WorksPhysics and Astronomy
When do machine-learning potentials need a configuration-dependent kernel? Learning long-range interactions with Fourier neural operators
Abstract
Long-range interactions in machine-learning interatomic potentials are commonly treated by coupling learned sources through either a prescribed electrostatic kernel or a learned but configuration-independent one. Whether such a kernel suffices depends on the underlying screened interaction. We address this question by using Fourier neural operators (FNOs) to learn how latent source fields, built from the descriptors of a local backbone (here MACE), interact across the simulation cell. For a synthetic ionic system with tunable screening, a prescribed Coulomb kernel underestimates the long-wavelength charge fluctuations by 43% under strong screening, whereas the FNO reproduces them within 3% and recovers 95-99% of the variance of the residual left by the local backbone. A linear FNO, which learns a configuration-independent kernel, performs equally well in the homogeneous case. Under inhomogeneous screening, rearranging the remote medium modulates the coupling between distant ions, a dependence that no configuration-independent kernel can represent. Among the models tested, only the nonlinear FNO responds to this modulation, and its force errors are 27-41% lower than those of configuration-independent models. In first-principles benchmarks of liquid water, lithium lanthanum zirconate, and Au2/MgO(001), MACE-FNO is comparable to a prescribed-kernel model for the bulk systems and more accurate for the interface. It further captures the reversal of the Au2 wetting preference induced by remote Al dopants, with deviations below 5 meV per cell from density functional theory. These results indicate that a configuration-dependent kernel, as learned by FNOs, becomes necessary once the screening medium varies, even though the energy remains quadratic in the charges for a fixed medium.
Introduction
Machine-learning interatomic potentials (MLIPs) have made accurate atomistic simulations accessible at length and time scales beyond first-principles electronic-structure calculations. A large class of neural-network MLIPs constructs the energy from local atomic environments. In message-passing models such as MACE [1], information is exchanged within a finite receptive field determined by the radial cutoff and the number of interaction layers. The spatial information available to each atom-centered representation is restricted to this finite receptive field.
This finite range poses a strong limitation when local MLIPs are applied to systems where long-range interactions become important, such as polar liquids, ionic solids, interfaces, and adsorbate systems [2, 3, 4]. Their use can also be problematic when nonlocal correlation arises, e.g., through a remote charge state or collective polarization beyond the receptive field of the model.
Long-range-aware MLIPs address this limitation through several complementary strategies. Environment-dependent atomic charges can be learned from reference charges or molecular moments and used in explicit electrostatic energies [13, 14, 15]. Global charge redistribution can instead be obtained through charge equilibration schemes [2, 16, 17]. Electronic response can also be represented through learned maximally localized Wannier centers, as in DPLR [18], or through a self-consistently updated long-range field, as in SCFNN [19]. Approaches that do not assign explicit atomic charges include long-distance equivariant descriptors [20, 21], Ewald-based long-range message passing [10, 22], density-based long-range descriptors [23], and attention-based global relay nodes [24].
A recent long-range-aware MLIP family learns environment-dependent sources from local descriptors and propagates them through a prescribed electrostatic kernel. Within this family, latent Ewald summation (LES) maps local invariant features to latent charge-like variables and evaluates their global coupling through a prescribed Coulomb or Ewald kernel [4]. The framework has been used to augment several local MLIPs and to recover polarization-related response from energy–force training [5, 25, 26]. Its source–kernel factorization is compact and interpretable when the missing interaction is well described as electrostatics. Park et al. extend this framework to latent atomic dipoles and quadrupoles, and induced charges and dipoles describe nonlocal charge redistribution and polarization through a non-self-consistent linear-response treatment [6]. The qNEP framework applies the same broad source–kernel principle within the neuroevolution potential (NEP) formalism [27]. It predicts environment-dependent atomic charges and couples them through Ewald summation or particle–particle particle–mesh electrostatics [7]. LES and qNEP demonstrate that local chemical descriptors can supply the sources of an explicit long-range operator.
| Method | Sources | Global propagation | Configuration-dependent propagation | Energy in sources | Coupling to local model | Training targets |
| LES [4, 5] | multichannel latent charges | prescribed Ewald | no | quadratic | additive | , () |
| Extended LES [6] | latent multipoles | prescribed Ewald | inductiona | quadratic | additive | , () |
| qNEP [7] | learned atomic charges | prescribed Ewald/PPPM | no | quadratic | additive | , , , b (BEC) |
| RSNN (RSP) [8] | latent charges | learned kernel of | no | quadratic | additive | , , , |
| SOG-Net [9] | multichannel latent charges | learned sum of Gaussians per channel | no | quadratic | additive | () |
| Ewald MP [10] | atomic embeddings | learned filters | yesc | nonquadratic | interleavedd | () |
| Neural P³M [12] | mesh features | learned nonlinear Fourier layer per block | yes | nonquadratic | interleaved | () |
| MACE–FNO, linear | latent source fields | learned spectral kernel with channel mixing | no | quadratic | additive | , |
| MACE–FNO, nonlinear | latent source fields | stacked nonlinear Fourier layers | yes | nonquadratic | additive | , |
Table 1. Comparison of long-range MLIPs that couple learned charges, multipoles, or latent source features through global operators. Charge-equilibration models, which determine the sources self-consistently, are not listed. Configuration-dependent propagation indicates whether the effective propagation between two sources can change with the rest of the configuration through induction or nonlinear updates for given sources in a fixed cell. The long-range energy is classified as quadratic or nonquadratic in the source amplitudes at fixed positions and response coefficients. Interleaved models, unlike additive ones, feed the propagated features back into message passing. Training targets are energies (), forces (), stresses (), total-charge penalties (), and Born effective charges (BEC), with parentheses marking targets used only for some systems. The last two rows correspond to the present work.
While a prescribed electrostatic kernel provides a strong physical prior, it confines global propagation to a chosen functional form. In a first-principles description, the bare interaction is Coulombic, and environment-dependent charges can partly represent the screening that emerges from electronic polarization [28]. The effective interaction between learned sources, however, need not take the Coulomb form between point charges. This is the case, e.g., for metallic Thomas–Fermi screening, charge transfer at interfaces, and implicit-solvent models. Learned propagation kernels offer an alternative. The reciprocal-space neural network (RSNN) couples latent atomic charge features through an isotropic reciprocal-space kernel represented by a fully connected network, which describes mixed Coulomb and Lennard-Jones interactions within a single learned form [8]. SOG-Net combines latent sources with Fourier-space kernels represented by sums of Gaussians for which the amplitudes and widths are learned to incorporate the long-range effects [9]. The long-range energy in both approaches is quadratic in the latent sources. A learned Fourier-space kernel also underlies Neural P³M, which uses nonlinear Fourier neural operator (FNO) blocks to account for global interactions [12]. Table I summarizes these approaches and how they propagate the local sources through global operators.
Building on the principle of learned global propagation, here we formulate a nonlocal energy correction to a complete local backbone using the FNO [29, 30] to propagate latent source fields derived from local atomic descriptors through spectral mixing and pointwise transformations. Unlike Neural P³M, the propagated fields do not feed back into the atomic message-passing layers, and the learned FNO supplies an additive correction to the local backbone. Compared with SOG-Net, which propagates each source channel independently through a sum of Gaussians, the FNO mixes the source channels at each wavevector and learns its spectral coupling for every wavelength without an assumed radial form. Stacked Fourier layers with pointwise nonlinearities further give rise to an energy that is nonquadratic in the sources, so that the effective propagation between the sources of two atoms is configuration dependent. For slabs, the operator couples sources at different heights without imposing periodicity along the surface normal. We first construct a synthetic ionic system with a Yukawa interaction to establish what the FNO learns from the residual interactions missed by the local backbone. We then introduce spatially inhomogeneous screening to assess the role of configuration-dependent propagation when the effective interaction depends on the surrounding medium. Finally, we evaluate energy and force accuracy for liquid water, (LLZO), and an dimer on . The latter also tests whether the learned correction captures the effect of remote dopants on the wetting preference of the adsorbate. Our results show that a linear FNO with a single configuration-independent propagation kernel already suffices to describe nonlocal interactions under homogeneous screening, without prior knowledge of their functional form. The nonlinear FNO, the effective propagation of which can adapt to the configuration, remains accurate for screening lengths between 2 and 20 Å and from weakly to strongly inhomogeneous screening.
Methods
MACE–FNO architecture
The central idea is to augment a local MLIP with a nonlocal energy correction learned through an FNO. Here we use MACE as a representative backbone to predict the local energy . The construction can be coupled to other local MLIPs as long as they provide differentiable and rotationally invariant atomic descriptors. The residual energy term is recast into a correction term from the learned field response. For a given atomic configuration, the total energy can be written as a sum of the two contributions
Figure 1(a) illustrates our MACE–FNO architecture. Following the particle–mesh method, we deposit atomic sources on a regular grid and evaluate global coupling through Fourier-space operations [31, 32]. For atom , MACE provides a rotationally invariant descriptor that encodes its local chemical environment. A trainable source head maps this descriptor to scalar source amplitudes , one per channel . These amplitudes are then distributed on a mesh using smooth cubic B-spline weights centered on the atomic position. Summing these contributions produces a source field at a mesh position . Motivated by overall electrostatic neutrality, we constrain each latent source field to have a zero mesh integral.
Figure 1. Schematics of the MACE–FNO architecture and the two operator variants. (a) Local MACE descriptors are mapped to latent atomic sources and deposited on a mesh. The FNO transforms the source fields into response fields, and their contraction gives the energy correction. Solid arrows show forward propagation, and dashed red and blue arrows indicate frozen-MACE and joint training, respectively. (b) Linear FNO applies a single spectral operation to the source channels. No pointwise branch or nonlinear activation is used. The inset illustrates the retained modes along one Fourier direction. (c) Nonlinear FNO maps the sources to hidden channels, applies successive Fourier layers, and projects back to the response channels. The dashed box shows a single Fourier layer, in which the spectral and pointwise branch outputs are summed before applying the GELU activation.
The FNO transforms the combined source field into a response field through a learned operator
where is the cell matrix, and denotes the learnable operator parameters. The response field contains channels, matching those of the source fields. We build by passing the output fields of one Fourier layer to the next. In layer , the input is a real-space vector containing feature values at mesh position . In particular, the first layer receives hidden fields obtained by a learned, pointwise mapping of the source channels . Within a single Fourier layer, spectral mixing proceeds sequentially through a Fourier transform, learned channel mixing, and an inverse transform
where the learned matrix combines the hidden channels at a reciprocal vector , and is the resulting real-space spectral response. This procedure is illustrated schematically in Figure 1(c). In real space, this operation corresponds to a convolution throughout the periodic unit cell. The response in one region can therefore depend on fields beyond the MACE receptive field. This provides the mechanism for learning nonlocal interactions using the FNO.
To complement the nonlocal spectral response, we further consider a pointwise contribution , as shown in Figure 1(c). is a learned matrix that is shared across all mesh points within a specific Fourier layer . With the standard Fourier-layer update, the input to layer is
where applies the Gaussian error linear unit (GELU) activation componentwise to the combined contributions.
For two-dimensional systems such as , we apply the Fourier transform only to the periodic in-plane coordinates and jointly mix all hidden channels and height points along the surface normal. Writing for the th height point, the spectral response becomes
where and label hidden channels, and is the number of height points. We refer to this formulation as full -mixing. The dependence of on both heights and wavelength allows the coupling between different slab heights to vary with the lateral length scale. An inverse in-plane transform gives for Eq. (4), with mixing channels at each mesh point. No Fourier transform or periodic wrapping is applied along . Alternatively, we separate channel and height mixing so that
where mixes channels at each height and is shared across all heights, and the learned matrices mix heights within each channel and are independent of the in-plane wavevector. Within one Fourier layer, the channel-wise mixing of Eq. (6) couples different heights only within the same channel, while a height in one channel couples to a different height in another channel only through successive Fourier layers. Appendix A 4 b analyzes the computational complexity of the two different mixing schemes, and Appendix B 3 compares the numerical accuracy and GPU inference costs.
Once the final response field is produced, the energy correction can be obtained by contracting the source and response fields over the mesh domain
where the dot product sums over the channels. We emphasize that the source () and response fields () are latent variables and need not represent physical charge densities or electrostatic potentials. Rather than prescribing electrostatics explicitly as in LES and qNEP, our approach learns both the sources and the global response operator. The total forces follow from
where is the position of atom .
The FNO hitherto defines a nonlinear mapping from the source to response fields due to the nonlinear activation between the Fourier layers. By replacing the stacked Fourier layers and their input and output mappings with a single spectral operation directly on the source channels [cf. Figure 1(b)], we can construct a linear FNO, denoted , for which the response satisfies , where is a learned matrix. The linear therefore omits the pointwise branch and nonlinear activation of the field operator. A linear operator with more than one Fourier layer would not be more expressive, since stacked layers without activation compose into a single wavevector-dependent channel-mixing matrix, apart from a local term that the backbone can represent. The difference in accuracy between the linear and nonlinear FNO therefore cannot be attributed to the depth of the operator. As the source head remains nonlinear while the contraction follows Eq. (7), the energy correction obtained with a linear FNO is quadratic only in the source fields.
The MACE–FNO optimization spans the parameter space defined by , , and , denoting the parameters of MACE, the source head, and , respectively. Within this space, we consider two training strategies initialized from a pretrained MACE model and fitted to reference energies and forces. In frozen-MACE training, we keep fixed and update and through backpropagation. This confines optimization to the source head and FNO, isolating the learned correction relative to an unchanged local baseline. In joint MACE–FNO training, we update all three parameter sets together by backpropagating the same energy–force loss through the full coupled model. Here, the MACE parameter gradients receive contributions from both its local energy prediction and the descriptor-dependent FNO correction. Through the latter path, the FNO correction can reshape the local descriptors as the local representation and global operator adapt together. These alternatives correspond to the two return routes in Figure 1(a). In the tables, the superscript in – and – denotes a frozen one- or two-interaction MACE backbone. FNO-augmented entries without this superscript denote joint training.
