thickness, we estimate the reference from the 22 Å cell, adding the response of the additional bulk water with thickness Å (see Sec. S1E of the Supplemental Material for details). Field-conditioned long-range models thus correctly describe the polarization of water under bias, whereas field-conditioned short-range models fail to do so.
Conditioning for metal–water interfaces. The metal–water–countercharge setups used in computational electrocatalysis are essentially the same constant-D problem as the capacitor model considered above, with one inert electrode replaced by a metal slab. However, two things change. First, the bias charge sets the number of electrons, so the bias response is the work function , where is in units of the elementary charge . Second, the electronic charge is free to move and can, for example, spill over into the electrolyte [2]. The position of the interfacial charge therefore depends on the interfacial structure. As shown in Sec. S1F of the Supplemental Material, at any charge,
where is the work function of a reference system such as the bare metal slab and is the change in dipole moment relative to this reference. is the distance between the centroid of the countercharge and the centroid of the excess charge on the metal and water atoms. Note that contains the electronic rearrangement of the metal surface as well as that of the water. The potential energy surface is then given exactly by
by
with . Assuming that depends only linearly on and that is constant, we recover the quadratic expansion [29]
with the zero-charge work function and capacitance
with . Compared to the capacitor model from Eq. (1), the metal–water–countercharge system introduces a linear term and a quadratic term , both of which are properties of the data and must be learned by the MLIP. The assumptions of constant and linear imply that the capacitance is constant. This holds only in a narrow bias window around the value of at which the water film has been equilibrated [41]. Outside that window, the electron excess or deficit is deposited in states belonging to the water, so that changes and is no longer constant.
Many existing strategies [17, 18, 19] condition an -symmetric model on a scalar to learn . These models cannot resolve the field direction for water molecules with the surface outside their effective cutoff, and therefore cannot give an orientation-dependent contribution to the bias response. While field conditioning can address the lack of directionality [29], the limitations of this approach near rough surfaces with an inhomogeneous field have not been tested. We now investigate the limitations of short-range models using our own implementation in the GRACE backbone, and show that the charge-conditioned long-range LOREM backbone overcomes them. Previous charge-conditioning models concatenate to the descriptor features [17, 18]; we instead apply a feature-wise linear modulation (FiLM) [42] layer, a more general way to condition a neural network on a global variable. The work function and the surface-normal components of the Born effective charges, , then follow by automatic differentiation. We denote the resulting models GRACE- and LOREM- (see Sec. S2A of the Supplemental Material for implementation details). To complete the comparison, we use the field-conditioned response analysis in -orientation (RAZOR) approach [29], which learns the zero-charge work function of Eq. (5) from latent atomic dipoles and the capacitance as a constant. While RAZOR originally used two backbones, our implementation, denoted GRACE-RAZOR, uses a single GRACE backbone (see Sec. S2B of the Supplemental Material for implementation details).
We train the three models (GRACE-RAZOR, GRACE-, and LOREM-) on a single water molecule above a Pt surface. We consider two surfaces: flat Pt(111) (Figure 2a) and stepped Pt(332) (Figure 2b). To teach the models the charge response, the data are labeled at different charges in a small window around (), using a plate 13 Å above the surface as countercharge. We train with energy, forces, and work function as learning targets (see Sec. S2C of the Supplemental Material for dataset and training details).

Figure 2. Work function of one water molecule above flat and stepped Pt. (a) Water at height above Pt(111), its dipole rotated by ; (b) water moved by across a step of Pt(332). Circles mark the cutoff . (c,e) against on Pt(111) at and 7 Å. (d,f) , relative to the mean, against on Pt(332); at Å GRACE- lies off scale (0.67 V).
For Pt(111), all models predict the correct work function when the water is inside the cutoff (Figure 2c). With the water beyond the cutoff (Figure 2e), the charge-conditioned GRACE- model fails to reproduce the response. The field-conditioned GRACE-RAZOR model still gives the correct response for the flat surface, even with the water outside the cutoff. Its field input carries the direction that the water molecule cannot infer from the geometry. For the stepped surface (Figure 2d,Figure 2f), however, GRACE-RAZOR fails to reproduce the response when the water molecule is moved across the step in the direction. A homogeneous applied field cannot describe the locally inhomogeneous field near the step. At Å, where the step changes by about 30 mV, the work-function error of GRACE-RAZOR along the curve is five times that of LOREM- (16 vs. 3 mV). In contrast, our charge-conditioned long-range model LOREM- follows the reference in all cases. Hence, charge-conditioned short-range models are limited by their effective cutoff, and field-conditioned ones by the homogeneous field near rough surfaces; long-range electrostatics lifts both limitations. For the long-range model, a scalar input suffices, because the model infers the surface normal from the atomic structure alone.
Charge-dependent properties.— To test the models over a wider charge range, we perform MD of the Pt(111)–water interface with GRACE-RAZOR and LOREM-q. Both models are trained on data from Ref. [41], with energy, forces, work function, and surface-normal Born effective charges as targets. A snapshot of the system is shown in Figure 3a. We compare with two references: our own AIMD (sampling time ps for each charge), and GRACE models trained at a single charge each (GRACE single-). For the latter reference, we generated additional training data by an active-learning procedure (see Sec. S3E of the Supplemental Material for details). Both references were obtained with the same computational settings as the literature data. We leave out e in training our MLIPs, as these data are prone to dielectric breakdown [41]. As the work-function label, we use the work function of the solvated jellium method [5], which differs from the charge derivative of the energy labels by about 0.3 V (see Sec. S3B of the Supplemental Material for a comparison of both labels).
