betaFeedback

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 D\mathbf{D} 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., D\mathbf{D}, 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 ±q\pm q, which fixes the displacement field at D=(q/A)z^\mathbf{D}=(q/A)\hat{\mathbf{z}}, where AA is the lateral area and z^\hat{\mathbf{z}} the surface normal. Here, we realize the same boundary condition by applying an electric field E0z^E_0\hat{\mathbf{z}} together with a dipole correction, so that D=ε0E0z^\mathbf{D}=\varepsilon_0E_0\hat{\mathbf{z}}. Note that E0E_0 is the field in the empty cell; the macroscopic field inside the water, E=(D−P)/ε0\mathbf{E}=(\mathbf{D}-\mathbf{P})/\varepsilon_0, is screened by the water polarization P\mathbf{P}. As shown in Sec. S1A of the Supplemental Material, the energy of this system can be written as

U(D)=U0−1ε0M⋅D=U0−MzE0,(1)U(\mathbf{D})=U_0-\frac{1}{\varepsilon_0}\mathbf{M}\cdot\mathbf{D}=U_0-M_zE_0, \tag*{(1)}

where U0U_0 is the energy at zero displacement field and M\mathbf{M} the cell dipole moment. The bias response of the energy is thus given by the cell dipole in zz, Mz=−∂U/∂E0M_z=-\partial U/\partial E_0. The capacitor is symmetric under z→−zz\to-z, 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 E0z^E_0\hat{\mathbf{z}} to the equivariant features of the model, similar to Ref. 28 (see Sec. S1B of the Supplemental Material for the implementation). With MzM_z obtained by automatic differentiation of U=∑iUiU=\sum_i U_i, the model represents the cell dipole as a sum of atomic dipoles, Mz=∑iμi,zM_z=\sum_i\mu_{i,z}. 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 UU, forces Fi\mathbf{F}_i, and cell dipole MzM_z 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 ∣E0∣≤0.2|E_0|\leq0.2 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 MzM_z at D=ε0E0=0D=\varepsilon_0E_0=0 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) ⟨Mz⟩\langle M_z\rangle from molecular dynamics at E0=D/ε0=0,±0.1,±0.2E_0=D/\varepsilon_0=0,\pm0.1,\pm0.2 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 ⟨cos⁡θ⟩(z)\langle\cos\theta\rangle(z) with θ\theta the angle between the water bisector and the zz-axis, at a field of D/ε0=0.1D/\varepsilon_0=0.1 V/Å; the shading marks the water at least 10 Å from both Ne layers; (g) ⟨Mz⟩\langle M_z\rangle against D/ε0D/\varepsilon_0; dashed line: the continuum value ε0ALw(1−1/ε)\varepsilon_0AL_w(1-1/\varepsilon). Error bars and bands are block estimates.

In zero-field MD, the short-range models show substantially larger fluctuations in MzM_z 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 MzM_z is related to the variance of its fluctuations (see Sec. S1D of the Supplemental Material for the derivation),

∂⟨Mz⟩∂E0∣E0=0=⟨∂Mz∂E0∣R⟩+βVar⁡(Mz),(2)\left.\frac{\partial\langle M_z\rangle}{\partial E_0}\right|_{E_0=0} = \left\langle\left.\frac{\partial M_z}{\partial E_0}\right|_{\mathbf{R}}\right\rangle +\beta\operatorname{Var}(M_z), \tag*{(2)}

where ∂Mz/∂E0∣R\partial M_z/\partial E_0|_R is the electronic response at fixed nuclear positions. Evaluating this first term by finite differences on a few snapshots labeled at E0=±0.1E_0=\pm0.1 V/Å and obtaining the variance from the zero-field AIMD trajectory yields a reference for ∂⟨Mz⟩/∂E0∣E0=0\partial\langle M_z\rangle/\partial E_0|_{E_0=0}. 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 ε0AΔL(1−1/ε)\varepsilon_0 A\Delta L(1-1/\varepsilon) of the additional bulk water with thickness ΔL≈9\Delta L\approx9 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 Φ=∂U/∂q\Phi=\partial U/\partial q, where qq is in units of the elementary charge ee. 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,

Φ=Φref−ΔMzε0A+qdε0A,(3)\Phi=\Phi_{\mathrm{ref}}-\frac{\Delta M_z}{\varepsilon_0 A}+\frac{qd}{\varepsilon_0 A}, \tag*{(3)}

where Φref\Phi_{\mathrm{ref}} is the work function of a reference system such as the bare metal slab and ΔMz=∑iΔμi,z\Delta M_z=\sum_i\Delta\mu_{i,z} is the change in dipole moment relative to this reference. d=zˉcc−zˉqd=\bar{z}_{\mathrm{cc}}-\bar{z}_q is the distance between the centroid zˉcc\bar{z}_{\mathrm{cc}} of the countercharge and the centroid zˉq\bar{z}_q of the excess charge qq on the metal and water atoms. Note that ΔMz\Delta M_z 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