The source construction, complete layer equations, and derivative expressions are given in Appendix A. The geometry-specific operators and symmetry treatment are detailed in Appendix A 4.
Spectral-response diagnostic
Although is not tied to a prescribed kernel, the learned response can still be characterized by the changes in the FNO energy upon Fourier perturbations to the latent source fields. For a nonzero reciprocal vector , we construct a cosine probe at that wavevector [33, 34], and add it to the deposited fields
where contains the perturbation amplitudes of the latent source channels. The probe has a zero mesh integral and is normalized to an RMS amplitude of one, preserving the zero mesh integral of the source fields. The learned response then follows from the channel-space energy curvature
where and label source channels, and is the cell volume. For a nonlinear FNO, this describes the energy curvature around the supplied fields and may vary between configurations.
Diagonalizing gives eigenvalues and corresponding eigenvectors that specify normalized combinations of source channels. Each eigenvalue is the quadratic energy cost per unit volume for a perturbation along its channel combination. We follow the largest positive eigenvalue , the strongest positive curvature at wavelength , and discard wavevectors at which no eigenvalue is positive. The radial dependence of this response can be summarized by the approximate power law
where , is a fitted prefactor, and is the spectral exponent over the chosen wavevector range. It is instructive to compare the spectral exponent with the wavevector dependence of a reference interaction, e.g., for a pure three-dimensional Coulomb kernel.
We apply the spectral analysis to the synthetic benchmark of Sec. III, for which the reference kernel is known. For the first-principles benchmarks, no reference kernel is available, and the fitted exponent of the latent response cannot be checked against the interaction it represents.
The fit targets the lowest nonzero wavevectors where the accessible wavelengths are among the longest. We express wavevector magnitudes as with . Within a shell of equal wavevector magnitude , we average the positive response to obtain . An unweighted linear regression of against yields as the negative slope. The upper cutoff admits the three lowest shells at . We average the structure-level exponents within a checkpoint before calculating means and sample standard deviations across training seeds. Note that these three-point fits are pertinent to the accessible low- response and they do not necessarily establish the asymptotic limit. Appendix A § 1.9.6 gives the finite-difference procedure and the detailed sampling and regression settings.
Implementation and training protocols
We implement MACE–FNO in PyTorch [35]. Following the equivariance treatment in EqGINO [36], we impose isotropic spectral coupling by sharing channel-mixing matrices among modes with the same physical wavelength (Appendix A 4). Our MACE–FNO code interfaces with ASE [37] for energy–force inference through a calculator interface.
The mode parameter includes integer indices along the Fourier-transformed directions. For orthorhombic cells, these indices give , where is the cell length along direction . The source-channel count specifies the number of deposited fields. The source head has 64 hidden units in all models. Appendix B 2 logs the sensitivity to the mesh resolution, mode setting, and source-channel count. Table II summarizes the principal FNO hyperparameters used for the synthetic and physical benchmarks.
| Parameter | Synthetic ionic system | Liquid water | LLZO | Au/MgO(001) |
| Mesh points | 44 | 24 | 32 | 24 |
| Mode parameter | 11 | 8 | 8 | 8 (in-plane) |
| Source channels | 4 | 16 | 16 | 16 |
| Fourier layers | 2 | 2 | 4 | 4 |
Table II. Principal FNO hyperparameters for the four benchmark systems. The synthetic settings apply to all three screening parameters . denotes the number of mesh points per periodic direction, and the number of Fourier layers. Bulk meshes contain points, while the slab mesh has points, with the last dimension resolving height.
For a given dataset and interaction depth, the frozen and joint runs start from the same MACE checkpoint and share the data split, mesh, and loss normalization. We perform independent FNO training runs per backbone with five random seeds. We report the mean and sample standard deviation of test root-mean-square errors (RMSEs) obtained by evaluating the five trained models separately. These statistics describe training variability for a given starting MACE model.
The source head and FNO are optimized with Adam [38]. We evaluate validation errors in batches of eight structures and select checkpoints by the combined normalized energy and force error. The step budgets, batch sizes, loss normalization, learning-rate schedule, and selection rules are given in Appendix A § 1.9.5. Double-precision arithmetic is used throughout training and inference.
Homogeneous screening beyond a prescribed kernel
Drawing on the homogeneous continuum description of screening in many bulk materials, we first construct a synthetic ionic system with spatially uniform screening. The system comprises disordered, charge-neutral NaCl-like configurations interacting through a short-range Lennard–Jones (LJ) potential and a Gaussian-smeared Yukawa potential with a tunable inverse screening length . Screening is made homogeneous by construction since is spatially constant and independent of the ionic configuration.
The short-range interaction of the synthetic systems is described by a smoothly truncated LJ interaction
where defines the well depth, and is the characteristic LJ distance that places the potential minimum at . The switching function is a quintic smoothstep that truncates the LJ interactions smoothly to zero beyond a cutoff
where , and following the MACE backbone. The switching function guarantees that the energy and its first two derivatives vanish continuously at the cutoff. As a result, the short-range interaction is confined by construction to the MACE receptive field without introducing a discontinuity.
The Gaussian-smeared Yukawa interaction is specified by the reciprocal-space kernel
where refers to the inverse screening length, and is a smearing parameter that regularizes the interaction at contact. The pointwise logarithmic slope of Eq. (14) has the analytical form
This exponent provides an analytical reference, the wavevector dependence of which can be varied through . Without appreciable Gaussian damping, it crosses approximately from zero when to two when . Here we use , , and , which correspond to screening lengths of 20 \AA, 5 \AA, and 2 \AA that straddle the backbone cutoff. This variation tests whether a learned operator can follow the change in screening without a prescribed kernel. LES keeps its Coulomb propagation kernel fixed, although its learned charges and local energy contribution can adapt to the different training targets.
To keep the nonlocal energy-variation scale constant across , we multiply the nonlocal energies and forces by a dimensionless factor , leaving the LJ contribution unchanged. The factors are 1.8512, 2.0483, and 2.9920 for , , and , respectively. They fix the standard deviation of the nonlocal energy per atom across the generated configurations at 5 meV atom. This rescaling changes the interaction strength but not its wavevector dependence or logarithmic slope. With this normalization, we calculate the periodic Yukawa energy as
Here is the Fourier amplitude of the reference charges at positions , and is the Coulomb unit-conversion factor. The sum includes both and up to Å. The term removes the unscaled self-interaction energies of the Gaussian-smeared charges.
The total reference energy of the synthetic ionic system is
where . The Yukawa contribution follows Eq. (16), which already accounts for the scaling through . The residual energy forms the target of the FNO correction, which includes both missing nonlocal contributions and remaining errors of the local backbone.
We model 1 146 atoms with charges in a periodic cubic cell with a side length of Å. This cell size gives a smallest nonzero wavevector magnitude of Å. The corresponding wavelength is about seven times the descriptor cutoff of the backbone. From random initial positions, we generate 40 independent Langevin molecular-dynamics (MD) trajectories at 400 K driven by the unscaled Yukawa interaction with Å and , and sample 30 configurations per trajectory after a 3 ps burn-in. These trajectories are subsequently assigned to training, validation, and testing, resulting in 660, 240, and 300 configurations, respectively. No trajectory is shared between subsets. The trajectories are then relabeled with the scaled energies and forces evaluated at each . The atomic configurations and the splits are thus identical across the three datasets. We note that the datasets at Å and Å carry the suppressed long-wavelength charge fluctuations of the nearly Coulombic fluid as the configurations are sampled at Å. This is adequate for a static benchmark, whereas the MD simulations of Sec. IIID sample each interaction in its own ensemble.
Throughout this benchmark, we use a simple one-interaction MACE model with a cutoff of 4.5 Å as the local backbone. Its hidden representation contains 128 scalar () and 128 vector () channels. We adopt the FNO settings in Table II. The Fourier layers have a hidden-channel width of 32. Spectral mixing uses a single group and couples all hidden channels. The static energy–force benchmark is evaluated on a single mesh without interlacing.
To first order, the agreement of the FNO correction and the residual energy can be evaluated through the residual variance . We calculate the residual variance across the 300 test configurations using the five FNO seeds on the same frozen backbone. The FNO correction explains 99.3%, 98.2%, and 95.1% of the residual variance for Å, Å, and Å, respectively, thereby confirming that the FNO closely follows the residual energy as a whole. Nevertheless, such an agreement does not address the wavelengths at which the FNO recovers the residual. Knowing the reference charges of the synthetic system, we now proceed to project the residual and FNO energies onto reciprocal-space shell features.
Shell-resolved residual recovery
For configuration , we construct the shell features from the known reference charges,
where is the cell volume, is the reference charge of atom at position , and accounts for the Gaussian smearing. The shell groups nonzero wavevectors of equal magnitude . Only one member of a pair enters the sum, giving the prefactor instead of . Thus measures the weighted charge-density modulation at with the reference charges.
We express the Yukawa reference, the frozen-backbone residual, and the FNO corrections in a common shell feature representation
where identifies the energy being analyzed. We resolve the lowest shells individually and include the higher- reference contribution in up to the cutoff Å. For the Yukawa reference, the coefficient corresponds essentially to the scaled Yukawa kernel , and thus no fitting is required. For the residual and FNO energies, we determine , , and simultaneously by ordinary least squares over the same 300 test configurations. Note that the choice of energy reference does not affect the fitted shell coefficients as the energy offset is described by . Within this shell representation, measures the interaction strength remaining after subtracting the MACE backbone at . The corresponding coefficient measures the strength supplied by the FNO correction.
Figure 2 shows that the residual coefficients are substantial at small wavevectors and generally decrease towards high . To examine the extent to which the reference Yukawa coefficient can be captured by the MACE backbone alone, we introduce . At the lowest shell Å, the local backbone only recovers of the reference for Å, leaving the interactions largely unaccounted for at this small wavevector. On the other hand, the backbone absorbs 93.8% of the reference for the high- tail above the 15 lowest shells. Figure 2 hence provides direct evidence of a pronounced wavevector dependence of the local MACE backbone, i.e., the responses at short wavelengths are better described than those at longer wavelengths. This is further supported by the dependence of the residual, which shows that the local backbone is more adequate as the screening strengthens ( and for Å and Å, respectively).
Figure 2. Shell-resolved recovery of the frozen-MACE residual in the synthetic ionic system. Columns correspond to Å, Å, and Å. (a–c) Raw coefficients. The gray dashed line gives the reference Yukawa coefficient . Black circles are coefficients fitted to for the MACE backbone (seed 17), whereas filled blue and purple circles are the mean coefficients for FNO and FNO, respectively, over five seeds on this frozen backbone. Error bars denote one sample standard deviation for both variants. (d–f) Deviations from the residual, , for the two FNO variants, with error bars normalized by the same reference coefficient. The horizontal zero lines mark exact agreement with the residual. The coefficients are shown for the 15 lowest shells. Lines are guides to the eye.
The FNO coefficients closely follow the overall magnitude and wavevector dependence of the residual across the resolved shells, irrespective of whether the field operator is linear or nonlinear [cf. Fig. 2(d–f)]. By reproducing the departure of the residual from the Yukawa reference, both FNO variants complement the uneven local description at the varying wavelengths without explicitly prescribing the Yukawa kernel or supervising the shell coefficients. Over the three lowest shells, the nonlinear FNO gives relative RMS deviations from the residual coefficients of 2.8%, 1.2%, and 12.0% for Å, 0.2 Å, and 0.5 Å, respectively. The corresponding deviations for FNO are 2.7%, 1.7%, and 17.9%. The largest deviation occurs at Å, where the residual exhibits a pronounced non-monotonic shell dependence as shown in Figure 2(c).
Spectral response of latent fields
We now probe the latent-field energy curvature using the spectral diagnostic described in Sec. II B. The fit is performed over the three lowest shells, the wavevector magnitudes of which span 0.196–0.340 Å. Table III compares the resulting exponent with , the reference exponent obtained as the negative slope of a log–log fit of the analytical smeared Yukawa kernel in Eq. (14) over the same shells. The screening-induced modification in the wavevector dependence of the kernel is clearly visible through . Over the sampled low- range, is close to the unscreened Coulomb value of two at Å and falls below 0.5 for Å. The trend is in agreement with the pointwise analytical exponent given in Eq. (15) although here represents a fitted slope over a finite window.
| (Å) | |||||
| 0.05 | 1.97 | 1.000 | FNO | 2.10 ± 0.07 | 1.000 ± 0.000 |
| FNO | 2.07 ± 0.05 | 0.999 ± 0.001 | |||
| 0.2 | 1.28 | 0.997 | FNO | 1.46 ± 0.06 | 0.996 ± 0.003 |
| FNO | 1.42 ± 0.04 | 0.994 ± 0.001 | |||
| 0.5 | 0.47 | 0.986 | FNO | 0.60 ± 0.12 | 0.965 ± 0.021 |
| FNO | 0.66 ± 0.04 | 0.961 ± 0.009 |
Table III. Reference and latent-field spectral exponents and log–log fit scores for the synthetic benchmark, fitted over the three shells at . Learned and are averaged over eight validation configurations per seed, then reported as means and sample standard deviations over five seeds of FNO and FNO on a frozen seed-17 backbone at a given . The diagnostics are evaluated at the validation-selected checkpoints.
The spectral exponents largely agree with for all three values, albeit being systematically larger by 0.1–0.2. Exact equality is admittedly not warranted by the present architecture. Indeed, Figure 2 shows that the local MACE backbone already describes a share of the reference Yukawa interaction even at the lowest shells, thus
modifying the learned target. More fundamentally, the perturbation in the spectral diagnostic applies directly to the deposited latent fields. It does not propagate through the source head, which also contributes to the full residual energy response by changing the latent fields as the atomic configuration varies. To put it into perspective, we extract from the -dependence of the residual coefficients , shown in Figure 2. We perform a log–log fit of over the three lowest shells, including the Gaussian factor for consistency with . This leads to and 2.28 at and 0.2 Å, respectively. Both values exceed as the backbone absorbs an increasing share of the reference from the first to the third shell. The learned spectral lies between and . The log–log fit at Å is precluded by the negative third-shell residual coefficient. In contrast to the spectral exponent , which characterizes the FNO energy response to latent-field perturbations, describes the wavevector dependence of the residual expressed through physical-charge shell features. The two analyses provide complementary descriptions, and accurate residual recovery does not require the two to coincide. For the present benchmark, the spectral diagnostic provides a numerically more robust characterization across the three screening strengths.