Figure 3. Constant-charge MD of Pt(111)–water. (a) Simulation snapshot of a Pt(111) slab with 24 water molecules and a solvated jellium countercharge [5]. (b–d) Oxygen density and (e–g) orientation profile against the height above the top Pt layer, at , 0, and . (h) Mean work function or electrode potential against (top axis: charge density ); its slope is the inverse capacitance. The work function of GRACE single- is evaluated with LOREM- on its trajectories. (i) Mean number of chemisorbed water molecules , counted as O atoms less than 2.6 Å above the top Pt layer.
Both GRACE-RAZOR and LOREM-q closely follow the water density and orientation of the references (Figure 3b–Figure 3g). The agreement between the charge- and field-conditioned models is expected from the simple model system (Figure 2). The surface is flat, and all water in the 10 Å film has the electrode surface within the effective cutoff of both models (at most 12 Å for two-layer GRACE). Both models also reproduce the AIMD electrode potential and the number of chemisorbed water molecules from to e (Figure 3h,Figure 3i).
Conclusion.— We have shown that a consistent MLIP of an electrified interface requires two ingredients: the bias as an input and long-range electrostatics. Conditioning alone is not sufficient: across all test cases, from water between identical Ne electrodes to water above flat and stepped Pt, only the long-range model responds correctly and consistently. At metal electrodes, a scalar electrode charge is a sufficient input for this model. Field-conditioned short-range models, the current state of the art [28, 29], remain limited by their effective cutoff. The limitation also affects properties at the electrode surface: in constant-potential simulations, an incorrect electrode potential drives the potentiostat toward the wrong charge. Our long-range model instead obtains the electrode potential from the same energy surface as the forces and can therefore be used directly in constant-potential simulations. Finally, our long-range models learn the electrostatic response without imposing its known functional form, Eq. (4). Building such physical constraints into the architecture is a natural next step toward simulations of electrolytes with solvated ions and is expected to further improve the transferability to larger and more realistic systems.
Acknowledgments.—P.L. and P.R. acknowledge funding from the German Research Foundation (DFG) under Germany’s Excellence Strategy – EXC 2089/2 – 390776260. L.d.K. and C.J.S. acknowledge funding from the European Union; this publication is part of the HeliEICat project that has received funding from the European Research Council (ERC) under the European Union’s Horizon research and innovation program (Grant agreement No. 101217879). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. The authors gratefully acknowledge the scientific support and HPC resources provided by the Leibniz Supercomputing Centre (LRZ) and by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under NHR projects b311bb and b314bb. The authors would like to thank Marcel F. Langer, Nicolas Bergmann, Henrik Stooß, Alexander Schlaich and Georg K. H. Madsen for insightful discussions.
Data availability.— The LOREM models have been trained using a branch of the public repository available at https://github.com/lucasdekam/lorem-jax/tree/film-field-conditioning. The code for training bias-conditioned GRACE models is available at https://github.com/lucasdekam/grace-tensorpotential/tree/charge-conditioning. The modified version of Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) with the GRACE interface to enable constant-charge simulations is available at https://github.com/lucasdekam/lammps/tree/grace. Further datasets, models, and analysis code to reproduce these findings will be made available upon publication in an open online data repository.
Water between Ne electrodes
Energy of the Ne–water–Ne capacitor under applied field
We label the Ne–water–Ne cell under an external field with a slab dipole correction. The cell is periodic in and , has length along and a volume . The displacement field is throughout the cell. At fixed , the energy of a periodic, insulating cell is [8]
where is the energy under ordinary periodic boundary conditions. The polarization is related to the dipole moment as . With and along , expanding the square gives
The first two terms form the energy at zero displacement field; is the dipole-correction energy [43]. The last term is the field energy of the empty cell; this term is left out of the energy labels. It does not depend on the atomic positions and therefore does not contribute to the forces. Leaving it out, as in the finite-field density-functional theory (DFT) energy below, makes equal to the dipole of the atoms and electrons, without the empty-cell term .
and also depend on the electronic state . A finite-field DFT calculation returns
with the minimizing state. Because is stationary, its dependence on does not enter the first derivative of , so that
(Hellmann–Feynman theorem). The DFT labels therefore satisfy
where both and are evaluated at and thus depend on the field through the electronic polarization.
Field-conditioned LOREM models
We implemented field conditioning in LOREM [40] following MACE-Field [28]. We contract the tensorial features with the applied field in a weighted tensor product,
where is the spherical harmonic representation of the applied field . The energy is constructed from the last-layer scalar features as . Because the cell dipole is obtained from automatic differentiation, the model effectively represents it as a sum over atomic dipole contributions, .
For a consistent comparison, the field-conditioned long-range and short-range models are the same LOREM model and differ only in the long-range term. Both use 64 scalar and 8 spherical features up to , 16 radial basis functions, a 5 Å cutoff and one message-passing step. The long-range model uses an Ewald sum over the latent charges (, scalar only) with a slab correction along , which is the boundary condition of the reference calculations. The short-range model keeps only its 10 Å effective cutoff. We train on energy, forces and with loss weights 1000 : 1 : 0.5, using the Muon optimizer [44] with a learning rate of , 20 warm-up epochs, cosine decay, gradient clipping at 1.0, batches of 7 structures and 400 epochs. We keep the model from the training step with the highest summed of energy, forces and dipole on the validation set (early stopping). We do not train on the Born effective charge here.
Training and simulation
Structures were obtained from an ab initio molecular dynamics (AIMD) trajectory from Ref. 33. The cell measures Å and holds 64 between two Ne layers, 216 atoms in total, with about 22 Å of water. It is periodic in and and non-periodic along . To label the data we run DFT calculations with VASP 6.4.3 [45, 46, 47, 48, 49] in a finite field with dipole correction. Other than that, the settings are the same as the reference AIMD [33]: RPBE [50] with D3 dispersion [51] (zero damping), a 500 eV plane-wave cutoff, a Monkhorst–Pack [52] grid, 0.1 eV Fermi smearing and an electronic convergence of eV. To avoid dielectric breakdown we discard labels with a Kohn–Sham gap below 1.5 eV or a smearing entropy above 1 meV.