U(q)=U0+∫0qΦ(q′) dq′,(4)U(q)=U_0+\int_0^q\Phi(q')\,\mathrm{d}q', \tag*{(4)}

with U0=U(q=0)U_0=U(q=0). Assuming that ΔMz\Delta M_z depends only linearly on qq and that dd is constant, we recover the quadratic expansion [29]

U(q)=U0+qΦ0+q22C0,(5)U(q)=U_0+q\Phi_0+\frac{q^2}{2C_0}, \tag*{(5)}

with the zero-charge work function and capacitance

Φ0=Φref−ΔMz,0ε0A,1C0=1ε0A(d−∂ΔMz∂q),(6)\Phi_0=\Phi_{\mathrm{ref}}-\frac{\Delta M_{z,0}}{\varepsilon_0 A},\qquad\frac{1}{C_0}=\frac{1}{\varepsilon_0 A}\left(d-\frac{\partial\Delta M_z}{\partial q}\right), \tag*{(6)}

with ΔMz,0=ΔMz(q=0)\Delta M_{z,0}=\Delta M_z(q=0). Compared to the capacitor model from Eq. (1), the metal–water–countercharge system introduces a linear term qΦrefq\Phi_{\mathrm{ref}} and a quadratic term q2/(2C0)q^2/(2C_0), both of which are properties of the data and must be learned by the MLIP. The assumptions of constant dd and linear ΔMz(q)\Delta M_z(q) imply that the capacitance C0C_0 is constant. This holds only in a narrow bias window around the value of qq 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 zˉq\bar{z}_q changes and dd is no longer constant.

Many existing strategies [17, 18, 19] condition an O(3)O(3)-symmetric model on a scalar qq to learn U(q)U(q). 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 qq 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 Φ=∂U/∂q\Phi= \partial U/\partial q and the surface-normal components of the Born effective charges, Zi,z∗=ε0A∂Fi/∂qZ^{*}_{i,z} = \varepsilon_{0} A \partial F_{i}/\partial q, then follow by automatic differentiation. We denote the resulting models GRACE-qq and LOREM-qq (see Sec. S2A of the Supplemental Material for implementation details). To complete the comparison, we use the field-conditioned response analysis in zz-orientation (RAZOR) approach [29], which learns the zero-charge work function Φ0\Phi_{0} of Eq. (5) from latent atomic dipoles and the capacitance C0C_{0} 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-qq, and LOREM-qq) 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 q=0q=0 (∣σ∣≤0.5 μC/cm2|\sigma| \le0.5\,\mu\mathrm{C}/\mathrm{cm}^{2}), 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).

Work function of one water molecule above flat and stepped Pt

Figure 2. Work function of one water molecule above flat and stepped Pt. (a) Water at height hh above Pt(111), its dipole rotated by θ\theta; (b) water moved by xx across a step of Pt(332). Circles mark the cutoff rcr_c. (c,e) Φ\Phi against θ\theta on Pt(111) at h=5h=5 and 7 Å. (d,f) ΔΦ\Delta\Phi, relative to the abab mean, against xx on Pt(332); at h=7h=7 Å GRACE-qq 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-qq 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 xx direction. A homogeneous applied field cannot describe the locally inhomogeneous field near the step. At h=5h=5 Å, where the step changes Φ\Phi by about 30 mV, the work-function error of GRACE-RAZOR along the curve is five times that of LOREM-qq (16 vs. 3 mV). In contrast, our charge-conditioned long-range model LOREM-qq 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 2×52 \times5 ps for each charge), and GRACE models trained at a single charge each (GRACE single-qq). 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 q=+1q = +1 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 3×4×33 \times4 \times3 Pt(111) slab with 24 water molecules and a solvated jellium countercharge [5]. (b–d) Oxygen density ρO\rho_{\mathrm{O}} and (e–g) orientation profile ρO⟨cos⁡θ⟩\rho_{\mathrm{O}}\langle\cos\theta\rangle against the height above the top Pt layer, at q=−1q=-1, 0, and +0.5 e+0.5\,e. (h) Mean work function or electrode potential against qq (top axis: charge density σ\sigma); its slope is the inverse capacitance. The work function of GRACE single-qq is evaluated with LOREM-qq on its trajectories. (i) Mean number of chemisorbed water molecules ⟨nchem⟩\langle n_{\mathrm{chem}}\rangle, 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 q=−1q = -1 to +0.75+0.75 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 E0z^E_0\hat{z} with a slab dipole correction. The cell is periodic in xx and yy, has length LL along zz and a volume Ω=AL\Omega= AL. The displacement field is D=ϵ0E0z^\mathbf{D} = \epsilon_0 E_0\hat{z} throughout the cell. At fixed D\mathbf{D}, the energy of a periodic, insulating cell is [8]