Despite a positive offset, the spectral exponent decreases systematically as screening strengthens in both linear and nonlinear FNOs, closely following the trend in as shown in Table III. The smaller is consistent with the flattening of the analytical Yukawa kernel over the sampled long-wavelength window at larger . As both FNO variants yield comparable exponents, we conclude that a linear operator alone is able to track the screening dependency in this homogeneous-screening benchmark.
Energy–force accuracy
Figure 3 shows the energy and force RMSEs obtained with the one-interaction MACE backbone, the frozen MACE–FNO, and the joint MACE–FNO, at the three inverse screening lengths. The figure and Table IV also include the corresponding linear-FNO models with the same seed set, training budgets, and checkpoint-selection rule. We further include MACE–LES and a two-interaction MACE backbone for comparison. The local MACE backbone uses a nonresidual first interaction block (RealAgnoisticInteractionBlock), which we also adopt for MACE–LES. The local MACE backbones and MACE–LES are trained in two stages, with an increased relative weight of the energy error during the second stage. MACE–LES models use the default values for the Gaussian smoothing length ( Å) and the reciprocal resolution parameter ( Å). In accord with the FNO models, we train the LES models independently with five random seeds, which sample variability in both the local backbone and the latent-charge prediction.
Figure 3. Accuracy of energy and forces for the synthetic benchmark. Energy and force RMSEs are shown in panels (a) and (b), respectively, at three inverse screening lengths. Both RMSE axes are logarithmic. Darker backgrounds indicate stronger screening as increases. Error bars denote sample standard deviations over five seeds, whereas MACE and MACE are single-seed baselines. Lines are guides to the eye.
| Model | Å Energy (meV atom) | Å Force (meV Å) | Å Energy (meV atom) | Å Force (meV Å) | Å Energy (meV atom) | Å Force (meV Å) |
| MACE | 1.30 | 71.37 | 0.93 | 56.03 | 0.45 | 31.95 |
| MACE | 0.69 | 29.16 | 0.42 | 15.95 | 0.16 | 10.32 |
| MACE–LES | 0.06 ± 0.00 | 7.96 ± 0.11 | 0.16 ± 0.00 | 12.59 ± 0.10 | 0.17 ± 0.00 | 13.38 ± 0.09 |
| MACE–FNO | 0.11 ± 0.00 | 16.81 ± 0.40 | 0.13 ± 0.00 | 16.51 ± 0.26 | 0.11 ± 0.00 | 13.76 ± 0.20 |
| MACE–FNO | 0.11 ± 0.01 | 16.16 ± 0.15 | 0.13 ± 0.02 | 15.77 ± 0.23 | 0.10 ± 0.00 | 13.36 ± 0.16 |
| MACE–FNO | 0.04 ± 0.00 | 6.45 ± 0.06 | 0.05 ± 0.01 | 6.74 ± 0.03 | 0.06 ± 0.00 | 8.28 ± 0.04 |
| MACE–FNO | 0.05 ± 0.01 | 6.59 ± 0.01 | 0.06 ± 0.01 | 6.97 ± 0.03 | 0.07 ± 0.01 | 8.56 ± 0.04 |
Table 4. Energy and force RMSEs on the synthetic benchmark for the three inverse screening lengths . Spreads are sample standard deviations over five seeds, whereas MACE and MACE are single-seed baselines. Bold means denote the lowest RMSE or a mean ± standard-deviation interval overlapping that of the lowest-RMSE model.
MACE uses the same cutoff and hidden representation as MACE. The numerical RMSE values are reported in Table IV.
With the frozen MACE backbone, nonlinear FNO lowers the energy RMSE by 78–91 % and the force RMSE by 58–77 % relative to the local baseline. The energy and force accuracy is largely maintained with linear FNO. By allowing the backbone to adapt in joint optimization, both FNO variants achieve lower energy and force RMSEs than MACE–LES and MACE across all three screening strengths.
While a second interaction layer in MACE extends the receptive radius from 4.5 to 9 Å, it remains less accurate than jointly trained MACE–FNO across all three screening lengths. Likewise, MACE–LES outperforms MACE at Å and Å, although its force RMSE is higher at Å. A global correction thus provides a clear advantage over a deeper local backbone, particularly in the weakly screened regime with pronounced long-range interactions. Compared to MACE, LES exhibits a contrasting trend with screening strength. It approaches the accuracy of the jointly trained FNO models at Å, and becomes less reliable as the reference interaction departs further from the unscreened Coulomb form. The fact that all the FNO models maintain high accuracy across varying screening strengths clearly supports the flexibility of FNO augmentation in learning the propagation kernel. In line with the residual recovery analysis and spectral diagnostics, we find that a linear FNO suffices for the present system, and introducing nonlinearity into the field operator provides no systematic improvement in energy–force accuracy.
Charge–charge correlations from molecular dynamics
To understand how the learned nonlocal response carries over to the equilibrium structure, we consider the static charge–charge structure factor [39, 40] in an ionic fluid. For nonzero reciprocal wavevectors, we calculate
where is the fixed reference charge of ion and is its position at time . The brackets refer to an average over production MD frames and over wavevectors of equal magnitude. The charge–charge structure factor therefore measures the charge-density fluctuations at a given wavelength. With a pure long-range Coulomb interaction, vanishes as for , whereas it stays finite at long wavelength when the system is subject to a screened interaction [28]. The long-wavelength behavior of therefore sheds light on how well a model reproduces the collective charge correlations associated with long-range interactions.
To this end, we perform MD simulations in the canonical ensemble (NVT) at 2800 K using the same cubic cell with Å. At this temperature, the synthetic ionic system is a diffusive fluid as evidenced by self-diffusion coefficients of about for both ion species using the reference potential. The self-diffusion coefficients are largely invariant to the screening parameter . We use a Langevin thermostat and a time step of 2 fs. For MD simulations with the synthetic ionic fluid, the FNO energy is averaged over eight mesh origins to reduce the spurious net force from the mesh. The models are nonetheless trained on a single mesh. For the MLIPs, is sampled over a trajectory of 1 ns (with an interval of 0.5 ps) after an equilibration of 2 ps. We estimate the standard error from ten consecutive 100 ps segments of the trajectory. The reference MD using the analytical potential is sampled by ten independent 100 ps replicas, with its standard errors taken over these replicas. None of the MACE–FNO trajectories shows an appreciable change in potential energy, which is within about 1 meV atom between the first and second halves of the 1 ns runs. The mean temperature lies within 10 K of the target. To verify energy conservation, we additionally run 100 ps NVE trajectories with a time step of 2 fs from the equilibrated states. The total energy of MACE–FNO drifts by at most 0.04 meV atom over 100 ps, whereas MACE shows no measurable drift (below 0.001 meV atom). This drift is comparable to the total-energy fluctuation of both models (about 0.04 meV atom) and below 1% of the kinetic-energy fluctuation. Distinct from the static benchmark checkpoints, all MLIPs for MD simulations are trained with a single seed on a separate dataset. Specifically, we generate 40 Langevin trajectories at 2800 K, and sample every 1 ps after a 10 ps burn-in. These configurations are then split into 660 training, 240 validation, and 300 test configurations.
Figure 4 shows on the 15 lowest shells obtained with the short-range and long-range MLIPs in comparison to the analytical reference potential. The reference MD shows the expected dependence on screening: at the lowest shell for and it increases about fivefold to 0.028 for as the shorter-ranged Yukawa interaction penalizes long-wavelength charge fluctuations less. Lacking the long-range interaction, the local MACE strongly overestimates by factors of 9.5 and 2.1 at the lowest shell for and , respectively, compared to the reference. The excess decays only gradually with . At , or a wavelength of about 8 Å, is still 11–13% above the reference. The deeper local MACE extends, but does not complete, the long-range response: at the lowest shell is still overestimated by a factor of 2.9 for and 1.2 for .
Thanks to the prescribed Coulomb kernel, the long-range MACE–LES reproduces the reference within 4% on all shells for . For , however, the long-wavelength charge fluctuations are suppressed too strongly, leading to an underestimated at the lowest shell reaching only 57% of the reference value. By contrast, linear and nonlinear MACE–FNO follow the reference within 2% and 3%, respectively, on the three lowest shells at both screening strengths. Since is not a training target, such an agreement indicates that the learned FNO infers the effective interaction successfully from energy and force labels alone. In particular, whereas LES adapts only its charges and local energy to a fixed Coulomb kernel, the FNO also learns the propagation itself.
Inhomogeneous screening and nonlinear response
With a spatially uniform across atomic configurations, the Yukawa interaction can be represented by a configuration-independent linear propagation kernel. We show this by writing the screened Poisson equation
Figure 4. Charge–charge structure factors of the synthetic ionic fluid at 2 800 K in a cubic cell ( Å). (a) Å; (b) Å. Insets show a zoomed-in view of the lowest shells. The exact reference denotes MD with the analytical interaction potential. Error bars indicate one standard error of the mean for MLIPs trained with a single seed (ten 100 ps blocks of a 1 ns trajectory) and the reference (ten 100 ps replicas), and are mostly smaller than the symbols.
where is the Gaussian-smeared physical charge density, and is omitted for simplicity. The reference potential satisfying Eq. (21) can be expressed in terms of the propagation kernel
where denotes the periodic simulation cell. As depends only on the separation, it is diagonal in Fourier space
At fixed screening, the potential thus depends linearly on , which further leads to the quadratic energy dependence . This explains the success of linear FNO for the homogeneous benchmark in Sec. III, and we expect an equally good description of the homogeneous reference by other sufficiently expressive, learned but configuration-independent kernels, such as those of SOG-Net [9] and RSNN [8].
In many systems, screening is neither uniform nor independent of the atomic configuration. In electrolytes near charged surfaces or dielectric interfaces [41, 42], spatial variations in ion concentration can produce a screening response that depends on the screening environment. Importantly, rearranging the screening medium can change the effective interactions between two charges, even when their local environments remain intact. Capturing this effect requires a configuration-dependent propagation that takes into account the screening medium beyond the local environment of both charges.
Here, we extend the synthetic system of Sec. III by incorporating a charge-neutral screening species into the NaCl-like ionic compound. Inspired by Debye–Hückel theory [43], in which the squared inverse screening length is proportional to ionic strength, we prescribe a screening law in which the local concentration of determines the inverse screening length,
where is the smoothed concentration relative to its cell average, is a normalized Gaussian of width Å summed over periodic images, and the number density ensures over the cell. We choose Å to keep the -depleted regions Coulombic, and Å so that 95% of the volume of the training configurations has Å. The local screening thus spans the range covered separately by the three homogeneous benchmarks of Sec. III. All the other parameters, such as the unit cell, the particle density, the ionic charges and their Gaussian smearing, and the species-independent LJ interaction, follow exactly the ones used in Sec. III.
With , the screened electrostatic energy reads
The reference energy is then . The constant refers to the self-energy of the smeared ionic charges in an -free medium (). Subtracting this constant ensures that isolated atoms keep zero energy, consistent with the isolated-atom reference energies used in training (as for in Eq. (16)). The medium-dependent part of the self-energy stays in , lowering the energy of an ion as the screening medium accumulates around it. As in Sec. III, the factor fixes the standard deviation of the nonlocal energy per atom across the training configurations at meV atom.
For a fixed arrangement of , remains linear in , and is quadratic in the charges, as for the homogeneous Yukawa interaction. In reciprocal space, however, the screened Poisson equation becomes
in which the Fourier components of the concentration field couple different wavevectors. Unlike above, no diagonal kernel can thus describe the response, and the operator changes whenever the screening medium rearranges. The inhomogeneity hence enters through the propagation of the response, not through a nonlinear dependence on the charges. Note that this construction is phenomenological (the particles prescribe the screening without carrying explicit charge), and it is not meant as a thermodynamically consistent coarse-graining of a specific electrolyte.
We solve Eq. (26) numerically in the plane-wave basis of Eq. (16) ( Å), with an iterative solver converged to near machine precision, and obtain the forces on both the ions and the particles from the converged potential. For a uniform medium, the solution reproduces the Yukawa energies and forces of Sec. III at , and finite differences confirm the forces in strongly inhomogeneous media (cf. Appendix B 4).
To generate configurations with inhomogeneous screening, we start with 80 frames obtained from the 40 MD trajectories of the pure ionic systems equilibrated at 2 800 K and Å as described in Sec. III D. Within every frame, we randomly relabel 286 of the 1 146 ions as the charge-neutral species . We then repeatedly propose transferring an label to a non- site, both chosen at random. Acceptance of the Monte Carlo (MC) move is subject to the Metropolis criterion based on the energy change in an external potential acting exclusively on . Here we use a randomized long-wavelength Fourier potential [44]
where denotes the Cartesian coordinate along axis , and K. The Gaussian random weights are normalized so that the sinusoidal sum has unit spatial variance, and the phases are uniform on . The amplitude , drawn uniformly between zero and four for every trajectory, controls the segregation of and thus the screening contrast. Upon acceptance, the original site recovers its template ionic species to maintain the initial charge ordering as much as possible. As this procedure generally leaves an excess of one ion species over the other, we restore charge neutrality by randomly relabeling ions of the more abundant species after all 60 MC sweeps are performed.