To generate the dataset, we first sample 1146 structures from the zero-field trajectory of Ref. 33. Of these, 712 enter the training set with their zero-field labels only. The remaining 434 are labeled under an applied field: 302 for training, 44 for validation and 88 for testing. Training and validation frames are labeled with a pair of positive and negative fields, with drawn uniformly from V/Å. Test frames get and V/Å. This gives 712 plus (two fields and the original zero-field label) training structures. Five test structures at V/Å did not pass the breakdown test and were left out. This constitutes round 0 (see Table S1).
| Round | Source | Geometries | Kept labels | Training | Validation | Test |
| 0 | AIMD, every 45th step | 434 | 1473 | 906 + 712 = 1618 | 132 | 435 |
| 1 | round-0 MD, both models | 500 | 398 | 1905 | 168 | 75 |
| 2 | round-1 MD, both models | 500 | 529 | 2291 | 215 | 96 |
Table S1. The Ne–water–Ne dataset per round. Training and validation sets are cumulative; each round has its own test set. Round 0’s training set includes the 712 zero-field-only geometries.
We then grow the dataset in two active-learning rounds. We run molecular dynamics (settings are given below) at , , V/Å for 110 ps (10 ps equilibration, 100 ps production), and sample 50 frames from each run (one every 2 ps) for both models, resulting in 500 additional geometries per round. These 500 frames were split evenly into 400 training, 50 validation and 50 test frames. Training and validation frames were labeled once at , with normally distributed with width 0.05 V/Å truncated at V/Å; test frames were labeled at the pair V/Å. In round 1, 104 of 110 labels from the short-range model trajectories at V/Å had to be discarded because the overpolarization led to dielectric breakdown.
The test set root-mean-squared errors (RMSEs) are shown in Table S2. In round 0, both models achieve similar errors, with the long-range model performing slightly better. In rounds 1 and 2, the short-range errors are about twice as large in energy and 1.5 times as large in forces, while its dipole errors stay close to those of the long-range model. The difference in the sampled dipoles (Figure 1) is therefore much larger than any difference in these test errors.
| Model | Test set | (meV) | (meV/Å) | (meÅ) |
| long-range | round 0 | 34.8 | 17.6 | 24.7 |
| round 1 | 49.2 | 19.9 | 26.9 | |
| round 2 | 42.9 | 19.3 | 25.9 | |
| short-range | round 0 | 41.4 | 21.1 | 30.5 |
| round 1 | 86.8 | 30.5 | 33.4 | |
| round 2 | 108.4 | 29.6 | 35.4 |
Table S2. Test RMSEs of the round-2 models on the test set of each round. Energies are per structure (216 atoms).
We run molecular dynamics using the Atomic Simulation Environment [53] at constant with a CSVR thermostat [54] (100 fs relaxation time) at 300 K and a 0.5 fs time step, keeping the Ne atoms fixed. In the 22 Å cell, we run the short-range and long-range models of every round at , and V/Å for 10 ps of equilibration and 100 ps of production. The 31 Å cell holds 83 in a 34 Å gap between the Ne layers ( Å). We run the round-2 models in this cell at and V/Å for 50 ps of equilibration and 150 ps of production. No models are trained on this cell. We compute and from the production runs, with on 1 Å bins and leaving out bins that hold less than half a molecule per frame on average. Error bars are block estimates over ten blocks.
Dipole response from fluctuations
In terms of the energy in Eq. (11), the mean cell dipole is
with . Differentiating with respect to and using Eq. (11), we obtain
With , this is Eq. (2) of the main text. The first term is the electronic response at frozen nuclei and the second (the variance) the orientational response of the nuclei, which follows from the equilibrium fluctuations of the dipole. Macroscopic electrostatics gives a closed-form estimate for the bias response. At fixed , linear response gives , so that . The dipole moment is , with the area and the thickness of the water slab. With ,
Taking the derivative at constant (frozen nuclei) leaves only the electronic response, so that is replaced by the high-frequency permittivity . Inserting these estimates in Eq. (2) gives an estimate for the variance,
which is used to plot the dashed Gaussian curve in Figure 1b,Figure 1c. Because is comparatively small for water, the choice of water permittivity has little effect; we use . With , \AA{} and the AIMD temperature of 298 K, this gives a standard deviation of 0.365 \AA{}, close to the 0.373 \AA{} of the AIMD trajectory itself.
For a more accurate estimate of the bias response, the electronic response at frozen nuclei is evaluated from the test set by central finite differences,
with V/\AA{} (two single-point calculations per snapshot), averaged over snapshots of the zero-field AIMD. The variance is also obtained from the AIMD trajectory. For the 22 \AA{} film, we obtain 5.26 for the frozen-nuclei term and 5.42 \AA{} per V/\AA{} for the variance term , so is 10.7 \AA{} per V/\AA{}, the slope of the dashed line in Figure 1d,Figure 1e.
Reference for the thick film
Eq. (15) also gives an estimate for the response of the thick film. With \AA{} we obtain an estimate of 14.15 \AA{} per V/\AA{}. However, with \AA{}, Eq. (15) gives 9.93 \AA{} per V/\AA{} for the 22 \AA{} cell, 7 % below the value 10.7 obtained from DFT and AIMD through Eq. (2). The response of interfacial water likely differs from that of a macroscopic continuum. Instead, we therefore take the response from the 22 \AA{} cell and add the response of the additional (bulk-like) water:
with \AA{} the difference between the water thicknesses of the two cells, which gives 14.90 \AA{} per V/\AA{}, the slope of the dashed line in Figure 1g. This value lies 5 % above the continuum estimate. The round-2 long-range model gives 14.29 \AA{} per V/\AA{}, between the two estimates, whereas the short-range model gives 25.74 \AA{} per V/\AA{}.