U(D)=UPBC+Ω2ϵ0∣D−P∣2,(7)U(\mathbf{D}) = U_{\mathrm{PBC}} + \frac{\Omega}{2\epsilon_0} \lvert\mathbf{D} - \mathbf{P} \rvert^2, \tag*{(7)}

where UPBCU_{\mathrm{PBC}} is the energy under ordinary periodic boundary conditions. The polarization P\mathbf{P} is related to the dipole moment M\mathbf{M} as M=ΩP\mathbf{M} = \Omega\mathbf{P}. With D\mathbf{D} and P\mathbf{P} along zz, expanding the square gives

U=UPBC+Mz22ϵ0Ω−MzDzϵ0⏟U0+ΩDz22ϵ0.(8)U = U_{\mathrm{PBC}} + \underbrace{\frac{M_z^2}{2\epsilon_0\Omega} - \frac{M_zD_z}{\epsilon_0}}_{U_0} + \frac{\Omega D_z^2}{2\epsilon_0}. \tag*{(8)}

The first two terms form the energy at zero displacement field; Mz2/(2ϵ0Ω)M_z^2/(2\epsilon_0\Omega) 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 −∂U/∂E0-\partial U/\partial E_0 equal to the dipole MzM_z of the atoms and electrons, without the empty-cell term ϵ0ΩE0\epsilon_0\Omega E_0.

U0U_0 and MzM_z also depend on the electronic state ψ\psi. A finite-field DFT calculation returns

U(R,E0)=min⁡ψ[U0(ψ;R)−Mz(ψ;R)E0],(9)U(\mathbf{R}, E_0) = \min_{\psi} \left[ U_0(\psi; \mathbf{R}) - M_z(\psi; \mathbf{R}) E_0 \right], \tag*{(9)}

with ψ∗\psi^* the minimizing state. Because ψ∗\psi^* is stationary, its dependence on E0E_0 does not enter the first derivative of UU, so that

∂U∂E0∣R=∂U∂ψ∣ψ∗∂ψ∗∂E0−Mz(ψ∗;R)=−Mz(ψ∗;R)(10)\left.\frac{\partial U}{\partial E_0}\right|_{\mathbf{R}} = \left.\frac{\partial U}{\partial\psi}\right|_{\psi^*}\frac{\partial\psi^*}{\partial E_0} - M_z(\psi^*;\mathbf{R}) = -M_z(\psi^*;\mathbf{R}) \tag*{(10)}

(Hellmann–Feynman theorem). The DFT labels therefore satisfy

U(R,E0)=U0(R,E0)−Mz(R,E0)E0,Mz=−∂U∂E0∣R,(11)U(\mathbf{R}, E_0) = U_0(\mathbf{R}, E_0) - M_z(\mathbf{R}, E_0) E_0,\qquad M_z = -\left.\frac{\partial U}{\partial E_0}\right|_{\mathbf{R}}, \tag*{(11)}

where both U0U_0 and MzM_z are evaluated at ψ∗(E0)\psi^*(E_0) 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 hi(l)\mathbf{h}_i^{(l)} with the applied field in a weighted tensor product,

hi(l)←hi(l)−hi(l)⊗wE0(1),(12)\mathbf{h}_i^{(l)} \leftarrow\mathbf{h}_i^{(l)} - \mathbf{h}_i^{(l)} \otimes_{\mathbf{w}} \mathbf{E}_0^{(1)}, \tag*{(12)}

where E0(1)\mathbf{E}_0^{(1)} is the l=1l = 1 spherical harmonic representation of the applied field E0z^E_0\hat{\mathbf{z}}. The energy is constructed from the last-layer scalar features as U=∑iUi(hi(0))U = \sum_i \mathcal{U}_i(\mathbf{h}_i^{(0)}). Because the cell dipole Mz=−∂U/∂E0M_z = -\partial U/\partial E_0 is obtained from automatic differentiation, the model effectively represents it as a sum over atomic dipole contributions, Mz=∑iμi,zM_z = \sum_i \mu_{i,z}.

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 l=2l = 2, 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 (l=0l = 0, scalar only) with a slab correction along zz, 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 MzM_z with loss weights 1000 : 1 : 0.5, using the Muon optimizer [44] with a learning rate of 2×10−42 \times10^{-4}, 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 R2R^2 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 8.464×9.773×44.068.464 \times9.773 \times44.06 Å3^3 and holds 64 H2O\mathrm{H_2O} between two Ne layers, 216 atoms in total, with about 22 Å of water. It is periodic in xx and yy and non-periodic along zz. To label the data we run DFT calculations with VASP 6.4.3 [45, 46, 47, 48, 49] in a finite field E0z^E_0\hat{z} 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 2×2×12 \times2 \times1 Monkhorst–Pack [52] grid, 0.1 eV Fermi smearing and an electronic convergence of 10−610^{-6} 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, ±E0\pm E_0 with ∣E0∣\lvert E_0\rvert drawn uniformly from [0.05,0.2][0.05, 0.2] V/Å. Test frames get ±0.1\pm0.1 and ±0.2\pm0.2 V/Å. This gives 712 plus 3×3023 \times302 (two fields and the original zero-field label) training structures. Five test structures at ±0.2\pm0.2 V/Å did not pass the breakdown test and were left out. This constitutes round 0 (see Table S1).