For each of the 80 relabeled configurations, we run Langevin MD at 2 800 K and sample 15 frames every 0.5 ps after a 6 ps burn-in. During MD, all particles interact through the LJ potential. The ions additionally experience the homogeneous Yukawa interaction of Eq. (16) at Å. The external potential acts only on to maintain its spatial pattern. For labeling, we recompute the reference energies and forces using the LJ interaction and the inhomogeneous screening model. Splitting by whole trajectories as in Sec. III, we obtain 660, 240, and 300 configurations for training, validation, and test, respectively. Over all configurations, the 5th and 95th percentiles of the spatial distribution of are 0.066 Å and 0.498 Å, respectively. We quantify the inhomogeneity of a configuration by the spatial standard deviation of over the cell, which, averaged over the frames of a trajectory, ranges from 0.039 to 0.167 Å across the 80 trajectories. This dataset thus spans weakly to strongly inhomogeneous screening environments.
Energy–force accuracy and screening inhomogeneity
Unlike for homogeneous screening (Sec. III), an accurate description of inhomogeneous screening is only attainable with the nonlinear FNO, as shown by the energy and force RMSEs in Table V. Joint MACE–FNO achieves the lowest energy and force RMSEs, 0.08 meV atom and 9.25 meV Å, which are 49 % and 27 % below those of joint MACE–FNO, respectively. The nonlinear FNO is also more accurate with a frozen backbone. As both FNO variants share the same source-head architecture, the consistent gain in accuracy can be attributed to the nonlinear field propagation for inhomogeneous screening.
| Model | Energy (meV atom) | Force (meV Å) |
| MACE | 1.49 | 51.14 |
| MACE | 0.50 | 19.58 |
| CACE | ||
| MACE–LES | ||
| CACE–SOG | ||
| MACE–FNO | ||
| MACE–FNO | ||
| MACE–FNO | ||
| MACE–FNO |
Table V. Energy and force RMSEs on the 300 test configurations of the inhomogeneous-screening benchmark. Spreads are sample standard deviations over five seeds, whereas MACE and MACE are single-seed baselines. The local CACE backbone uses a cutoff of 4.5 Å without message passing. Bold values mark the lowest RMSE.
As introduced in § 1, SOG-Net [9] propagates latent sources linearly through a learned sum of Gaussians, as the linear FNO does through its learned spectral kernel. Here we assess CACE–SOG, a SOG-Net implementation with the Cartesian atomic cluster expansion (CACE) local backbone [45] on the inhomogeneous-screening benchmark. To match the receptive radius of MACE, we use a cutoff of 4.5 Å for the CACE backbone with no message passing. We use four latent source channels, each propagated through a shared kernel parametrized as a sum of twelve trainable Gaussian terms. The energy–force weights are 1:100 for the first 45 epochs, and increase to 100:100 thereafter up to 1 000 epochs. The local CACE baseline follows the same two-stage schedule. Table V shows that the force RMSE obtained with CACE–SOG agrees with that of joint MACE–FNO within the seed spread, while the energy RMSE is larger to some extent. Given that the improvement over the local CACE baseline is comparable to that of MACE–FNO over MACE, we conclude that a configuration-independent propagation saturates at this level irrespective of the kernel form and the local backbone.
To examine how these differences depend on the degree of inhomogeneity, Figure 5(a) groups the per-configuration force RMSEs into four windows of the spatial standard deviation of . The error of joint MACE–FNO depends weakly on the inhomogeneity, and its median rises by only 4 % (from 9.00 to 9.35 meV Å) between the least and the most inhomogeneous windows. Over the same range, the medians rise by 14 % for MACE–FNO and by 11 % for CACE–SOG (mostly beyond the second window), and by 39 % for the local MACE, whereas those of MACE–LES stay between 15.4 and 16.0 meV Å without a systematic trend. Hence, the gap between linear and nonlinear FNO widens from 2.9 to 4.3 meV Å, and the two learned fixed kernels are more accurate than LES. The rising MACE error under higher screening inhomogeneity in turn suggests that, as the screening medium segregates, a larger fraction of the ions resides in weakly screened regions, in which the forces depend on changes and on the medium beyond the receptive field of the local model.
Figure 5. Distributions of (a) per-configuration force RMSEs and (b) signed energy errors per atom (model minus reference) in four windows of the spatial standard deviation of . The solid horizontal line in (b) marks zero error. Darker background shading indicates greater screening inhomogeneity. The equally spaced groups are labeled by their observed screening-spread ranges. Grouping is based on the per-configuration screening spread. Violins pool configurations and training seeds, with widths normalized independently. White dots and thick bars indicate medians and interquartile ranges. All methods have five seeds except the single-seed MACE baseline. Only jointly trained FNO models are shown.
The energy RMSE shows no consistent monotonic trend with screening inhomogeneity when evaluated over the same four windows. The signed energy errors in Figure 5(b), however, reveal systematic shifts that the RMSE obscures. While the models with a learned kernel (FNO and CACE–SOG) maintain a mean signed energy error below 0.09 meV atom in magnitude, MACE–LES records a signed error of 1.60 meV atom in the least inhomogeneous window, gradually shifting to meV atom in the most inhomogeneous one. In fact, this offset accounts for most of the LES energy RMSE in Table V, which exceeds even that of the local MACE. The same trend is observed with the two local MACE models. Since the errors are defined as model minus reference, this shift indicates that LES and the local models underestimate the extent to which the reference energy rises as the screening becomes more inhomogeneous. Indeed, segregation at a higher degree of inhomogeneity reduces the energetic stabilization provided by the screening medium. We see this via Eq. (49), which reduces to when the ionic positions are held fixed. A local change of the concentration modifies the energy in proportion to , the squared screened potential of the ions. Therefore, moving away from the ions, where is largest, raises the energy. MACE–LES and the two local MACE models capture this medium-dependent energy increase only partly, missing about 7–19 % of its variation across the test configurations, whereas all learned-kernel models reproduce it to within 1 %.
Thus, the energy and force errors reflect different aspects of the inhomogeneous screening. The energy offset follows the overall screening stabilization of the ions, which the learned kernels appear to capture well, whereas the increasing force errors with inhomogeneity point to a missing dependence of the interionic coupling on the screening medium. We now proceed to probe this dependence directly in Sec. IV B.
Remote screening medium and interionic coupling
In a system with inhomogeneous screening, changes in the screening medium can modify the coupling between two ions even when they occur far from both ions. To probe this dependence directly, we measure how the force on an ion responds to a small displacement of an ion when only the medium far from both ions is rearranged. The interionic response is the interatomic force-constant matrix , the elements of which are
where and are the position and force components of ion , , and is a Cartesian unit vector.
We use Å for the displacement of in the central finite differences.
From one configuration of each of the 20 test trajectories, we select two ion pairs with minimum-image separations between 10 and 12 Å. We freeze all species within 10 Å of either ion of the probed pair. Outside these regions, we resample the species by Metropolis swaps at 2800 K using the potential of the MD for over 1600 sweeps. Since the LJ interaction is invariant under the swaps, the acceptance depends only on the homogeneous Yukawa interaction among the ions and the external potential acting on . For each probe set, we generate five remote environments in this way: four of them are biased by an independently drawn field with (cf. Eq. (27)) towards segregation of into different long-wavelength patterns away from the probed ions, whereas the fifth is unbiased, in favor of a uniform distribution. Together with the base configuration, each probe set thus provides six remote environments that differ only beyond 10 Å from the probed ions. The swaps preserve the composition and charge neutrality. The resampled configurations are representative of the test set, and their reference energies relative to those of the base configurations are within the standard deviation of 8.4 meV atom along the corresponding trajectories. We measure the medium-induced response by , the deviation of from its mean over the six environments within a probe set
where labels the six remote environments in , and refers to either an MLIP model () or the numerical reference (). Note that pooling over a probe set is equivalent to comparing all 15 pair-wise differences of between its six environments. In the reference, the rearrangements of the remote medium change by 0.11 meV Å (RMS of ), about 8% of itself (RMS of 1.42 meV Å). This remote modulation stems entirely from the screened interaction as the LJ interaction between the probed ions is the same in all environments. As a null control, we generate two additional configurations per probe set in which only remote ions are exchanged at a fixed medium distribution. Here the screened Poisson equation depends only on the medium and remains fixed during ionic displacements. The quadratic screened energy thus gives a contribution to that shows no dependence on the remote charges. Indeed, the reference shows no response to this null control, and any change of the model would be a spurious response to the remote charges.
We compare with in Figure 6, pooling all matrix elements of the six environments in the 40 probe sets. For each model, we report the slope of a linear fit as a measure of how strongly a model follows the reference modulation. Only the nonlinear MACE–FNO attains a slope close to unity (cf. Figure 6(g,Figure 6h)). The local models and all the other long-range variants considered here show no appreciable response, albeit for different reasons. With a receptive field cutoff Å in MACE, the matrix elements of are effectively zero since no atomic energy can depend on both ions when they are by construction more than apart. The deeper local MACE couples atoms with Å, and thus yields a nonzero for the probed pairs with an RMS of 1.13 meV Å. The small slope nevertheless shows that MACE is unable to capture the reference modulation. Indeed, a response to the remote medium requires atomic-energy contributions centered on other atoms to depend jointly on the probed pair and the rearranged medium. This is a more restrictive requirement than coupling the two ions alone, and the maximum coupling range of 18 Å does not ensure that shared local environments capture the medium dependence of this coupling.
Figure 6. Response of the interionic coupling to the remote screening medium. Elements of the model response are plotted against those of the reference on equal axes shared by all panels. Each panel pools the six environments of all 40 probe sets. All five model seeds are shown except for the single-seed MACE and MACE baselines. Solid lines show linear fits, and dashed lines indicate . The reported uncertainties in denote one sample standard deviation across seeds.
MACE–LES, CACE–SOG, and the linear MACE–FNO couple the probed ions through a long-range kernel, as manifested by the RMS of 1.4–1.6 meV Å, comparable to that of the reference. Although atoms contribute to the latent source field, the independence of
these kernels from the screening medium prevents remote rearrangements from modifying the propagation operator coupling the probed ions. Hence, these fixed-kernel models capture an effective interionic coupling but miss its modulation by the remote screening medium, leaving constant across the six environments.
In the linear FNO, remote sources only add to the field. In the nonlinear FNO, they also modify the effective propagation between the probed ions, so responds to the remote-medium rearrangements. With joint training, the nonlinear FNO attains a seed-averaged slope of , and the slope within each probe set is positive in all 40 sets for every seed. A comparable slope is obtained with the frozen MACE backbone (). We further quantify the element-wise agreement by the coefficient of determination of the predictions, , which vanishes for a model without response and reaches unity only when every matrix element is reproduced. The joint nonlinear FNO achieves , higher than that of the frozen one ().
The scatter in MACE–FNO is in part a spurious response to remote ionic charges at a fixed distribution of the screening medium. In the null control, the nonlinear FNO changes by an RMS of 0.03–0.05 meV Å, about half that of the scatter. This spurious sensitivity presumably arises because the sources of remote ions can modify the effective propagation in the nonlinear FNO as those of do, and the architecture does not enforce the invariance of the reference to remote ionic rearrangement. For itself, the joint nonlinear FNO gives the lowest RMSE (0.29 meV Å, compared with 0.41 for the joint linear FNO, 0.43 for CACE–SOG, and 0.64 meV Å for LES). Note that none of the models is trained on force constants, and the probe provides an independent test of the learned coupling. It supports the interpretation of the force errors in Sec. IV A: among the models studied here, only the nonlinear FNO captures a substantial part of the reference modulation of the interionic coupling by the remote screening medium.
First-principles benchmarks
Alongside the synthetic benchmark, we consider liquid water, lithium lanthanum zirconate (LLZO), and Au/MgO(001) with remote Al dopants. These benchmarks are collected from existing datasets using first-principles references.
Liquid water represents a prototypical system in which local hydrogen bonding combines with long-range electrostatic coupling between polar water molecules [19]. This coupling influences collective polarization and dielectric screening, which can be misrepresented by short-range MLIPs [4]. We examine whether our MACE–FNO can improve energy and force predictions beyond a compact local MACE description. The liquid-water dataset released with the qNEP study [7] contains reference energies and forces calculated at the meta-GGA SCAN level [46]. The structures contain 128 water molecules (384 atoms) in an aligned cubic cell. Removing exact duplicate geometries from the released 1 388-entry training pool results in 523 distinct configurations to be used in this study. A cell-length-stratified split with seed 17 assigns 471 to training and 52 to validation. From the released 500-entry validation file, we select the 134 unique geometries absent from the source training pool as our test set. All local MACE, MACE–FNO, and MACE–LES models share this geometry-disjoint split.
LLZO is a garnet-type solid electrolyte with mobile lithium ions in an oxide framework. In LLZO, lithium ions interact electrostatically with other lithium ions and the extended La-Zr-O network, giving rise to long-range interactions that go beyond their immediate coordination shell. For LLZO, we take the 1 978 reference structures originally generated by Yan and Zhu [47] and redistributed by Fan et al. with the qNEP study [7]. The reference calculations employ the PBEsol functional [48]. The structures contain 192 atoms per periodic cell. The unit cell sizes and shapes vary across the dataset. As the release provides no predefined training and test partition, we assign consecutive source blocks of 20 structures using
seed 17, obtaining 1 578 training, 200 validation, and 200 test structures. Since the release does not identify the trajectories from which the structures are drawn, test blocks may share a trajectory with training blocks. The local baselines, frozen and joint MACE–FNO, and MACE–LES models share this data partition.
The Au dimer adsorbed on MgO(001) extends our benchmarks from bulk systems to a two-dimensional surface system. Depending on its axis orientation relative to the MgO surface, the dimer can adopt either a wetting (parallel) or a non-wetting (upright) configuration. Theoretical studies have shown that electron-donor doping can stabilize flat Au clusters on MgO through charge transfer [49, 50]. Interestingly, the geometry preference of the dimer can be modified by subsurface Al dopants located more than 10 Å away [2]. Such a long-range correlation provides a direct test of sensitivity to remote chemical changes. We use 5 000 configurations with DFT energies and forces generated by Ko et al. [2] using the PBE functional [51]. We adopt the split of 4 500 training and 500 test structures released with the LES study of Kim et al. [25]. For FNO and LES training, we reserve 225 training structures for internal validation, leaving 4 275 for fitting.