Work function of a metal–water–countercharge cell
In this section we keep the elementary charge explicit, so that the work function (or absolute electrode potential) is an energy; the main text uses units with . We consider a metal slab covered by a water film, with an excess charge on the electrode and a counter charge above the water (Figure S1); is the surface normal and the lateral area. The Fermi level is the electrochemical potential of the electrons, , where is their chemical potential in the metal and the electrostatic potential inside it. The work function is the energy to move an electron from the Fermi level to the vacuum just outside the counter charge, where the electrostatic potential is the outer potential :
Figure S1. Planar-averaged potential energy of an electron, , in a Pt(111)/water cell from Ref. [41] with a solvated jellium counter charge (GPAW-SJM), at (gray) and (black), with the vacuum level above the jellium () taken as zero. Part of the structure is shown on top and its scale matches the -axis. Dashed lines mark the Fermi level for the two different charges. The implicit solvent is indicated by the blue shaded region; the position of the jellium is marked by the purple bar. The step at the cell boundary is the dipole correction.
with the potential drop from the metal to the vacuum. Because and , taking gives , which is of the main text with .
To obtain , let be the charge density averaged over the lateral cell, including the electrode charge and the counter charge. We take inside the metal and in the vacuum above the counter charge. The field vanishes at both points, so by Gauss’s law the enclosed charge is zero: . To obtain , we integrate Poisson’s equation twice between and
where we switched the order of integration over the triangular integration domain. The integral is the -component of the dipole moment of the cell per unit area: .
We compare with a reference system, such as the bare slab at , with work function . For a slab thick enough that its interior is unperturbed, is the same in both, so , which depends on the charge density difference with respect to the reference system. Since vanishes below , the integral can be extended over the whole cell:
We partition into contributions of the atoms and the counter charge , . For atom at height , we define its charge and dipole,
with . Splitting for each atom, and introducing the centroids of the interfacial charge and of the counter charge,
the integral in Eq. (21) becomes
with and ; the terms in cancel because the electrode charge and the counter charge are equal and opposite. Hence
which is Eq. (3) of the main text with . The atomic dipoles contain the electronic rearrangement of both the metal surface and the water. The partition into atomic contributions is not unique; an MLIP that predicts from a sum over atoms learns one such partition.
Metal–water interfaces
Charge conditioning
We implement charge conditioning in LOREM and GRACE using feature-wise linear modulation (FiLM) [42], which transforms the descriptor features using two small neural networks and that take the conditioning parameter or as input,
where indicates element-wise multiplication. The work function is obtained by automatic differentiation in the same backward pass as the forces. The column of the Born effective charge tensors, , requires an additional backward pass. For the 108 atom Pt(111)/water cell and a 0.5 fs time step, with the work function written every 100 steps, our LOREM- production runs with ASE reach 3.0 ns/day (70 steps per second) on one NVIDIA A100.
GRACE-RAZOR
The response analysis in -orientation (RAZOR) approach [29, 41] uses two backbones, one for and one for . We implement the same logic in a single GRACE backbone, which improves computational efficiency and simplifies the integration in the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) code for molecular dynamics. By forming latent atomic dipoles from the last-layer features and contracting them with , the atomic contribution to the energy becomes
where is the number of atoms in the cell. Summed over atoms, this recovers Eq. (5) of the main text. Forces, work function and Born effective charges follow from automatic differentiation. is degenerate with a constant in ; we fix it at the clean Pt(111) value of 5.75 V, and learn only , initialized at the mean of the training set. The area is a fixed parameter of the model rather than read from the cell, because the LAMMPS interface does not pass the cell to the model.
For the 108 atom Pt(111)/water cell and a 0.5 fs time step, with the work function written every step, the production runs of this work reach 5.5 ns/day (127 steps per second) on one NVIDIA A100.
One water above Pt(111) and Pt(332)
The structures consist of one water molecule above a four-layer Pt slab, flat Pt(111) () or stepped Pt(332) (, one step per cell), in a cell of 34.81 Å along . The water geometries are 20 gas-phase molecules of different dipole magnitude. For training, we create 1200 frames per surface. For each frame, we draw a random water geometry, give it a uniformly random orientation, and place it at a height Å above the surface, with the lateral position sampled from a uniform distribution. We keep the 964 frames with Å: for higher waters, the highest occupied orbital approaches the Fermi level at positive charge, so that electrons start to leave the water, an effect not expected in a typical simulation setup. We split them at random into 868 training and 96 validation frames per surface. For the test curves, we choose a water geometry with median dipole; on Pt(111), its dipole is rotated by the polar angle at fixed position; on Pt(332), the water is moved by across the step at fixed dipole.
We label the dataset with ABACUS 3.10.1 [6] in its LCAO mode, choosing settings that provide internal consistency but not necessarily absolute accuracy. All models are trained and tested on the same labels, so the comparison between the architectures does not depend on the accuracy of these settings; only the absolute work functions are not converged. We use PBE [55], no dispersion, SG15 ONCV [56, 57] pseudopotentials, DZP orbitals (Pt and O 7 bohr, H 6 bohr), a 100 Ry grid, Gaussian smearing of 0.02 Ry, and Monkhorst–Pack [52] grids of (Pt(111)) and (Pt(332)). Each frame carries a charge drawn uniformly from e on Pt(111). For Pt(332), we use the same surface-charge density range. The counter charge sits on a plate 2.5 Å above the highest atom of all structures of a surface, behind which a barrier of 1.5 Ry keeps the electrons out of the vacuum; the plate and barrier are fixed in space. The work function is read from the planar-averaged potential in the vacuum beyond the plate. On a charge stencil, from the energies matches this work function to 0.5 mV, so the label is the charge derivative of the energy label. The small basis keeps each calculation at 2–3 minutes on one node; it shifts the work function by a constant and changes the step effect by about 5%, so the labels are consistent rather than converged.