RoundSourceGeometriesKept labelsTrainingValidationTest
0AIMD, every 45th step4341473906 + 712 = 1618132435
1round-0 MD, both models500398190516875
2round-1 MD, both models500529229121596

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 E0=0E_0 = 0, ±0.1\pm0.1, ±0.2\pm0.2 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 E0+δE_0 + \delta, with δ\delta normally distributed with width 0.05 V/Å truncated at ∣δ∣≤0.08\lvert\delta\rvert\le0.08 V/Å; test frames were labeled at the pair E0±0.05E_0 \pm0.05 V/Å. In round 1, 104 of 110 labels from the short-range model trajectories at E0±0.2E_0 \pm0.2 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.

ModelTest setEE (meV)FF (meV/Å)MzM_z (meÅ)
long-rangeround 034.817.624.7
round 149.219.926.9
round 242.919.325.9
short-rangeround 041.421.130.5
round 186.830.533.4
round 2108.429.635.4

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 E0E_0 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 E0=0E_0 = 0, ±0.1\pm0.1 and ±0.2\pm0.2 V/Å for 10 ps of equilibration and 100 ps of production. The 31 Å cell holds 83 H2O\mathrm{H_2O} in a 34 Å gap between the Ne layers (Lw=31.34L_w = 31.34 Å). We run the round-2 models in this cell at E0=0E_0 = 0 and ±0.1\pm0.1 V/Å for 50 ps of equilibration and 150 ps of production. No models are trained on this cell. We compute ⟨Mz⟩\langle M_z\rangle and ⟨cos⁡θ⟩(z)\langle\cos\theta\rangle(z) from the production runs, with ⟨cos⁡θ⟩(z)\langle\cos\theta\rangle(z) 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

⟨Mz⟩=1Z∫Mz(R,E0)e−βU(R,E0) dR,Z=∫e−βU(R,E0) dR,(13)\langle M_z\rangle= \frac{1}{Z}\int M_z(\mathbf{R},E_0)e^{-\beta U(\mathbf{R},E_0)}\,\mathrm{d}\mathbf{R},\qquad Z=\int e^{-\beta U(\mathbf{R},E_0)}\,\mathrm{d}\mathbf{R}, \tag*{(13)}

with β=1/kBT\beta= 1/k_{\mathrm{B}}T. Differentiating with respect to E0E_0 and using Eq. (11), we obtain

∂⟨Mz⟩∂E0=⟨∂Mz∂E0∣R⟩−β⟨Mz∂U∂E0⟩+β⟨Mz⟩⟨∂U∂E0⟩=⟨αzz⟩+β(⟨Mz2⟩−⟨Mz⟩2).(14)\begin{aligned} \frac{\partial\langle M_z\rangle}{\partial E_0}=\left\langle\left.\frac{\partial M_z}{\partial E_0}\right|_{\mathbf{R}}\right\rangle-\beta\left\langle M_z\frac{\partial U}{\partial E_0}\right\rangle+\beta\langle M_z\rangle\left\langle\frac{\partial U}{\partial E_0}\right\rangle\\ &=\langle\alpha_{zz}\rangle+\beta\left(\langle M_z^2\rangle-\langle M_z\rangle^2\right). \tag*{(14)} \end{aligned}

With αzz=∂Mz/∂E0∣R\alpha_{zz}=\left.\partial M_z/\partial E_0\right|_{\mathbf{R}}, 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 D=ε0E+P\mathbf{D}=\varepsilon_0\mathbf{E}+\mathbf{P}, linear response gives D=εE\mathbf{D}=\varepsilon\mathbf{E}, so that P=(1−1/ε)D\mathbf{P}=(1-1/\varepsilon)\mathbf{D}. The dipole moment is M=ALwP\mathbf{M}=AL_{\mathrm{w}}\mathbf{P}, with AA the area and LwL_{\mathrm{w}} the thickness of the water slab. With D=ε0E0z^\mathbf{D}=\varepsilon_0E_0\hat{\mathbf{z}},

∂⟨Mz⟩∂E0=ε0ALw(1−1ε).(15)\frac{\partial\langle M_z\rangle}{\partial E_0}=\varepsilon_0AL_{\mathrm{w}}\left(1-\frac{1}{\varepsilon}\right). \tag*{(15)}

Taking the derivative at constant R\mathbf{R} (frozen nuclei) leaves only the electronic response, so that ε\varepsilon is replaced by the high-frequency permittivity ε∞\varepsilon_\infty. Inserting these estimates in Eq. (2) gives an estimate for the variance,