For liquid water and LLZO, we train the local MACE baselines from scratch with a 4.5 Å cutoff and 128 scalar hidden features. For Au/MgO(001), we adopt the one-interaction local MACE checkpoint released with the LES study [25] and train a two-interaction baseline. Both use a 5.5 Å cutoff and 128 scalar and 128 vector channels per interaction.
In addition to the one-interaction backbone, we augment the two-interaction backbone with FNO and LES to assess whether global corrections provide further accuracy gains beyond those achieved by extending the local receptive field. The two-interaction FNO extension follows the same fitting protocol as the one-interaction runs, including the loss scales, learning-rate schedule, and checkpoint-selection rule. The two-interaction source head receives 256 invariant descriptors from the two interactions instead of 128. The source-head hidden width and all FNO settings are unchanged.
We adopt the FNO settings in Table II. For all three systems, we use a hidden-channel width for the Fourier layers and mix all hidden channels through a single spectral channel group (). For Au/MgO(001), we report the results obtained with full -mixing over 24 height points within the spectral matrix Eq. (5). The channel-wise mixing results are presented in Appendix B.3.
To reduce spurious energy and force changes under rigid translation relative to the mesh, we average the FNO energy corrections over four in-plane mesh origins for Au/MgO(001). These correspond to all combinations of zero and half-grid shifts along the two periodic directions. This interlacing is applied during both training and inference. Liquid water and LLZO are evaluated on a single mesh, since interlacing changes their energy and force RMSEs only marginally (Appendix B.1).
Energy and force accuracy
As summarized in Table VI, FNO augmentation substantially improves the accuracy of the one-interaction backbone for liquid water, LLZO, and Au/MgO(001). With the frozen MACE backbone, nonlinear FNO already reduces the energy RMSE by 40–51 % and the force RMSE by 23–43 % for the two bulk systems. The gains are particularly sizable for Au/MgO(001), reaching 97 % for energy and 90 % for forces. Training the backbone and FNO jointly further lowers the force RMSE by 9–37 % in all three systems and the energy RMSE by 10 % and 37 % for water and LLZO, respectively. Most of the improvement over the local baseline is, however, already obtained with linear FNO. With joint training, MACE–FNO accounts for 86–100 % of the RMSE reduction achieved by MACE–FNO. Nonlinearity mainly benefits the forces of LLZO and Au/MgO(001), whose RMSEs decrease by a further 18 % and 23 %, respectively, while both RMSEs of water change by only 3 %.
| Model | Liquid water Energy (meV atom) | Liquid water Forces (meV Å) | LLZO Energy (meV atom) | LLZO Forces (meV Å) | Au/MgO(001) Energy (meV atom) | Au/MgO(001) Forces (meV Å) |
| MACE | 1.58 | 106.35 | 0.58 | 70.50 | 2.31 | 56.99 |
| MACE–LES | 0.90 ± 0.01 | 71.01 ± 0.12 | 0.24 ± 0.02 | 25.29 ± 0.19 | 0.11 ± 0.01 | 6.42 ± 0.04 |
| MACE–FNO | 0.79 ± 0.02 | 84.45 ± 0.23 | 0.39 ± 0.01 | 44.80 ± 0.26 | 0.07 ± 0.00 | 7.61 ± 0.21 |
| MACE–FNO | 0.77 ± 0.04 | 81.79 ± 1.06 | 0.35 ± 0.00 | 40.29 ± 0.29 | 0.07 ± 0.00 | 5.70 ± 0.45 |
| MACE–FNO | 0.72 ± 0.02 | 73.87 ± 0.97 | 0.27 ± 0.01 | 31.21 ± 0.89 | 0.07 ± 0.00 | 6.72 ± 0.24 |
| MACE–FNO | 0.70 ± 0.03 | 71.78 ± 0.48 | 0.22 ± 0.01 | 25.47 ± 0.35 | 0.07 ± 0.00 | 5.17 ± 0.28 |
| MACE | 0.40 | 35.75 | 0.10 | 12.18 | 0.15 | 3.35 |
| MACE–LES | 0.34 ± 0.02 | 32.38 ± 0.05 | 0.09 ± 0.01 | 10.32 ± 0.13 | 0.11 ± 0.03 | 3.61 ± 0.22 |
| MACE–FNO | 0.33 ± 0.01 | 35.37 ± 0.02 | 0.09 ± 0.00 | 11.37 ± 0.02 | 0.07 ± 0.00 | 2.99 ± 0.09 |
| MACE–FNO | 0.31 ± 0.01 | 35.19 ± 0.05 | 0.09 ± 0.00 | 11.41 ± 0.02 | 0.07 ± 0.00 | 2.89 ± 0.09 |
| MACE–FNO | 0.34 ± 0.02 | 35.42 ± 0.04 | 0.09 ± 0.00 | 10.95 ± 0.04 | 0.07 ± 0.00 | 2.93 ± 0.17 |
| MACE–FNO | 0.30 ± 0.01 | 35.12 ± 0.06 | 0.09 ± 0.00 | 10.99 ± 0.02 | 0.07 ± 0.00 | 2.94 ± 0.06 |
Table VI. RMSEs of the energy and forces for liquid water, LLZO, and Au/MgO(001). FNO and LES entries are five-seed means with sample standard deviations, whereas MACE and MACE are single-seed baselines. For Au/MgO(001), all FNO entries use dense full -mixing. Bold lettering marks the lowest RMSE values within their error margins, separately for each backbone depth.
Increasing the depth of the local backbone largely reduces the errors of MACE, lowering the energy RMSE by 75–93 % and the force RMSE by 66–94 %. MACE is also more accurate than all MACE–FNO models except for the energy of Au/MgO(001), for which the FNO models give lower RMSEs. Adding FNO to this deeper MACE backbone further improves the accuracy, although the gains become smaller, particularly for the forces. The jointly trained MACE–FNO lowers the energy RMSE by 24 %, 6 %, and 56 % and the force RMSE by 2%, 10%, and 12% for water, LLZO, and Au/MgO(001), respectively. It is clear that joint training becomes less important as the receptive field of the backbone extends. Whereas joint training reduces the energy and force RMSEs of MACE–FNO by up to 37 %, it changes the RMSEs of MACE–FNO by less than 4 %. Nonlinearity likewise loses its advantage with the deeper backbone, as the force RMSEs of the linear and nonlinear FNO differ by less than 4 % for all three systems. The gains from nonlinearity with MACE in LLZO and Au/MgO(001) thus point to the limited one-interaction backbone rather than to configuration-dependent screening. The pretrained MACE representation, in turn, already supplies most of what the FNO correction requires.
We now compare FNO with LES on the same local backbones. The accuracies of the two approaches are comparable for the two bulk systems. With MACE, FNO gives a lower energy RMSE for water and agrees with LES within the seed spread for LLZO, while LES gives a slightly lower force RMSE for water. With MACE, FNO retains the lower energy RMSE for water, and the force RMSEs are lower with LES for both systems. The learned propagation thus reaches an accuracy comparable to that of the prescribed Coulomb kernel in LES for the two bulk systems, where learned charges can partly represent the polarization.
The advantage of FNO is more prominent for Au/MgO(001), where the energy RMSEs are 36–41 % lower than those of LES in all settings on both backbones, and the force RMSEs are 17–20 % lower with the MACE backbone. With the shallow MACE backbone, only the nonlinear FNO with dense full -mixing lowers the force RMSE relative to LES (by 19 % with joint training), whereas the linear FNO and the nonlinear FNO with channel-wise -mixing give higher force errors (cf. Table VI and XIII). Note that dense full -mixing involves about 20 million FNO parameters, compared with 0.08 million FNO parameters in the channel-wise -mixing and 0.46 million in the MACE backbone (cf. Sec. VI and Appendix B 3). The higher accuracy attained with FNO may reflect the charge transfer at interfaces as discussed in § 1, for which the effective interaction between learned sources need not take the Coulomb form.
Remote-dopant wetting preference
Remote Al dopants in the MgO(001) substrate present a stringent test for nonlocal MLIPs. We define the wetting preference for the adsorbed Au dimer as the relative energy of the wetting (parallel) and non-wetting (upright) configurations
On undoped MgO, a non-wetting dimer configuration is favored according to the PBE calculations [2]. Al dopants in the substrate lead to a sign reversal in , thereby indicating a stabilized wetting configuration. This reversal cannot be described by local MACE models with a receptive field too short to couple the dimer to the dopants [5, 25]. With a cutoff of 5.5 Å, the one-interaction MACE couples only atoms within 11 Å of each other, whereas the nearest dimer–dopant separation is 11.1–11.4 Å in the non-wetting and wetting geometries. The second interaction layer of MACE extends this coupling range to 22 Å. The reach of the local backbones is directly manifested in the wetting preferences shown in Table VII. While MACE predicts essentially the same on both substrates, MACE reproduces the sign reversal upon Al doping.
| Model | Undoped | Al-doped | ||||||
| DFT-PBE | 0.0 | 934.8 | 934.8 | 0.0 | −66.9 | −66.9 | ||
| MACE | 218.9 | 637.9 | 419.0 | −515.8 | −248.2 | 170.8 | 419.0 | +486.0 |
| MACE–LES | −51.7 ± 2.1 | 890.0 ± 2.1 | 941.7 ± 4.2 | +6.9 | −25.9 ± 1.7 | −103.1 ± 2.6 | −77.2 ± 4.1 | −10.3 |
| MACE–LES | 919.3 | −15.5 | −56.5 | +10.4 | ||||
| MACE–FNO | −44.2 ± 1.2 | 881.6 ± 0.4 | 925.8 ± 1.2 | −9.0 | −28.3 ± 0.3 | −95.3 ± 1.2 | −67.0 ± 1.0 | 0.0 |
| MACE–FNO | −46.5 ± 0.3 | 883.8 ± 0.6 | 930.2 ± 0.5 | −4.6 | −27.0 ± 0.2 | −96.6 ± 0.7 | −69.6 ± 0.8 | −2.6 |
| MACE–FNO | −45.3 ± 0.6 | 883.2 ± 0.7 | 928.5 ± 0.5 | −6.4 | −28.0 ± 0.5 | −95.9 ± 0.6 | −67.8 ± 0.9 | −0.9 |
| MACE–FNO | −46.7 ± 0.3 | 883.9 ± 0.4 | 930.6 ± 0.5 | −4.2 | −27.3 ± 0.2 | −97.0 ± 0.9 | −69.6 ± 0.9 | −2.7 |
| MACE | −33.0 | 869.7 | 902.7 | −32.1 | −43.4 | −80.5 | −37.0 | +29.9 |
| MACE–LES | −38.6 ± 5.4 | 876.3 ± 5.9 | 914.9 ± 11.2 | −19.9 | −38.4 ± 5.5 | −88.7 ± 5.3 | −50.4 ± 10.7 | +16.6 |
| MACE–FNO | −47.1 ± 1.2 | 883.8 ± 0.3 | 930.9 ± 1.3 | −3.9 | −29.0 ± 0.9 | −96.2 ± 1.2 | −67.2 ± 1.5 | −0.3 |
| MACE–FNO | −47.0 ± 0.6 | 884.2 ± 0.6 | 931.2 ± 0.4 | −3.6 | −27.6 ± 0.6 | −96.9 ± 1.2 | −69.3 ± 1.4 | −2.4 |
| MACE–FNO | −47.1 ± 1.1 | 884.1 ± 0.4 | 931.1 ± 0.9 | −3.7 | −28.4 ± 0.5 | −97.7 ± 3.1 | −69.3 ± 3.0 | −2.4 |
| MACE–FNO | −46.0 ± 0.9 | 885.0 ± 0.7 | 931.0 ± 0.2 | −3.9 | −27.9 ± 1.1 | −96.1 ± 0.9 | −68.3 ± 0.5 | −1.3 |
Table VII. Relative total energies and wetting preferences for Au/MgO(001). Energies are in meV per cell. The non-wetting () and wetting () energies are evaluated at relaxed geometries obtained with the corresponding model. The MgO substrate is kept fixed as in the DFT reference throughout. The force tolerance is 0.01 eV Å. The DFT non-wetting energy is taken as the common zero for all methods on the same substrate. and . Spreads are sample standard deviations over five seeds, whereas MACE and MACE are single-seed baselines. All FNO entries use dense full -mixing.
The sign reversal is also reproduced when augmenting MACE with FNO or LES. The nonlinear MACE–FNO predicts wetting preferences within 5 meV per cell of the DFT values on both substrates, with seed spreads below 1 meV. For comparison, MACE–LES and MACE deviate by up to 10 meV and 32 meV, respectively. Linear FNO is at least as accurate on doped MgO but deviates by up to 9 meV on undoped MgO. Finally, with the deeper MACE backbone, FNO augmentation reduces the deviations further to below 4 meV, while MACE–LES still deviates by up to 20 meV.
Table VII shows that all MLIPs except MACE place both endpoints below the reference DFT-PBE energies. For the FNO models, these offsets amount to 44–53 meV on undoped and 27–31 meV on Al-doped MgO, equivalent to 0.25–0.48 meV atom, about three to seven times the test energy RMSE. The offsets are already present at the DFT geometries, as relaxation changes the endpoint energies of these models by less than 0.2 meV. We note that the dataset consists of 1 250 perturbed configurations around each of the four endpoints but not the relaxed endpoints themselves. The offsets therefore indicate a common error in extrapolating to the basin minima. With the FNO models, the offsets are found to be more consistent between the wetting and non-wetting configurations, thereby largely canceling in .