GRACE models use two layers with a 6 Å cutoff; LOREM uses a 5 Å cutoff with one message-passing step plus the long-range Ewald term. All are trained on energy, forces and work function. Table S3 lists the validation errors and the errors on the curves of Figure 2. GRACE-RAZOR and LOREM- fit the validation energies and forces equally well, but on Pt(332) the work-function error of GRACE-RAZOR is four times larger, and only LOREM- follows the step at both heights.
| surface | model | (meV / atom) | (meV / Å) | (mV) | curves, (mV), | curves, (mV), |
| Pt(111) | LOREM- | 0.17 | 3.5 | 1.9 | 3.7 | 1.2 |
| Pt(111) | GRACE-RAZOR | 0.14 | 3.1 | 2.9 | 2.5 | 3.2 |
| Pt(111) | GRACE- | 0.52 | 46.6 | 470 | 76 | 666 |
| Pt(332) | LOREM- | 0.10 | 2.4 | 1.8 | 3.1 | 0.9 |
| Pt(332) | GRACE-RAZOR | 0.10 | 2.5 | 7.7 | 16.4 | 9.9 |
| Pt(332) | GRACE- | 0.40 | 40.3 | 386 | 432 | 667 |
Table S3. Errors of the models in Figure 2: RMSE on the 96 validation frames per surface, and RMSE of on the curves drawn in Figure 2, at and 7 Å (inside and beyond the GRACE cutoff).
Pt(111)–water
Training data
The Pt(111)–water models are trained on the published data of Ref. 41: a three-layer Pt(111) slab (36 Pt) with 24 water molecules, lateral area 82.31 Å, at bias charges , computed with GPAW and the solvated jellium method (SJM) [5, 58] using PBE-D3(BJ) [51, 55, 59], Γ-point only.
For this system, the charge response is only linear within a window of about 0.5 around the charge it was equilibrated at (Figure S2). Outside that window, electrons are deposited at or removed from states of the water (dielectric breakdown) [41]. For structures equilibrated at , the window is centered near , and at they already sit on the upper end of the linear region. For the models and results described in the main text, we therefore train and run MD only up to .
Figure S2. Work function of the Pt(111)–water models. (a) of single structures from Ref. 41, each equilibrated by MD at a charge (dots) and recomputed with GPAW-SJM at charges from to . The dotted line marks the highest training charge, . (b) from MD with the models trained on and from AIMD—the same as Figure 3h. The work function of the GRACE single- series is computed with LOREM-. (c) from MD (here, two replicas only) with the same architectures trained on (same symbols as in b), with the AIMD converted to by subtracting the average offset (Sec. S3B). At , both GRACE-RAZOR replicas and one LOREM- replica dissociate the water and are excluded. Error bars: standard errors over replicas; for the AIMD, as described in Sec. S3D.
Work-function labels
The data contain two work function labels for each structure. is the work function returned by GPAW-SJM. is obtained by finite differences of the energy on a charge stencil of , which Ref. 41 uses for its labels. For an energy surface , both should be equal, but they differ by . A model with trained on and on the SJM energies therefore receives labels that are not exactly consistent.
The difference is not an artifact of the finite-difference derivative. On 28 AIMD-sampled structures at to , we recomputed with a stencil of and obtained (standard deviation over structures), against from the labels of the same structures. We have not identified the origin of ; it implies that a term in the GPAW-SJM energy is absent from the Kohn–Sham potential that determines the Fermi level. The per-charge means of change by only per .
We train the models of the main text on , the quantity reported by the AIMD reference. Figure S2,Figure S2 compares them with the same architectures trained on . For the comparison in panel c, the AIMD is converted to by subtracting the measured . Trained on , both models follow the AIMD within , as shown in the main text (Figure 3). Trained on , both lie – above the converted AIMD; since the two models agree closely with each other, this reflects an inconsistency in the data labels rather than in the architecture.
Test errors
Table S4 compares the models on the 395 test structures with of Ref. 41. The LOREM- model most accurately predicts the work function and Born effective charges.
| model, label | (meV/atom) | (meV/Å) | (V) | (V) | () |
| LOREM-, | 0.55 | 25.5 | 0.315 | 0.071 | 0.014 |
| GRACE-RAZOR, | 0.47 | 19.2 | 0.321 | 0.086 | 0.019 |
| LOREM-, | 0.54 | 25.5 | 0.086 | 0.317 | 0.015 |
| GRACE-RAZOR, | 0.44 | 18.9 | 0.095 | 0.313 | 0.019 |
Table S4. RMSE on the 395 structures with of the test set of Ref. 41. errors shown against both labels, with the label each model was trained on in bold. : the Born effective charges along the surface normal, , per component over all atoms.
Molecular dynamics
The MD runs use the cell of the training data, with the two bottom Pt layers fixed, periodic boundary conditions in and , and the harmonic wall on O of Ref. 41 acting from 21 Å. They run at 350 K with a CSVR thermostat [54] (100 fs relaxation time) and a 0.5 fs time step: four replicas of 250 ps per charge, each started from a frame of the published RAZOR-qQ trajectory at that charge. GRACE runs in LAMMPS and LOREM-q in ASE. We discard the first 30 ps of each run and keep a frame every 100 fs.
The AIMD reference uses GPAW-SJM with the settings of the training data and D3(BJ) forces, a time step of 1 fs and the same thermostat settings as described above. At each charge, we ran two 7 ps trajectories, started from frames of two independent MLIP trajectories, and discard the first 2 ps of each.
Error bars of the MLIP series are standard errors over the replicas; for the AIMD, with two chains per charge, we use a block standard error scaled by the factor that makes block errors match the scatter between independent MLIP runs (1.71), or half the drift between the two halves of a chain, whichever is larger.