Var⁡(Mz)=ε0ALwkBT(1ε∞−1ε),(16)\operatorname{Var}(M_z)=\varepsilon_0AL_{\mathrm{w}}k_{\mathrm{B}}T\left(\frac{1}{\varepsilon_\infty}-\frac{1}{\varepsilon}\right), \tag*{(16)}

which is used to plot the dashed Gaussian curve in Figure 1b,Figure 1c. Because 1/ε1/\varepsilon is comparatively small for water, the choice of water permittivity has little effect; we use ε=78.5\varepsilon=78.5. With ε∞=1.89\varepsilon_\infty=1.89, Lw=22.0L_{\mathrm{w}}=22.0 \AA{} and the AIMD temperature of 298 K, this gives a standard deviation of 0.365 ee \AA{}, close to the 0.373 ee \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,

αzz≈Mz(+ΔE)−Mz(−ΔE)2ΔE(17)\alpha_{zz}\approx\frac{M_z(+\Delta E)-M_z(-\Delta E)}{2\Delta E} \tag*{(17)}

with ΔE=0.1\Delta E=0.1 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 ee \AA{} per V/\AA{} for the variance term βVar⁡(Mz)\beta\operatorname{Var}(M_z), so ∂⟨Mz⟩/∂E0∣E0=0\left.\partial\langle M_z\rangle/\partial E_0\right|_{E_0=0} is 10.7 ee \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 Lw=31.34L_{\mathrm{w}}=31.34 \AA{} we obtain an estimate of 14.15 ee\AA{} per V/\AA{}. However, with Lw=22.0L_{\mathrm{w}}=22.0 \AA{}, Eq. (15) gives 9.93 ee\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:

∂⟨Mz⟩∂E0∣31 A˚=∂⟨Mz⟩∂E0∣22 A˚+ε0AΔLw(1−1ε),(18)\left.\frac{\partial\langle M_z\rangle}{\partial E_0}\right|_{31\ \text{\AA}}=\left.\frac{\partial\langle M_z\rangle}{\partial E_0}\right|_{22\ \text{\AA}}+\varepsilon_0A\Delta L_{\mathrm{w}}\left(1-\frac{1}{\varepsilon}\right), \tag*{(18)}

with ΔLw=9.34\Delta L_{\mathrm{w}}=9.34 \AA{} the difference between the water thicknesses of the two cells, which gives 14.90 ee\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 ee\AA{} per V/\AA{}, between the two estimates, whereas the short-range model gives 25.74 ee\AA{} per V/\AA{}.

Work function of a metal–water–countercharge cell

In this section we keep the elementary charge ee explicit, so that the work function Φ\Phi (or absolute electrode potential) is an energy; the main text uses units with e=1e = 1. We consider a metal slab covered by a water film, with an excess charge qq on the electrode and a counter charge −q-q above the water (Figure S1); zz is the surface normal and AA the lateral area. The Fermi level is the electrochemical potential of the electrons, ϵF=μm−eϕm\epsilon_{\mathrm{F}} = \mu^{\mathrm{m}} - e\phi^{\mathrm{m}}, where μm\mu^{\mathrm{m}} is their chemical potential in the metal and ϕm\phi^{\mathrm{m}} 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 ψ\psi:

Figure S1. Planar-averaged potential energy of an electron, −eφ(z)-e\varphi(z), in a Pt(111)/water cell from Ref. [41] with a solvated jellium counter charge (GPAW-SJM), at q=−0.25q=-0.25 (gray) and +0.25e+0.25e (black), with the vacuum level above the jellium (ψ\psi) taken as zero. Part of the structure is shown on top and its scale matches the zz-axis. Dashed lines mark the Fermi level ϵF=μm−eφm\epsilon_{\mathrm{F}}=\mu_{\mathrm{m}}-e\varphi_{\mathrm{m}} 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.

Φ=−eψ−ϵF=−μm+eχ,χ=ϕm−ψ,(19)\Phi= -e\psi- \epsilon_{\mathrm{F}} = -\mu^{\mathrm{m}} + e\chi,\qquad\chi= \phi^{\mathrm{m}} - \psi, \tag*{(19)}

with χ\chi the potential drop from the metal to the vacuum. Because q=−eΔNeq = -e\Delta N_{\mathrm{e}} and ϵF=∂U/∂Ne\epsilon_{\mathrm{F}} = \partial U/\partial N_{\mathrm{e}}, taking ψ=0\psi= 0 gives ∂U/∂q=−ϵF/e=Φ/e\partial U/\partial q = -\epsilon_{\mathrm{F}}/e = \Phi/e, which is Φ=∂U/∂q\Phi= \partial U/\partial q of the main text with e=1e = 1.