Inference time and memory consumption
To put the computational cost of MACE–FNO into context, we compare its linear and nonlinear variants with MACE–LES and the two local baselines on a single NVIDIA A100-SXM4-40GB GPU. Table VIII compares double-precision energy–force inference for the three first-principles benchmarks, with one structure evaluated at a time. Runtime and memory usage are measured with PyTorch 2.7.1 and MACE 0.3.16, without compilation or cuEquivariance acceleration.
| Model | Runtime (ms) | Peak tensor memory (MiB) |
| Liquid water (384 atoms) | ||
| MACE | 27 | 1 399 |
| MACE | 38 | 1 870 |
| MACE–LES | 31 | 1 399 |
| MACE–FNO | 36 | 1 399 |
| MACE–FNO | 40 | 1 400 |
| LLZO (192 atoms) | ||
| MACE | 14 | 298 |
| MACE | 19 | 457 |
| MACE–LES | 17 | 298 |
| MACE–FNO | 22 | 299 |
| MACE–FNO | 31 | 305 |
| Au/MgO(001) (110 atoms) | ||
| MACE | 15 | 461 |
| MACE | 29 | 1 189 |
| MACE–LES | 18 | 461 |
| MACE–FNO (full ) | 57 | 4 473 |
| MACE–FNO (full ) | 175 | 7 683 |
| MACE–FNO (channel-wise) | 32 | 463 |
Table VIII. Energy–force inference runtime and peak tensor memory. Only joint FNO is shown, using seed 17 and . For Au/MgO(001), linear FNO uses dense full channel–height mixing, and the nonlinear FNO entries compare channel-wise and dense full -mixing. All slab FNO entries use four mesh origins and the same benchmarking protocol.
After five warm-up calls per structure, we report the median of 90 GPU-synchronized calls across three rounds with randomized model order. The timings include graph construction, energy–force evaluation, and result transfers, excluding model loading, file input, and stress evaluation. Predictions are checked against the ASE calculators. We measure memory in a fresh process per model and report the maximum live-tensor allocation over nine energy–force evaluations after warm-up. This includes model tensors and force intermediates, but excludes unused allocator blocks and external CUDA allocations. All memory values are expressed in MiB ( bytes). The FNO models use the production hyperparameters listed in Table II, which prioritize predictive accuracy with relatively generous mesh resolutions and network capacities and have not been optimized for inference speed. Frozen and joint nonlinear FNO models differ by less than 1% in runtime and have identical tensor-memory peaks.
Overall, LES remains faster than the FNO methods for all three systems. For the two bulk systems, the cost of FNO augmentation is comparable to that of LES, particularly with linear FNO. The tensor-memory peaks of both FNO variants lie within 3% of those of MACE and MACE–LES, and are 25% and 33% lower than those of MACE for liquid water and LLZO, respectively. The linear FNO takes 17% and 31% longer than LES for the two systems, and the nonlinear FNO 31% and 82% longer. Relative to MACE, the nonlinear FNO takes 6% and 63% longer. FNO augmentation of a shallow MACE backbone thus offers a memory advantage over a deeper local backbone at the expense of increased evaluation time. Since the mesh grows with the cell volume at fixed resolution, whether this advantage persists in larger cells remains to be tested.
For Au/MgO(001), the nonlinear FNO with dense full -mixing gives the highest accuracy among the MACE-based models but is also by far the most expensive. Its runtime is 503% longer than that of MACE, and its tensor-memory peak is 6.5 times larger. Switching from four mesh origins to one lowers the runtime by about 66%. The displaced meshes are evaluated as a batch, sharing the local MACE and source-bead calculations. This single-mesh comparison changes only the inference setting and does not imply unchanged prediction accuracy. The effects of interlacing on translation sensitivity and prediction accuracy are quantified in Appendix B1.
We note that the computational cost associated with Au/MgO can be largely mitigated at a modest loss of accuracy. The linear FNO, which maintains full -mixing, lowers the runtime and tensor-memory peak to 57 ms and 4 473 MiB, while its force RMSE increases from 5.17 meV Å to 6.72 meV Å. Ultimately, channel-wise -mixing reduces the runtime to 32 ms, only 11% longer than that of MACE, with a tensor-memory peak essentially equal to that of MACE. Its force RMSE of 7.4 meV Å for seed 17 is higher than the 5.0 meV Å obtained with full -mixing, but it reproduces the wetting preferences equally well, within 5 meV of DFT (Appendix B3). A low-rank factorization of the full channel–height mixing matrices could also reduce the cost of full -mixing, which we leave for future work.
Conclusions and outlook
We have introduced FNO augmentation as a distinct route to nonlocal MLPs by learning the global coupling between latent source fields. Added as a nonlocal correction term to a complete local backbone, the learned operator captures long-range interactions accurately beyond the local receptive field, and, where a reference kernel is known, the learned response can be compared directly with it through spectral analysis in reciprocal space. A central strength of our framework is its flexibility in describing long-range interactions without adhering to a specific kernel. A prescribed Coulomb kernel is a useful prior when the residual interaction is largely electrostatic, but its predictive power becomes limited once the effective interaction between the sources departs from the Coulomb form, e.g., through screening set by the environment. The learned operator overcomes this limitation by adapting its wavevector dependence to the interaction at hand. On a synthetic ionic system with tunable screened Coulomb interactions, the FNO recovers 95–99 % of the variance of the residual left by the local MACE backbone and, with joint training, consistently outperforms a prescribed-kernel long-range model and a deeper local model at all three screening lengths tested. It further reproduces the long-wavelength charge–charge structure factor of the ionic fluid in both the nearly Coulombic and the strongly screened regimes. Note that low energy and force errors do not by themselves guarantee correct long-wavelength charge fluctuations, which underlie dielectric response and ion solvation. Reproducing these correlations in its own sampled ensemble without explicitly including the structure factor in the training indicates that the learned operator reproduces the collective consequences of the residual interaction, not only the training labels.
With inhomogeneous screening, the interaction can no longer be described by a configuration-independent propagation kernel. Among the models considered here, only the nonlinear FNO captures the modulation of the coupling between distant ions by the remote screening medium, which LES, CACE–SOG, and the linear FNO leave unchanged. Its force RMSE is also 27–41 % lower than those of these models.
We expect learned operators of this kind to be particularly useful where the effective interaction between learned sources departs from the Coulomb form or varies with the surrounding configuration. Among the first-principles benchmarks, Au₂/MgO(001) provides an example, in which remote dopants beyond the reach of the local backbone reverse the wetting preference of the adsorbed dimer through charge transfer [49, 50]. The FNO captures this reversal more accurately than both LES and a deeper local backbone. For liquid water and LLZO, FNO augmentation also substantially lowers the errors of the one-interaction backbone, but reaches an accuracy comparable to that of LES, consistent with environment-dependent charges partly representing the screening from electronic polarization [28]. Note that these first-principles benchmarks are restricted to static energies and forces.
Beyond the present benchmarks, learned operators could thus be relevant to electrolytes and solvated interfaces near electrodes, doped and defective oxides and semiconductors, and metal–insulator interfaces where charge transfer and image charges reshape the long-range response. Finally, connecting the learned response to a physically defined polarization could enable predictions of Born effective charges through its derivatives with respect to atomic positions [52]. The present MACE–FNO implementation does not identify physical charge densities in the latent fields. An extension would require a consistent polarization definition and additional physical constraints or response supervision. These observables could both constrain the learned operator and test its description of nonlocal polarization response.
Acknowledgments
Computational resources have been provided by the supercomputing facilities of UCLouvain (CISM/UCL) and the Consortium des Équipements de Calcul Intensif en Fédération Wallonie-Bruxelles (CECI), funded by the Fonds de la Recherche Scientifique de Belgique (F.R.S.-FNRS) under Convention 2.5020.11 and by the Walloon Region. We also acknowledge computational resources made available on Lucia, the Tier-1 supercomputer funded by the Walloon Region under grant agreement no. 1910247. GPT-6 Astra (OpenAI) was used to assist with code prototyping and data curation under the direction of the authors, who reviewed and verified the resulting changes.
Data and code availability
We provide benchmark datasets and trained inter-atomic potentials on Zenodo at https://doi.org/10.5281/zenodo.2349502. For the synthetic benchmark, the record includes the homogeneous Yukawa and heterogeneous-screening datasets, scripts for data generation and reference evaluation, remote-medium probes, and MD trajectories. The same record contains liquid-water, LLZO, and Au₂/MgO datasets along with the splits used in this work. Model checkpoints and training configurations are included for all benchmarks.
The MACE–FNO source code is available at https://github.com/wch3n/mace_fno.
FNO formulation and implementation
Fourier representation of a nonlocal operator
A neural operator maps one field to another, with parameters shared over the spatial domain [30]. Its Fourier construction starts from a linear integral operator,
where is the input field and is the real-space coupling between source and observation points. A general discretization requires an matrix for mesh points. On a periodic domain, a translation-invariant kernel depends only on and defines a convolution.
Fourier waves then diagonalize the operator
where , and is the unnormalized kernel integral. The volume factor is absorbed into in this convention.
An FNO learns the multiplier within a selected spectral band [29]. For multiple channels, becomes a matrix that mixes channel coefficients at the same wavevector. The operation remains nonlocal because each coefficient and Fourier basis function spans the complete domain. For example, the homogeneous three-dimensional Coulomb kernel gives for , with boundary conditions and the zero mode treated separately. Our learned multipliers are not constrained to this form.
On a mesh with dimensions , we define
where labels channels, labels mesh points, and labels integer reciprocal modes. The inverse transform contains , with . Our operator maps source fields to hidden fields, applies Fourier layers, and projects back to response fields
where forms the initial hidden fields from the source channels and, when included, the cell features . These features have the same values at every mesh point and are defined in Appendix A 4 a. The bracket groups the inputs to . A Fourier layer combines channel-wise mixing through with local channel mixing through , followed by the GELU activation. The learned maps and convert the final hidden fields into the response fields.
The selected Fourier modes determine which wavelengths enter the spectral branch, while the local branch acts without this spectral truncation. The source and hidden channels provide multiple learned fields, without assigned chemical or electrostatic identities. Different Fourier modes are coupled through GELU, and the full operator is not generally described by a single linear convolution. For slabs, Eq. (34) is supplemented by explicit height mixing, as described in Appendix A 4.
Invariant sources and particle–mesh assignment
For atom , the descriptor collects the scalar () product outputs of the MACE interaction layers independent of orientation. The source head converts these features into source amplitudes
where labels a structure of atoms, and layer normalization LN rescales the descriptors with learned scale and shift parameters. The matrix and bias , followed by the sigmoid linear unit (SiLU) activation, produce 64 hidden features, which the bias-free matrix maps to the source amplitudes. Subtracting the mean makes in each channel, removing its uniform, zero-wavevector component.
We spread each amplitude smoothly over nearby mesh points using cubic B-spline weights , distributing it over 64 points in bulk and slab models. The finite-height treatment is given in Appendix A 4. Summing the atomic contributions gives
where is the voxel volume. The weights sum to one and are continuous through their second derivatives. This assignment follows particle–mesh electrostatics [31, 32], but the amplitudes and fields remain latent variables.
Energy derivatives, forces, and backpropagation
We sum the products of sources and responses over channels and mesh points to obtain the residual energy
This energy vanishes for zero sources but is not generally quadratic in , as the response is nonlinear. Exact energy scaling under cell replication also depends on the mode band, cell features, and source mean subtraction, not just the volume factor.
Changing the source fields changes both factors in the energy expression. At fixed cell, differentiation therefore gives
where describes how the response changes with the sources, and denotes its adjoint under the spatial inner product. For a linear self-adjoint map, this expression reduces to . We impose neither this property nor Poisson’s equation.
Forces follow by differentiating with respect to atomic positions. Moving an atom changes both where its source is deposited and the source amplitudes obtained from the MACE descriptors
The first term changes the deposition weights and slab height window at fixed amplitudes. The second follows changes in the descriptors, source head outputs, and channel means, with indexing atoms. Both terms are needed with frozen MACE, since its descriptors still depend on atomic positions. In joint training, changing the MACE parameters affects both the local energy and the descriptors supplied to the FNO
Backpropagation combines these contributions, including derivatives through the force calculation, to update jointly the local model and FNO.
Geometry-specific operators
Metric-aware EqGINO for bulk and slabs
An unconstrained FNO learns independent channel-mixing matrices for different wavevectors, so that rotating a field need not rotate the predicted response. Following EqGINO [36], we impose isotropic weight sharing, in which modes of the same wavelength share a matrix, , with the in-plane wavevector magnitude for slabs. Here we adopt this spectral principle in place of the full EqGINO encoder–decoder architecture. The response transforms with the input field under rotations and reflections that preserve the cell, mesh, and selected mode band. For slabs, this statement applies to in-plane transformations. This construction does not guarantee equivariance under arbitrary rotations on a fixed mesh or invariance under arbitrary changes of lattice basis. Isotropy constrains the spectral matrices. The linearized response of the nonlinear operator about a given atomic configuration can nevertheless be anisotropic, as the spatial arrangement of the sources can break rotational symmetry.
Since the same integer Fourier index represents different wavelengths in cells of different sizes or shapes, we evaluate the learned matrices at physical wavevector magnitudes. Their values are learned at the reciprocal-space shells of a fixed cubic or square reference cell and interpolated by a natural cubic spline in , which reduces to independent shell weights at the reference cell. Outside the knot interval, we extrapolate linearly along the endpoint tangents.
All systems in this work use one spectral channel group (). In the nonlinear FNO, the local matrix also mixes all channels. For liquid water, the cell features supply with in Å, and LLZO additionally receives the six independent components of to describe the cell shape. With the lattice vectors stored as rows of , these features are constant across the mesh and invariant under rigid rotation of the whole cell. The fixed-cell slab has no cell features.
Slab geometry and full z-mixing
The full -mixing in Eq. (5) treats the hidden channel and height as a compound index . With and , each Fourier layer learns a matrix at each radial shell. We interpolate its entries in using the same shell-spline construction as for the bulk spectral matrices. The pointwise matrix remains and is shared across all mesh points.