GRACE single- reference
The single- reference consists of nine standard two-layer GRACE models (cutoff 6 Å, effective cutoff at most 12 Å), one per charge from −1 to +1 in steps of 0.25 , trained on energy and forces only. Each is trained on 1000 to 1375 structures at its own charge, about half of them from Ref. 41 and the rest new GPAW-SJM labels with the same settings as the original dadta.
The new structures come from our MD at 400 K with previous charge-conditioned models at that charge, sampled every 0.5 ps to 1 ps to create a pool of candidates. From these we select frames with the maximum set coverage (MSC) algorithm [60] of the QUESTS package [61], applied to SOAP descriptors [62] of the interfacial O atoms computed with Featomic [63], and starting from the existing labels at that charge, to choose O environments least represented in the dataset. The descriptor vectors are normalized to have norm 1, and the kernel bandwidth parameter in QUESTS is set to 0.085 so that the overlap between physisorbed and chemisorbed O’s is about 0.1. Frames with a work function outside 2.5 < < 7 V are excluded from training, as they likely exhibit dielectric breakdown. MD with the first single- models showed adsorbed OH or O at negative charges, which is unphysical. We therefore ran one additional round of active learning, sampling the single- trajectories in the same way.
The production single- MD was run with the same settings as the charge-conditioned GRACE- models (Sec. S3D), with four 250 ps replicas per charge. As these models have no work-function output, the potential of the single- series is computed with LOREM- on its frames.
References
- [1]A. Mathlanker, J. Guo, B. R. Goldsmith, J. B. Varley, and N. Govindarajan, Estimating potential-dependent physicochemical properties at metal–electrolyte interfaces using machine learning interatomic potentials, ACS Electrochemistry 2, 1176 (2026).DOI
- [2]L. Li, T. Eggert, K. Reuter, and N. G. Hörmann, Electron Spillover into Water Layers: A Quantum Leap in Understanding Capacitance Behavior, J. Am. Chem. Soc. 147, 22778 (2025).DOI
- [3]M. Zhang, Y. Chen, M. M. Melander, and J. Huang, Electrochemical electron transfer: Key concepts, theories, and parameterization via atomistic simulations, Chemical Reviews 126, 7407 (2026).arxiv.org/abs/2510.24635
- [4]S. Sundentral, M. Todorova, M. W. Finnis, and J. Neugebauer, First-principles approach to model electrochemical reactions: understanding the fundamental mechanisms behind mg corrosion, Physical Review Letters 120, 246801 (2018).DOI
- [5]G. Kastlunger, P. Lindgren, and A. A. Peterson, Controlled-Potential Simulation of Elementary Electrochemical Reactions: Proton Discharge on Metal Surfaces, J. Phys. Chem. C 122, 12771 (2018).DOI
- [6]W. Zhou, D. Zheng, Q. Liu, D. Lu, Y. Liu, P. Lin, Y. Huang, X. Peng, Y. J. Bao, C. Cai, Z. Jin, J. Wu, H. Zhang, G. Jin, Y. Ji, Z. Shen, X. Liu, L. Sun, Y. Cao, M. Sun, J. Liu, T. Chen, R. Liu, Y. Li, H. Han, X. Liang, T. Bao, Z. Deng, T. Liu, N. Chen, H. Ren, X. Zhang, Z. Liu, Y. Fu, M. Liu, Z. Li, T. Wen, Z. Tang, Y. Xu, W. Duan, X. Wang, Q. Gu, F.-Z. Dai, Q. Zheng, Y. Zhong, H. Xiang, X. Gong, J. Zhao, Y. Zhang, Q. Ou, H. Jiang, S. Liu, B. Xu, S. Xu, X. Ren, L. He, L. Zhang, and M. Chen, ABACUS: An electronic structure analysis package for the AI era, J. Chem. Phys. 163, 192501 (2025).arxiv.org/abs/2501.08697
- [7]S. Mattoso, J. Yang, F. Deimnbeck, A. Abdelkawy, C. Freysoldt, S. Wippermann, M. Todorova, and J. Neugebauer, User-Defined Electrostatic Potentials in DFT Supercell Calculations: Implementation and Application to Electrified Interfaces, J. Chem. Theory Comput. 22, 2597 (2026).arxiv.org/abs/2510.23328
- [8]M. Stengel, N. A. Spaldin, and D. Vanderbilt, Electric displacement as the fundamental variable in electronic-structure calculations, Nat. Phys. 5, 304 (2009).DOI
- [9]C. Zhang and M. Sprik, Finite field methods for the supercell modeling of charged insulator/electrolyte interfaces, Phys. Rev. B 94, 245309 (2016).DOI
- [10]K. Brezina, L. Ahrens-Iwers, and M. Rossi, Electrochemical Molecular Dynamics within the Finite-Field Framework (2026), arXiv:2610.23291 [physics.chem-ph].arxiv.org/abs/2610.23291
- [11]N. Bonnet, T. Morishita, O. Sugino, and M. Otani, First-principles molecular dynamics at a constant electrode potential, Physical Review Letters 109, 266101 (2012).pubmed.ncbi.nlm.nih.gov/23368585
- [12]F. Deibenbeck and S. Wippermann, Dielectric Properties of Nanoconfined Water from Ab Initio Thermopotentiostat Molecular Dynamics, J. Chem. Theory Comput. 19, 1035 (2023).DOI
- [13]R. Sechi, G. Kastlunger, A. Bhowmik, and H. A. Hansen, First-principles molecular dynamics with potential and charge fluctuations applied to au (111) in alkaline solutions, Journal of Chemical Theory and Computation 21, 5279 (2025).DOI
- [14]M. Sun, L. Fu, and S. Xu, Electrochemical Potential Fluctuation Matters in Rate Constant Calculations for Proton-Coupled Electron Transfer, J. Chem. Theory Comput. 22, 1539 (2026).DOI
- [15]A. Hagopian, M. T. M. Koper, and K. Doblhoff-Dier, Assessing the potential of zero charge in ab initio molecular dynamics simulations, J. Chem. Phys. 164, 174711 (2026).DOI
- [16]A. R. N. Aouichaoui, D. Tisi, I. E. Castelli, and A. Bhowmik, Machine learning potentials for electrified interfaces, RSC Appl. Interfaces 10.1039/d6lf00149a (2026).DOI
- [17]R. Wang, S. Fang, Q. Huang, and Y. Liu, Constant-Potential Machine Learning Force Field for the Electrochemical Interface, J. Chem. Theory Comput. 21, 7628 (2025).DOI
- [18]M. Sun, B. Jin, X. Yang, and S. Xu, Probing nuclear quantum effects in electrocatalysis via a machine-learning enhanced grand canonical constant potential approach, Nat. Commun. 16, 3600 (2025).DOI
- [19]L. Chen, Y. Tian, X. Hu, S. Chen, H. Wang, X. Zhang, and Z. Zhou, Constant-Potential MD with Neural Network Potentials Reveals Cation Effects on CO₂ Reduction at Au-Water Interfaces, JACS Au 6, 868 (2026).DOI
- [20]S. Sahel-Schackis, K.-i. Nomura, A. Nakano, M. F. Kling, and T. Linker, EquiFILM: Charge-Conditioned Equivariant Force Fields via Feature-wise Linear Modulation (2026), arXiv:2607.05559.