To obtain χ\chi, let ρ(z)=A−1∫ρ(r) dx dy\rho(z) = A^{-1}\int\rho(\mathbf{r})\,dx\,dy be the charge density averaged over the lateral cell, including the electrode charge and the counter charge. We take zmz_{\mathrm{m}} inside the metal and zvacz_{\mathrm{vac}} in the vacuum above the counter charge. The field vanishes at both points, so by Gauss’s law the enclosed charge is zero: ∫zmzvacρ(z′) dz′=0\int_{z_{\mathrm{m}}}^{z_{\mathrm{vac}}}\rho(z')\,dz' = 0. To obtain χ=ϕm−ψ\chi= \phi^{\mathrm{m}} - \psi, we integrate Poisson’s equation twice between zmz_{\mathrm{m}} and zvacz_{\mathrm{vac}}

−χ=ψ−ϕm=−1ϵ0∫zmzvacdz∫zmzρ(z′) dz′=1ϵ0∫zmzvaczρ(z) dz,(20)-\chi= \psi- \phi^{\mathrm{m}} = -\frac{1}{\epsilon_0}\int_{z_{\mathrm{m}}}^{z_{\mathrm{vac}}} dz\int_{z_{\mathrm{m}}}^{z}\rho(z')\,dz' = \frac{1}{\epsilon_0}\int_{z_{\mathrm{m}}}^{z_{\mathrm{vac}}}z\rho(z)\,dz, \tag*{(20)}

where we switched the order of integration over the triangular integration domain. The integral is the zz-component of the dipole moment of the cell per unit area: χ=−Mz/(ϵ0A)\chi= -M_z/(\epsilon_0 A).

We compare with a reference system, such as the bare slab at q=0q = 0, with work function Φref\Phi_{\mathrm{ref}}. For a slab thick enough that its interior is unperturbed, μm\mu^{\mathrm{m}} is the same in both, so Φ−Φref=eΔχ\Phi- \Phi_{\mathrm{ref}} = e\Delta\chi, which depends on the charge density difference Δρ\Delta\rho with respect to the reference system. Since Δρ\Delta\rho vanishes below zmz_{\mathrm{m}}, the integral can be extended over the whole cell:

Φ=Φref−eϵ0A∫(z−zm)Δρ(r) d3r.(21)\Phi= \Phi_{\mathrm{ref}} - \frac{e}{\epsilon_0 A}\int(z-z_{\mathrm{m}})\Delta\rho(\mathbf{r})\,d^3r. \tag*{(21)}

We partition Δρ\Delta\rho into contributions Δρi\Delta\rho_i of the atoms and the counter charge ρcc\rho_{\mathrm{cc}}, Δρ=∑iΔρi+ρcc\Delta\rho= \sum_i\Delta\rho_i + \rho_{\mathrm{cc}}. For atom ii at height ziz_i, we define its charge and dipole,

Δqi=∫Δρi d3r,Δμi,z=∫(z−zi)Δρi d3r,(22)\Delta q_i = \int\Delta\rho_i\,d^3r,\qquad\Delta\mu_{i,z} = \int(z-z_i)\Delta\rho_i\,d^3r, \tag*{(22)}

with ∑iΔqi=q=−∫ρcc d3r\sum_i\Delta q_i = q = -\int\rho_{\mathrm{cc}}\,d^3r. Splitting z−zm=(z−zi)+(zi−zm)z-z_{\mathrm{m}} = (z-z_i)+(z_i-z_{\mathrm{m}}) for each atom, and introducing the centroids of the interfacial charge and of the counter charge,

zˉq=1q∑iziΔqi,zˉcc=−1q∫zρcc d3r,(23)\bar{z}_q = \frac{1}{q}\sum_i z_i\Delta q_i,\qquad\bar{z}_{\mathrm{cc}} = -\frac{1}{q}\int z\rho_{\mathrm{cc}}\,d^3r, \tag*{(23)}

the integral in Eq. (21) becomes

∫(z−zm)Δρ d3r=∑i[Δμi,z+(zi−zm)Δqi]−q(zˉcc−zm)=ΔMz−qd,(24)\int(z-z_{\mathrm{m}})\Delta\rho\,d^3r = \sum_i\left[\Delta\mu_{i,z}+(z_i-z_{\mathrm{m}})\Delta q_i\right]-q(\bar{z}_{\mathrm{cc}}-z_{\mathrm{m}})=\Delta M_z-qd, \tag*{(24)}

with ΔMz=∑iΔμi,z\Delta M_z=\sum_i\Delta\mu_{i,z} and d=zˉcc−zˉqd=\bar{z}_{\mathrm{cc}}-\bar{z}_q; the terms in zmz_{\mathrm{m}} cancel because the electrode charge and the counter charge are equal and opposite. Hence

Φ=Φref−eΔMzϵ0A+eqdϵ0A,(25)\Phi= \Phi_{\mathrm{ref}}-\frac{e\Delta M_z}{\epsilon_0 A}+\frac{eqd}{\epsilon_0 A}, \tag*{(25)}

which is Eq. (3) of the main text with e=1e = 1. The atomic dipoles Δμi,z\Delta\mu_{i,z} contain the electronic rearrangement of both the metal surface and the water. The partition into atomic contributions is not unique; an MLIP that predicts Φ\Phi 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 γ\gamma and β\beta that take the conditioning parameter qq or σ=q/A\sigma= q/A as input,

hi←(1+γ(q))⊙hi+β(q),(26)\mathbf{h}_i \leftarrow(1+\gamma(q)) \odot\mathbf{h}_i+\beta(q), \tag*{(26)}

where ⊙\odot indicates element-wise multiplication. The work function Φ=∂U/∂q\Phi= \partial U/\partial q is obtained by automatic differentiation in the same backward pass as the forces. The zz column of the Born effective charge tensors, Zi,z∗=ε0A ∂Fi/∂qZ^{*}_{i,z}=\varepsilon_0 A\,\partial F_i/\partial q, 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-qq production runs with ASE reach 3.0 ns/day (70 steps per second) on one NVIDIA A100.