Channel-wise -mixing Eq. (6) instead combines an matrix at each radial shell with one wavevector-independent height matrix per hidden channel. For one spectral group and radial shells, including the zero mode, the numbers of mixing parameters per Fourier layer are
With , , and for , these counts are 5013504 and 17920, respectively, a factor of about 280. Both formulations also have pointwise weights per layer and identical source, lifting, and projection maps. Appendix B3 compares their accuracy and the measured GPU cost of the complete models.
For Au/MgO(001), the 24 height points span a fixed 22 Å window centered on the instantaneous mean atomic height. Deposition weights extending beyond the height window are assigned to the nearest boundary grid point, preserving the total deposited source. The height window is chosen to ensure that all atoms are enclosed.
Training objective and model selection
For a batch of structures containing atoms, we minimize
where is the total energy, the atom count, and the force component on atom . The scales and balance the two errors, with in every fit. For the synthetic ionic system, and , , and for , , and , respectively. Liquid water and LLZO have and eV Å. For Au/MgO, the one-interaction baseline training errors set eV atom and eV Å. The same scales apply to both backbone depths.
We select FNO checkpoints by minimizing the validation score
where the validation RMSEs use energies per atom and Cartesian force components. We validate every 100 steps for the synthetic system and every 500 steps for the materials benchmarks. Joint training selects the backbone and FNO parameters from the same step.
The source head and FNO start with an Adam learning rate of for the synthetic ionic system and for the materials benchmarks. The learning rate is halved after four validation evaluations without improvement, down to . The synthetic FNO runs use eight structures per training batch, and step budgets of 100 000 and 200 000 for frozen and joint training, respectively. For the first-principles benchmarks, both FNO variants use the step budgets in Table XI, except for joint training on the one-interaction water backbone, which uses 200 000 steps. The effective batch sizes are two structures for liquid water and four for LLZO and Au/MgO(001). For nonlinear full -mixing, we accumulate gradients over four single-structure batches to limit peak memory, preserving this effective batch size. The weights are fixed during the MACE warm-up. Shared implementation and training choices are given in Sec. II C of the main text. System-specific settings accompany the benchmarks in Secs. III and § 1.4.
| (a) Liquid water | (b) LLZO | (c) Au/MgO(001) | ||||||||||||||||
| Energy | Forces | Energy | Forces | Energy | Forces | |||||||||||||
| 24 | 8 | 16 | 2 | 0.53 ± 0.03 | 74.55 ± 0.56 | 32 | 8 | 16 | 4 | 0.34 ± 0.00 | 39.13 ± 0.26 | 24 | 24 | 8 | 16 | 4 | 0.05 ± 0.00 | 8.76 ± 0.08 |
| 32 | 8 | 16 | 2 | 0.53 ± 0.01 | 75.13 ± 0.77 | 40 | 8 | 16 | 4 | 0.34 ± 0.01 | 39.45 ± 0.90 | 32 | 24 | 8 | 16 | 4 | 0.06 ± 0.00 | 9.11 ± 0.10 |
| 24 | 10 | 16 | 2 | 0.56 ± 0.02 | 76.24 ± 0.48 | 32 | 10 | 16 | 4 | 0.34 ± 0.01 | 40.09 ± 0.86 | 24 | 24 | 10 | 16 | 4 | 0.06 ± 0.00 | 8.95 ± 0.27 |
| 24 | 8 | 8 | 2 | 0.55 ± 0.02 | 74.36 ± 0.88 | 32 | 8 | 8 | 4 | 0.35 ± 0.02 | 39.93 ± 1.35 | 24 | 24 | 8 | 8 | 4 | 0.06 ± 0.00 | 9.32 ± 0.50 |
| 24 | 8 | 32 | 2 | 0.55 ± 0.06 | 74.11 ± 1.08 | 32 | 8 | 32 | 4 | 0.35 ± 0.02 | 39.26 ± 0.55 | 24 | 24 | 8 | 32 | 4 | 0.06 ± 0.00 | 8.80 ± 0.10 |
| 24 | 8 | 16 | 4 | 0.68 ± 0.05 | 77.06 ± 0.39 | 32 | 8 | 16 | 6 | 0.34 ± 0.01 | 38.75 ± 0.09 | 24 | 24 | 8 | 16 | 6 | 0.05 ± 0.00 | 8.54 ± 0.21 |
| 24 | 16 | 8 | 16 | 4 | 0.06 ± 0.00 | 9.53 ± 0.29 | ||||||||||||
| 24 | 32 | 8 | 16 | 4 | 0.06 ± 0.00 | 8.57 ± 0.15 |
Table XI. Validation RMSEs of MACE–FNO for liquid water, LLZO, and Au/MgO(001). , , , and follow Table X. Bulk meshes have points, and slab meshes have points. The first row in a panel gives the production values of the listed hyperparameters, followed by one-parameter variations. Entries give means and sample standard deviations over FNO seeds 29, 41, and 53. The backbone, data split, and training budget are fixed within each panel. The Au/MgO(001) data use the channel-wise height mixing. Energy and force RMSEs are in meV atom and meV Å, respectively, at the lowest-loss validation checkpoints.
The local MACE baselines use validation-selected stage-two checkpoints throughout. The exception is MACE2 at Å, for which the stage-one checkpoint gives lower validation energy and force RMSEs. For the first-principles benchmarks, the local models trained here use energy–force weights of 1:100 and 1 000:100, and initial learning rates of 0.01 and 0.001, in stages one and two, respectively. We use AMSGrad and an exponential moving average with decay 0.99. Element-specific energy offsets are obtained by least-squares fitting to the training energies and held fixed during model training. The batch sizes are four structures for water and LLZO and ten for Au/MgO(001), with epoch budgets of 400, 1 000, and 1 000, respectively. Stage two begins halfway through each budget. For Au/MgO(001), we train only MACE2 and use the released MACE1 checkpoint [25]. Table VI reports raw test energy RMSEs without an offset correction.
For LES, we select checkpoints by the stage-two validation loss. The synthetic models use a 1 200-epoch budget. For the first-principles benchmarks, the loss weights, learning rates, and optimizer settings match the local MACE fits. Atomic energy offsets are fitted as for the local MACE models. The batch sizes are four structures for water and LLZO and ten for Au/MgO(001). All three systems use a 1 000-epoch budget, with stage two starting halfway through training. These LES models use a Gaussian width of 1 Å, mesh-spacing parameter d1 = 2 Å, and self-interaction removal.
Spectral-response diagnostic
For the diagnostic in Sec. II B, we keep atoms, descriptors, and model parameters fixed and perturb the deposited fields. For a nonzero integer mode n, the normalized cosine probe is
Angle brackets average over mesh points , giving a probe with zero mean and RMS equal to one. For cell vectors stored as rows of , the physical wavevector is . The equivalent cosine modes are counted once.
Writing , we evaluate the Hessian from positive and negative perturbations. Let select channel , with , , and for . Central differences give
where is defined in Sec. II B. We set the perturbation amplitude to a fraction of the typical source-field magnitude, measured by its RMS over all channels and mesh points. For the synthetic benchmark, we use , i.e., 5% of this magnitude.
We diagonalize the Hessians independently at the sampled wavevectors before averaging their positive over shells. Modes with no positive eigenvalue are recorded but excluded from the logarithmic fit, without discarding any model or structure. Equal-magnitude shells have relative tolerance and absolute tolerance Å.
For a given structure, we fit the logarithms of the finite, positive shell averages by unweighted least squares with a free intercept. The negative slope gives , with equal weights across shells. We first average the structure-level exponents separately for the trained models, then report the mean and sample standard deviation across training seeds. Eigenvalues depend on the learned channel basis, so their scaling alone does not identify physical charges.
Numerical validation and computational cost
Mesh interlacing and translation sensitivity
Mesh discretization can make the predicted energy depend on the placement of the grid relative to the atoms.
| Model | Mesh origins | (meV atom) | (meV Å) | Energy RMSE (meV atom) | Force RMSE (meV Å) | Time (ms) | Tensor peak (MiB) |
| Liquid water | |||||||
| MACE–FNO | 1 | 0.058 | 9.573 | 0.76 | 82.12 | 40 | 1 400 |
| MACE–FNO | 8 | 0.001 | 0.463 | 0.77 | 81.85 | 44 | 1 400 |
| MACE–FNO | 1 | 0.059 | 9.496 | 0.66 | 71.87 | 40 | 1 400 |
| MACE–FNO | 8 | 0.001 | 0.453 | 0.67 | 71.58 | 44 | 1 400 |
| LLZO | |||||||
| MACE–FNO | 1 | 0.045 | 3.815 | 0.35 | 40.36 | 32 | 305 |
| MACE–FNO | 8 | 0.001 | 0.388 | 0.35 | 40.29 | 41 | 985 |
| MACE–FNO | 1 | 0.021 | 2.571 | 0.22 | 25.01 | 32 | 305 |
| MACE–FNO | 8 | 0.001 | 0.232 | 0.22 | 24.95 | 41 | 985 |
| Au/MgO(001) | |||||||
| MACE–FNO | 1 | 0.281 | 11.579 | 0.24 | 12.99 | 32 | 462 |
| MACE–FNO | 4 | 0.011 | 0.461 | 0.07 | 8.88 | 33 | 463 |
| MACE–FNO | 1 | 0.311 | 10.496 | 0.27 | 11.52 | 32 | 462 |
| MACE–FNO | 4 | 0.009 | 0.422 | 0.07 | 7.37 | 33 | 463 |
Table IX. Effect of mesh interlacing on MACE–FNO and MACE–FNO with fixed inference weights. and measure translation sensitivity. RMSEs are relative to DFT. Costs refer to energy–force inference.
To reduce this dependence, we average FNO energies over shifted mesh origins for a given set of atomic positions. For mesh points along a periodic cell vector , the mesh spacing is . The allowed offsets are zero or half this spacing, , chosen independently along the periodic directions, giving
where denotes the atomic positions and specifies offsets in units of the mesh spacings. The quantity is the residual energy evaluated on that mesh. We redeposit the sources and evaluate the same FNO separately for the different mesh origins. There are four origins for a slab () and eight for periodic bulk ().
We compare the effect of interlacing using the same validation-selected, seed-17 frozen and joint MACE–FNO checkpoints with the hyperparameters in Table II. The Au/MgO checkpoints use the channel-wise height mixing Eq. (6). The trained weights remain unchanged. Energy and force RMSEs are evaluated using the full test sets covering 134 water, 200 LLZO, and 500 Au/MgO structures.
For each benchmark system, we additionally sample 32 structures without replacement and apply 16 rigid translations per structure in order to study the translation sensitivity. For a given structure associated with atomic positions , we apply the same displacement to every atom for translation , relative to the fixed mesh origins.
The shift is in-plane for a slab. We then evaluate
where and label atoms and Cartesian components, respectively, and is the atom count.
Table IX shows reduced translation sensitivity for every checkpoint, by factors of 24–45 for and 10–25 for . Only the slab shows appreciable DFT–RMSE improvement. Bulk changes stay below 0.005 meV atom and 0.30 meV Å.
The interlacing cost measurements follow the protocol in Sec. VI. We process the shifted meshes as a single FNO batch, while the MACE backbone and source head run only once per structure. Graph construction and result transfers are also shared across mesh origins. Latency increases by about 9% for water, 28% for LLZO, and 2% for Au/MgO. Peak tensor memory changes by less than 1% for water and Au/MgO and increases by 223% for LLZO.
Hyperparameter sensitivity and convergence
We assess the mesh resolution , the mode parameter , the source-channel count , and the number of Fourier layers with a frozen MACE backbone. Tables X and XI summarize the synthetic and physical benchmarks, respectively. denotes the number of mesh points per periodic direction. The slab height-grid size is specified separately. The mode parameter includes integer indices along the Fourier-transformed directions. The complete parameter screens and model-selection records remain available with the source data.
| \multicolumn{2}{c}{} | \multicolumn{2}{c}{} | \multicolumn{2}{c}{} | Seeds | |||||||
| Energy | Forces | Energy | Forces | Energy | Forces | |||||
| 44 | 11 | 4 | 2 | 5/5/5 | ||||||
| 44 | 11 | 8 | 2 | 0.15 | 15.79 | 0.14 | 13.27 | 3/1/1 | ||
| 44 | 11 | 16 | 2 | 0.13 | 15.53 | 0.14 | 13.09 | 3/1/1 | ||
| 44 | 11 | 4 | 4 | 0.14 | 16.10 | 0.14 | 13.54 | 5/1/1 | ||
| 44 | 8 | 4 | 4 | 0.13 | 16.13 | – | – | – | – | 1/–/– |
| 32 | 11 | 4 | 4 | 0.13 | 16.80 | – | – | – | – | 1/–/– |
Table X. Validation RMSEs of MACE–FNO for the synthetic ionic system. is the number of mesh points per direction ( points in total), the mode parameter, the number of source channels, and the number of Fourier layers. The first row gives the production settings of Table II, followed by the tested variants. All fits share the frozen seed-17 backbone at a given and a 100 000-step budget. RMSEs are evaluated at validation-selected checkpoints. Multi-seed entries give means and sample standard deviations. Energy and force RMSEs are in meV atom and meV Å, respectively. Seed counts follow the order of the three columns. Dashes mark untested combinations.
For the synthetic system, we compare the FNO settings of Table II (, two Fourier layers, a mesh, and ) with variants in which one or two of these settings change. All variants are trained with the frozen seed-17 MACE backbone at the corresponding and the 100 000-step protocol of the synthetic benchmark, and tbl:10 reports the validation RMSEs at the selected checkpoints. Changes in the force RMSE are quoted as differences paired by FNO seed, with the standard error of the mean difference.
None of the variants lowers the force RMSE by more than about 3%. Increasing the number of source channels from four to eight changes it by for , whereas sixteen channels lower it by , and by 2.0–2.6% for single seeds at the other two , at about three times the training cost per step. Four Fourier layers increase the force RMSE by , and neither the lower mode parameter nor the coarser mesh, both tested with four layers, improves on the production setting. The energy RMSEs show no systematic change across the variants, and the test RMSEs give the same ranking. The production setting thus represents a compromise between accuracy and cost, with sixteen source channels as the only variant that lowers the force error appreciably.