- [21]O. T. Unke, S. Chmiela, M. Gastegger, K. T. Schütt, H. E. Sauceda, and K.-R. Müller, SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects, Nat. Commun. 12, 7273 (2021).DOI
- [22]X. Xie, K. A. Persson, and D. W. Small, Incorporating Electronic Information into Machine Learning Potential Energy Surfaces via Approaching the Ground-State Electronic Energy as a Function of Atom-Based Electronic Populations, J. Chem. Theory Comput. 16, 4256 (2020).arxiv.org/abs/2003.01893
- [23]T. W. Ko, J. A. Finkler, S. Goedecker, and J. Behler, A fourth-generation high-dimensional neural network potential with accurate electrostatics including non-local charge transfer, Nat. Commun. 12, 398 (2021).DOI
- [24]I. Batatia, W. J. Baldwin, D. Kuryla, J. Hart, E. Kasar, A. M. Elena, H. Moore, M. J. Gawkowski, B. X. Shi, V. Kapil, P. Kourtis, I.-B. Magdău, and G. Csányi, MACÉ-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry (2026), arXiv:2602.19411 [physics].arxiv.org/abs/2602.19411
- [25]A. Gao and R. C. Remsing, Self-consistent determination of long-range electrostatics in neural network potentials, Nat. Commun. 13, 1572 (2022).DOI
- [26]K. Joll, P. Schienbein, K. M. Rosso, and J. Blumberger, Machine learning the electric field response of condensed phase systems using perturbed neural network potentials, Nat. Commun. 15, 8192 (2024).DOI
- [27]S. Falletta, A. Cepellotti, A. Johansson, C. W. Tan, M. L. Descoteaux, A. Musialian, C. J. Owen, and B. Kozinsky, Unified differentiable learning of electric response, Nat. Commun. 16, 4031 (2025).arxiv.org/abs/2403.17207
- [28]B. A. A. Martin, A. M. Ganose, V. Kapil, T. Li, and K. T. Butler, General Learning of the Electric Response of Inorganic Materials, PRX Energy 1, 013006 (2026).DOI
- [29]N. Bergmann, N. Bonnet, N. Marzari, K. Reuter, and N. G. Hörmann, Machine Learning the Energetics of Electrified Solid-Liquid Interfaces, Phys. Rev. Lett. 135, 146201 (2025).DOI
- [30]I. J. Parker, M. J. Hoffmann, W. J. Baldwin, S. Han, S. Gupta, K. D. Fong, A. Michaelides, C. Schran, S. De, and G. Csányi, False Metallization in Short-Ranged Machine Learned Interatomic Potentials (2026), arXiv:2603.04228 [physics.chem-ph].arxiv.org/abs/2603.04228
- [31]L. B. T. de Kam, J.-X. Zhu, A. Mathanker, K. Doblhoff-Dier, and N. Govindarajan, Modeling the inner electric double layer using short-range machine learning potentials: A critical assessment, J. Chem. Phys. 165, 184709 (2026).DOI
- [32]B. Sumić, R. Vasdev, S. K. Ethirajan, J. Yang, C. S. Cucinotta, R. G. Hennig, K. Reuter, S. Ringe, M. Todorova, C. Freysoldt, and J. Neugebauer, Electrostatic Phenomenology Benchmarks for Machine-Learned Interatomic Potentials in Electrochemistry: Beyond the Energy-Force Metric (2026), arXiv:2608.14153 [cond-mat.mtrl-sci].arxiv.org/abs/2608.14153
- [33]J. Yang, B. Li, S. Mattoso, A. Abdelkawy, M. Todorova, and J. Neugebauer, Mace-Reflect-Constrained Learning of Atomic Charges for Accurate Electrostatic Potentials at Electrochemical Interfaces, Phys. Rev. Lett. 136, 118001 (2026).DOI
- [34]F. Grasselli, K. Rossi, S. de Gironcoli, and A. Grisafi, Long-range electrostatics in atomistic machine learning: A physical perspective (2026), arXiv:2602.11071 [cond-mat.mtrl-sci].