GRACE-RAZOR

The response analysis in zz-orientation (RAZOR) approach [29, 41] uses two backbones, one for U0U_0 and one for Φ\Phi. 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 μi\boldsymbol{\mu}_i from the l=1l=1 last-layer features and contracting them with D=(q/A)z^\mathbf{D}=(q/A)\hat{\mathbf{z}}, the atomic contribution to the energy becomes

Ui=U0,i+q(ΦrefN−μi,zε0A)+12q21C0N,(27)U_i=U_{0,i}+q\left(\frac{\Phi_{\mathrm{ref}}}{N}-\frac{\mu_{i,z}}{\varepsilon_0 A}\right)+\frac{1}{2}q^2\frac{1}{C_0N}, \tag*{(27)}

where NN 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. Φref\Phi_{\mathrm{ref}} is degenerate with a constant in ∑iμi,z\sum_i\mu_{i,z}; we fix it at the clean Pt(111) value of 5.75 V, and learn only C0−1C_0^{-1}, initialized at the mean ∂2U/∂q2\partial^2 U/\partial q^2 of the training set. The area AA 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) (9.749×8.443 A˚9.749 \times8.443\,\text{\AA}) or stepped Pt(332) (13.200×8.443 A˚13.200 \times8.443\,\text{\AA}, one step per cell), in a cell of 34.81 Å along zz. 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 h∈[5,10]h \in[5, 10] Å above the surface, with the lateral position sampled from a uniform distribution. We keep the 964 frames with h≤9h \le9 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 θ\theta at fixed position; on Pt(332), the water is moved by xx 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 4×4×14 \times4 \times1 (Pt(111)) and 3×4×13 \times4 \times1 (Pt(332)). Each frame carries a charge qq drawn uniformly from [−0.030,+0.025][-0.030, +0.025] e on Pt(111). For Pt(332), we use the same surface-charge density range. The counter charge −q-q 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, ∂U/∂q\partial U/\partial q from the energies matches this work function to 0.5 mV, so the label Φ\Phi 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-qq 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-qq follows the step at both heights.

surfacemodelEE (meV / atom)FF (meV / Å)Φ\Phi (mV)curves, Φ\Phi (mV), h=5h = 5curves, Φ\Phi (mV), h=7h = 7
Pt(111)LOREM-qq0.173.51.93.71.2
Pt(111)GRACE-RAZOR0.143.12.92.53.2
Pt(111)GRACE-qq0.5246.647076666
Pt(332)LOREM-qq0.102.41.83.10.9
Pt(332)GRACE-RAZOR0.102.57.716.49.9
Pt(332)GRACE-qq0.4040.3386432667

Table S3. Errors of the models in Figure 2: RMSE on the 96 validation frames per surface, and RMSE of Φ\Phi on the curves drawn in Figure 2, at h=5h = 5 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 3×43 \times4 Pt(111) slab (36 Pt) with 24 water molecules, lateral area 82.31 Å2^2, at bias charges ∣q∣≤1 e|q| \le1\,e, 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 0.5ee 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 qMD=+1eq_{\mathrm{MD}}=+1e, the window is centered near +0.75e+0.75e, and at +1e+1e 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 q≤+0.75eq\leq+0.75e.