For liquid water, LLZO, and Au/MgO(001), Table XI compares one-parameter variations around the production settings. We average over three paired FNO seeds at a fixed training budget within each panel, using the same data splits and backbones as the first-principles benchmarks. We select checkpoints with (43) and report raw energy RMSEs without offset corrections. All fits have hidden width and a source head with 64 hidden units. The Au/MgO(001) convergence data are obtained with the channel-wise height mixing in (6), using a fixed 22 Å window. All fits use one spectral channel group (). Loss normalization and interlacing remain fixed within a system.
Finer periodic meshes and a larger mode parameter give no systematic improvement. Changing the source-channel count from 16 to 8 or 32 also provides no consistent benefit. Adding two Fourier layers lowers the mean force RMSE by about 1% for LLZO and 3% for Au/MgO(001). For water, the same increase in depth raises both mean RMSEs at the fixed training budget. For the slab, increasing from 24 to 32 lowers the force RMSE by about 2%, while reducing it to 16 increases the error by about 9%. These matched-budget comparisons support the production settings as a practical accuracy–cost compromise, without establishing a unique converged optimum.
Full and channel-wise -mixing
We compare channel-wise and dense full -mixing for jointly trained MACE–FNO models with seed 17 (Table XII). Both use the same starting backbone, training and validation structures, and four Fourier layers with . The mesh contains points, with , , and four in-plane mesh origins. Both models completed 240 000 training steps.
Reference solver for inhomogeneous screening
We solve (26) by expanding the potential in plane waves up to , including . In a neutral cell, the zero-mode potential vanishes for a uniform medium. An inhomogeneous medium instead couples it to the nonzero modes through
where denotes the cell average. The zero mode must therefore be included even when the total charge is zero.
The medium field is a sum of Gaussians centered on the particles [Eq. (24)], and its Fourier coefficients contain the smoothing factor . We evaluate these coefficients analytically from the particle positions and truncate the medium-field expansion where its Gaussian smoothing factor falls below , which leads to . For Å, Å.
| Mixing | Energy (meV atom) | Forces (meV Å) | Undoped | Al-doped | FNO | Total | Runtime (ms) | Peak tensor memory (MiB) |
| Channel-wise | 0.0725 | 7.37 | 930.47 | −69.23 | 83 008 | 545 984 | 32.4 | 462.9 |
| Full | 0.0691 | 4.98 | 929.89 | −69.32 | 20 065 344 | 20 528 320 | 175.1 | 7 683.2 |
Table XII. Accuracy and GPU cost of channel-wise and full -mixing for joint MACE–FNO on Au/MgO(001). FNO parameter counts include the source head, and total counts also include the backbone.
We evaluate the convolution in Eq. (26) by forming the product in real space, rather than directly in reciprocal space. Both expansions are transformed to a grid of points per axis, multiplied pointwise, and transformed back, keeping only the modes of the basis [53]. The product contains integer mode components up to , where and are the largest absolute integer mode components of the two expansions (17 and 13 for the 32 Å cell and Å). On a finite grid, modes beyond its resolution are aliased onto lower ones, and the condition ensures that none of them falls into the retained band. To satisfy this condition, all labels use . We solve the linear system by conjugate gradients, preconditioned by the diagonal inverse of the uniform-medium operator, . A relative residual of is reached in 22–26 iterations per configuration.
To reduce the sensitivity of the energy to the finite solver tolerance, we use the stationary expression
where . This expression equals Eq. (25) at the exact solution, and its error is second order in the solver residual. At convergence, stationarity removes the terms involving from the energy variation [54].
The forces can therefore be evaluated without differentiating the solver iterations. Displacing ion changes , whereas displacing screening particle changes . Their screened force contributions are
The coefficients of needed for the forces are obtained on the same alias-free grid. We multiply these contributions by and add the Lennard–Jones forces to obtain the reference forces. The constant is evaluated in the -free medium using the nonzero modes of the same plane-wave basis, so that the self-energy subtraction is consistent with the Fourier cutoff. One energy and force evaluation takes about 0.7 s on a single CPU core.
Before labeling, we verified the uniform Yukawa limit, grid independence for , 64, and 96, symmetries under translation and cell-compatible rotations, and the vanishing net force. Analytic forces agree with finite differences within eV Å.
References
References
- [1]I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, in Advances in Neural Information Processing Systems, Vol. 35 (2022) pp. 11423–11436.
- [2]T. W. Ko, J. A. Finkler, S. Goedecker, and J. Behler, Nat. Commun. 12, 398 (2021).
- [3]O. T. Unke, S. Chmiela, M. Gastegger, K. T. Schütt, H. E. Sauceda, and K.-R. Müller, Nat. Commun. 12, 7273 (2021).
- [4]B. Cheng, npj Comput. Mater. 11, 80 (2025).
- [5]D. S. King, D. Kim, P. Zhong, and B. Cheng, Nat. Commun. 16, 8763 (2025).
- [6]Y. Park, D. Kim, D. S. King, N. H. Đào, R. Savoj, S. Hamel, X. Wang, and B. Cheng, Polarizable atomic multipoles for learning long-range electrostatics (2026), arXiv:2605.05746 [cond-mat.mtrl-sci].arxiv.org/abs/2605.05746
- [7]Z. Fan, B. Tang, E. Berger, E. Berger, E. Fransson, K. Xu, Y. Zan, Z. Liu, Z. Song, H. Dong, S. Chen, L. Li, Z. Wang, Y. Zhu, J. Wiktor, and P. Erhart, J. Chem. Theory Comput. 22, 4787 (2026).
- [8]R. Guo, H. Yu, L. Hong, S. Chen, X. Gong, and H. Xiang, Phys. Rev. B 113, 174101 (2026).
- [9]Y. Ji, J. Liang, and Z. Xu, Phys. Rev. Lett. 135, 178001 (2025).
- [10]A. Kosmala, J. Gasteiger, N. Gao, and S. Günnemann, in Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 202 (2023) pp. 17544–17563.
- [11]J. Gasteiger, S. Giri, J. T. Margraf, and S. Günnemann, in Machine Learning for Molecules Workshop, NeurIPS (2020).
- [12]Y. Wang, C. Cheng, S. Li, Y. Ren, B. Shao, G. Liu, P.-A. Heng, and N. Zheng, in Advances in Neural Information Processing Systems, Vol. 37 (2024) pp. 120336–120365.
- [13]O. T. Unke and M. Meuwly, J. Chem. Theory Comput. 15, 3678 (2019).
- [14]A. E. Sifain, N. Lubbers, B. T. Nebgen, J. S. Smith, A. Y. Lokhov, O. Isayev, A. E. Roitberg, K. Barros, and S. Tretiak, J. Phys. Chem. Lett. 9, 4495 (2018).
- [15]S. Gong, Y. Zhang, Z. Mu, Z. Pu, H. Wang, X. Han, Z. Yu, M. Chen, T. Zheng, Z. Wang, L. Chen, Z. Yang, X. Wu, S. Shi, W. Gao, W. Yan, and L. Xiang, Nat. Mach. Intell. 7, 543 (2025).
- [16]Y. Shidu, F. Pellegrini, E. Küçükbenli, R. Lot, and S. de Gironcoli, npj Comput. Mater. 10, 47 (2024).
- [17]I. Batatia, W. J. Baldwin, D. Kurylа, J. Hart, E. Ka-soar, A. M. Elena, H. Moore, M. J. Gawkowski, B. X. Shi, V. Kapil, P. Kourtis, I.-B. Magdău, and G. Csányi, MACE-POLAR-1: A polarisable electrostatic foundation model for molecular chemistry (2026), arXiv:2602.19411 [physics.chem-ph].arxiv.org/abs/2602.19411
- [18]L. Zhang, H. Wang, M. C. Muniz, A. Z. Panagiotopoulos, R. Car, and W. E, J. Chem. Phys. 156, 124107 (2022).
- [19]A. Gao and R. C. Remsing, Nat. Commun. 13, 1572 (2022).
- [20]A. Grisafi and M. Ceriotti, J. Chem. Phys. 151, 204105 (2019).
- [21]K. K. Huguenin-Dumittan, P. Loche, H. Ni, and M. Ceriotti, J. Phys. Chem. Lett. 14, 9612 (2023).
- [22]E. Rumiantsev, M. F. Langer, T.-E. Sodjargal, M. Ceriotti, and P. Loche, Trans. Mach. Learn. Res. (2026).
- [23]C. Faller, M. Kaltak, and G. Kresse, J. Chem. Phys. 161, 214701 (2024).
- [24]A. Caruso, J. Venturin, L. Giambagli, E. Rolando, Z. El-Machachi, F. Noé, and C. Clementi, Nat. Commun. 17, 1855 (2026).
- [25]D. Kim, X. Wang, S. Vargas, P. Zhong, D. S. King, T. Jaffrelot Inizan, and B. Cheng, J. Chem. Theory Comput. 21, 12709 (2025).
- [26]P. Zhong, D. Kim, D. S. King, and B. Cheng, npj Comput. Mater. 11, 384 (2025).
- [27]Z. Fan, Z. Zeng, C. Zhang, Y. Wang, K. Song, H. Dong, Y. Chen, and T. Ala-Nissila, Phys. Rev. B 104, 104309 (2021).
- [28]F. Grasselli, K. Rossi, S. de Gironcoli, and A. Grisafi, Long-range electrostatics in atomistic machine learning: A physical perspective (2026), arXiv:2602.11071v2 [cond-mat.mtrl-sci].arxiv.org/abs/2602.11071
- [29]Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, in International Conference on Learning Representations (2021).
- [30]N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar, J. Mach. Learn. Res. 24, 1 (2023).
- [31]T. Darden, D. York, and L. Pedersen, J. Chem. Phys. 98, 10089 (1993).
- [32]U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, and L. G. Pedersen, J. Chem. Phys. 103, 8577 (1995).
- [33]S. Baroni, S. de Gironcoli, A. Dal Corso, and P. Giannozzi, Rev. Mod. Phys. 73, 515 (2001).
- [34]T. Dornheim, M. Böhme, Z. A. Moldabekov, J. Vorberger, and M. Bonitz, Phys. Rev. Research 3, 033231 (2021).
- [35]A. Paszke, S. Gross, F. Massa, A. Lerер, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, in Advances in Neural Information Processing Systems, Vol. 32 (2019) pp. 8024–8035.
- [36]S. Kim, J. Song, S. Shin, G. Cho, S. Kim, and C. Park, EqGINO: Equivariant geometry-informed Fourier neural operators for 3D PDEs (2026), arXiv:2606.03260 [cs.LG].arxiv.org/abs/2606.03260
- [37]A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dulak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C. Jennings, P. B. Jensen, J. Kermode, J. R. Kitchin, E. L. Kolsbjerg, J. Kubal, K. Kaasbjerg, S. Lysgaard, J. B. Maronsson, T. Maxson, T. Olsen, L. Pastewka, A. Peterson, C. Rostgaard, J. Schiøtz, O. Schütt, M. Strange, K. S. Thygesen, T. Vegge, L. Vilhelmsen, M. Walter, Z. Zeng, and K. W. Jacobsen, J. Phys.: Condens. Matter 29, 273002 (2017).
- [38]D. P. Kingma and J. Ba, in International Conference on Learning Representations (2015).
- [39]A. P. Costepitake and R. Evans, J. Phys. C: Solid State Phys. 15, 4961 (1982).
- [40]J.-P. Hansen and I. R. McDonald, Theory of Simple Liquids: with Applications to Soft Matter, 4th ed. (Academic Press, 2013).
- [41]S. Buyukdagli and R. Blossey, J. Phys.: Condens. Matter 28, 343001 (2016).
- [42]N. R. Aluru, F. Aydin, M. Z. Bazant, D. Blankschtein, A. H. Brozena, J. P. de Souza, M. Elimelech, S. Faucher, J. T. Fourkas, V. B. Koman, M. Kuehne, H. J. Kulik, H.- K. Li, Y. Li, Z. Li, A. Majumdar, J. Martis, R. P. Misra, A. Noy, T. A. Pham, H. Qu, A. Rayabharam, M. A. Reed, C. L. Ritt, E. Schwegler, Z. Siwy, M. S. Strano, Y. Wang, Y.-C. Yao, C. Zhan, and Z. Zhang, Chem. Rev. 123, 2737 (2023).
- [43]P. Debye and E. Hückel, Phys. Z. 24, 185 (1923).
- [44]F. Sammüller, S. Hermann, D. de las Heras, and M. Schmidt, Proc. Natl. Acad. Sci. U.S.A. 120, e2312484120 (2023).
- [45]B. Cheng, npj Comput. Mater. 10, 157 (2024).
- [46]J. Sun, A. Ruzsinszky, and J. P. Perdew, Phys. Rev. Lett. 115, 036402 (2015).
- [47]Z. Yan and Y. Zhu, Chem. Mater. 36, 11551 (2024).
- [48]J. P. Perdew, A. Ruzsinszky, G. I. Csonka, O. A. Vydrov, G. E. Scuseria, L. A. Constantin, X. Zhou, and K. Burke, Phys. Rev. Lett. 100, 136406 (2008).
- [49]N. Mammen, S. Narasimhan, and S. de Gironcoli, J. Am. Chem. Soc. 133, 2801 (2011).
- [50]N. Mammen and S. Narasimhan, J. Chem. Phys. 149, 174701 (2018).
- [51]J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
- [52]X. Gonze and C. Lee, Phys. Rev. B 55, 10355 (1997).
- [53]J. P. Boyd, Chebyshev and Fourier Spectral Methods, 2nd ed. (Dover Publications, Mineola, NY, 2001).
- [54]M. K. Gilson, M. E. Davis, B. A. Luty, and J. A. McCammon, J. Phys. Chem. 97, 3591 (1993).