- [35]J.-X. Zhu and J. Cheng, Machine Learning Potential for Electrochemical Interfaces with Hybrid Representation of Dielectric Response, Phys. Rev. Lett. 135, 018003 (2025).DOI
- [36]X. Wang, J. Chen, Z. Zeng, F. Stein, J. Lim, and B. Cheng, Ion-modulated structure, proton transfer, and capacitance in the Pt-water electric double layer, Npj Comput. Mater. 10.1038/s41524-026-02230-7 (2026).arxiv.org/abs/2509.13727
- [37]J. I. Siepmann and M. Sprik, Influence of surface topology and electrostatic potential on water/electrode systems, J. Chem. Phys. 102, 511 (1995).DOI
- [38]A. Grisafi, A. Bussy, M. Salanne, and R. Vuilleumier, Predicting the charge density response in metal electrodes, Phys. Rev. Mater. 7, 125403 (2023).DOI
- [39]A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Graph Atomic Cluster Expansion for Semilocal Interactions beyond Equivariant Message Passing, Phys. Rev. X 14, 021036 (2024).arxiv.org/abs/2311.16326
- [40]E. Rumiantsev, M. F. Langer, T.-E. Sodjarugal, M. Cerriotti, and P. Loche, Learning long-range representations with equivariant messages, Trans. Mach. Learn. Res. (2026).arxiv.org/abs/2507.19382
- [41]N. Bergmann, K. Reuter, and N. G. Hörmann, Erasing dielectric breakdown artifacts to machine-learn charged Pt–water interfaces, J. Chem. Phys. 164, 174110 (2026).DOI
- [42]E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville, FiLM: Visual Reasoning with a General Conditioning Layer, Proc. AAAI Conf. Artif. Intell. 32, 10.1609/aaai.v32i1.11671 (2018).arxiv.org/abs/1709.07871
- [43]I.-C. Yeh and M. L. Berkowitz, Ewald summation for systems with slab geometry, J. Chem. Phys. 111, 3155 (1999).DOI
- [44]J. Liu, J. Su, X. Yao, Z. Jiang, G. Lai, Y. Du, Y. Qin, W. Xu, E. Lu, J. Yan, Y. Chen, H. Zheng, Y. Liu, S. Liu, B. Yin, W. He, H. Zhu, Y. Wang, J. Wang, M. Dong, Z. Zhang, Y. Kang, H. Zhang, X. Xu, Y. Zhang, Y. Wu, X. Zhou, and Z. Yang, Muon is Scalable for LLM Training (2025).arxiv.org/abs/2502.16982
- [45]G. Kresse and J. Hafner, Ab initio molecular dynamics for liquid metals, Physical review B 47, 558 (1993).pubmed.ncbi.nlm.nih.gov/10004490
- [46]G. Kresse and J. Hafner, Ab initio molecular-dynamics simulation of the liquid-metal–amorphous-semiconductor transition in germanium, Physical Review B 49, 14251 (1994).pubmed.ncbi.nlm.nih.gov/10010505
- [47]G. Kresse and J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set, Computational materials science 6, 15 (1996).DOI
- [48]G. Kresse and J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Physical review B 54, 11169 (1996).pubmed.ncbi.nlm.nih.gov/9984901
- [49]G. Kresse and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Physical review B 59, 1758 (1999).DOI
- [50]B. Hammer, L. B. Hansen, and J. K. Nørskov, Improved adsorption energetics within density-functional theory using revised Perdew-Burke-Ernzerhof functionals, Physical review B 59, 7413 (1999).DOI
- [51]S. Grimme, J. Antony, S. Ehrlich, and H. Krieg, A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu, The Journal of chemical physics 132 (2010).DOI
- [52]H. J. Monkhorst and J. D. Pack, Special points for Brillouin-zone integrations, Physical review B 13, 5188 (1976).
- [53]A. Hjorth Larsen, J. Jørgen Mortensen, J. Blomqvist, I. E. Castelli, R. Christiansen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, et al., The atomic simulation environment—a python library for working with atoms, Journal of Physics: Condensed Matter 29, 273002 (2017).DOI
- [54]G. Bussi, D. Donadio, and M. Parrinello, Canonical sampling through velocity rescaling, The Journal of chemical physics 126 (2007).DOI
- [55]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Physical review letters 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
- [56]M. Schlipf and F. Gygi, Optimization algorithm for the generation of oncv pseudopotentials, Computer Physics Communications 196, 36 (2015).DOI
- [57]D. R. Hamann, Optimized norm-conserving Vanderbilt pseudopotentials, Physical Review B—Condensed Matter and Materials Physics 88, 085117 (2013).DOI
- [58]J. J. Mortensen, A. H. Larsen, M. Kuisma, A. V. Ivanov, A. Taghizadeh, A. Peterson, A. Haldar, A. O. Dohn, C. Schäfer, E. Ö. Jónsson, et al., Gpaw: An open python package for electronic structure calculations, The Journal of Chemical Physics 160 (2024).DOI
- [59]S. Grimme, S. Ehrlich, and L. Goerigk, Effect of damping function in dispersion corrected density functional theory, J. Comput. Chem. 32, 1456 (2011).DOI
- [60]B. Yu, V. Lordi, and D. Schwalbe-Koda, Maximizing efficiency of dataset compression for machine learning potentials with information theory, The Journal of Chemical Physics 164 (2026).arxiv.org/abs/2511.10561
- [61]D. Schwalbe-Koda, S. Hamel, B. Sadigh, F. Zhou, and V. Lordi, Model-free estimation of completeness, uncertainties, and outliers in atomistic machine learning using information theory, Nature Communications 16, 4014 (2025).DOI
- [62]A. P. Bartók, R. Kondor, and G. Csányi, On representing chemical environments, Physical Review B—Condensed Matter and Materials Physics 87, 184115 (2013).DOI
- [63]F. Bigi, J. W. Abbott, P. Loche, A. Mazitov, D. Tisi, M. F. Langer, A. Goscinski, P. Pegolo, S. Chong, R. Goswami, P. Febrer, S. Chorna, M. Kellner, M. Ceriotti, and G. Fraux, Metatensor and metatomic: Functional libraries for interoperable atomistic machine learning, J. Chem. Phys. 164, 064113 (2026).