Figure S2. Work function of the Pt(111)–water models. (a) ΦSJM\Phi_{\mathrm{SJM}} of single structures from Ref. 41, each equilibrated by MD at a charge qMDq_{\mathrm{MD}} (dots) and recomputed with GPAW-SJM at charges from −1.5-1.5 to +1.5e+1.5e. The dotted line marks the highest training charge, +0.75e+0.75e. (b) ⟨ΦSJM⟩(q)\langle\Phi_{\mathrm{SJM}}\rangle(q) from MD with the models trained on ΦSJM\Phi_{\mathrm{SJM}} and from AIMD—the same as Figure 3h. The work function of the GRACE single-qq series is computed with LOREM-qq. (c) ⟨Φfd⟩(q)\langle\Phi_{\mathrm{fd}}\rangle(q) from MD (here, two replicas only) with the same architectures trained on Φfd\Phi_{\mathrm{fd}} (same symbols as in b), with the AIMD converted to Φfd\Phi_{\mathrm{fd}} by subtracting the average offset δ(q)=ΦSJM−Φfd\delta(q)=\Phi_{\mathrm{SJM}}-\Phi_{\mathrm{fd}} (Sec. S3B). At +1e+1e, both GRACE-RAZOR replicas and one LOREM-qq 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. ΦSJM\Phi_{\mathrm{SJM}} is the work function returned by GPAW-SJM. Φfd\Phi_{\mathrm{fd}} is ∂U/∂q\partial U/\partial q obtained by finite differences of the energy on a charge stencil of ±0.25e\pm0.25e, which Ref. 41 uses for its labels. For an energy surface U(R,q)U(R,q), both should be equal, but they differ by δ=ΦSJM−Φfd≈0.3 V\delta=\Phi_{\mathrm{SJM}}-\Phi_{\mathrm{fd}}\approx0.3\ \mathrm{V}. A model with Φ=∂U/∂q\Phi=\partial U/\partial q trained on ΦSJM\Phi_{\mathrm{SJM}} 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 q=−1q=-1 to +1e+1e, we recomputed ∂U/∂q\partial U/\partial q with a stencil of ±0.05e\pm0.05e and obtained δ=0.28±0.09 V\delta=0.28\pm0.09\ \mathrm{V} (standard deviation over structures), against 0.27±0.08 V0.27\pm0.08\ \mathrm{V} from the ±0.25e\pm0.25e labels of the same structures. We have not identified the origin of δ\delta; 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 δ\delta change by only 0.03±0.02 V0.03\pm0.02\ \mathrm{V} per ee.

We train the models of the main text on ΦSJM\Phi_{\mathrm{SJM}}, the quantity reported by the AIMD reference. Figure S2,Figure S2 compares them with the same architectures trained on Φfd\Phi_{\mathrm{fd}}. For the comparison in panel c, the AIMD is converted to Φfd\Phi_{\mathrm{fd}} by subtracting the measured δ(q)\delta(q). Trained on ΦSJM\Phi_{\mathrm{SJM}}, both models follow the AIMD within 0.1 V0.1\ \mathrm{V}, as shown in the main text (Figure 3). Trained on Φfd\Phi_{\mathrm{fd}}, both lie 0.150.15–0.3 V0.3\ \mathrm{V} 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 q≤+0.75eq\leq+0.75e of Ref. 41. The LOREM-qq model most accurately predicts the work function and Born effective charges.

model, Φ\Phi labelEE (meV/atom)FF (meV/Å)Φfd\Phi_{\mathrm{fd}} (V)ΦSJM\Phi_{\mathrm{SJM}} (V)Z∗Z^{*} (ee)
LOREM-qq, ΦSJM\Phi_{\mathrm{SJM}}0.5525.50.3150.0710.014
GRACE-RAZOR, ΦSJM\Phi_{\mathrm{SJM}}0.4719.20.3210.0860.019
LOREM-qq, Φfd\Phi_{\mathrm{fd}}0.5425.50.0860.3170.015
GRACE-RAZOR, Φfd\Phi_{\mathrm{fd}}0.4418.90.0950.3130.019

Table S4. RMSE on the 395 structures with q≤+0.75eq \le+0.75e of the test set of Ref. 41. Φ\Phi errors shown against both labels, with the label each model was trained on in bold. Z∗Z^{*}: the Born effective charges along the surface normal, Ziα∗=Aε0 ∂Fiα/∂qZ^{*}_{i\alpha}=A\varepsilon_{0}\,\partial F_{i\alpha}/\partial q, 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 xx and yy, 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-qq reference

The single-qq 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 ee in steps of 0.25 ee, 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 < Φ\Phi < 7 V are excluded from training, as they likely exhibit dielectric breakdown. MD with the first single-qq models showed adsorbed OH or O at negative charges, which is unphysical. We therefore ran one additional round of active learning, sampling the single-qq trajectories in the same way.

The production single-qq MD was run with the same settings as the charge-conditioned GRACE-qq models (Sec. S3D), with four 250 ps replicas per charge. As these models have no work-function output, the potential of the single-qq series is computed with LOREM-qq on its frames.

References

  1. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [43]I.-C. Yeh and M. L. Berkowitz, Ewald summation for systems with slab geometry, J. Chem. Phys. 111, 3155 (1999).DOI
  44. [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. [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. [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. [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. [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. [49]G. Kresse and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Physical review B 59, 1758 (1999).DOI
  50. [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. [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. [52]H. J. Monkhorst and J. D. Pack, Special points for Brillouin-zone integrations, Physical review B 13, 5188 (1976).
  53. [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. [54]G. Bussi, D. Donadio, and M. Parrinello, Canonical sampling through velocity rescaling, The Journal of chemical physics 126 (2007).DOI
  55. [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. [56]M. Schlipf and F. Gygi, Optimization algorithm for the generation of oncv pseudopotentials, Computer Physics Communications 196, 36 (2015).DOI
  57. [57]D. R. Hamann, Optimized norm-conserving Vanderbilt pseudopotentials, Physical Review B—Condensed Matter and Materials Physics 88, 085117 (2013).DOI
  58. [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. [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. [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. [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. [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. [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).

Paper details

vlnc:W_n5t5qn82
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents