Accurate, transferable, and verifiable machine-learned interatomic potentials for layered materials Authors: Johnathan D Georgaras, Akash Ramdas, Chung Hsuan Shan, Elena Halsted, Berwyn Berwyn, Tianshu Li, Felipe H da Jornada Published: 2026 ABSTRACT AbstractTwisted layered van der Waals materials often exhibit unique electronic and optical properties absent in their non-twisted counterparts. Unfortunately, predicting such properties is hindered by the difficulty in determining the atomic structure in materials displaying large moiré domains. Here, we introduce a split machine-learned interatomic potential (MLIP) and dataset curation approach that separates intralayer and interlayer interactions and significantly improves model accuracy, yielding roughly a tenfold improvement in energy and force predictions relative to conventional models. We further demonstrate that traditional MLIP validation metrics – force and energy errors – are inadequate for moiré structures and develop a holistic, physically-motivated metric based on the distribution of stacking configurations. This metric effectively compares the entirety of large-scale moiré domains between two structures instead of relying on conventional measures evaluated on smaller commensurate cells. Finally, we establish that one-dimensional, rather than two-dimensional, moiré structures can serve as efficient surrogate systems for validating MLIPs, permitting validation protocols against explicit DFT calculations. Applying our framework to HfS2/GaS bilayers reveals that accurate structural predictions directly translate into reliable electronic properties. Our model-agnostic approach integrates with various intralayer and interlayer interaction models, enabling computationally tractable relaxation of moiré materials, from bilayer to complex multilayers, with rigorously validated accuracy.Subject terms: Atomistic models, Electronic properties and materials, Computational methods Introduction Stacking individual layers of two-dimensional (2D) materials with small misalignments in rotation, strain, or lattice geometry forms moiré superlattices: structures characterized by periodic unit-cell repetitions that produce long-range modulation patterns spanning tens to hundreds of nanometers. These superlattices host spatially varying stacking-dependent potentials, often leading to flat electronic bands that drive correlated macroscopic phenomena such as unconventional superconductivity1–3, exotic topological phases4–6, and correlated insulating states7–10. Moiré patterns also host unique quasiparticles and collective excitations such as moiré-confined excitons, phonons, and spatially trapped states11–14. Predicting and controlling these emergent phenomena necessitates models capable of capturing the strain-induced lattice distortions and distribution of stacking configurations intrinsic to moiré materials. The conventional approach for modeling moiré structures typically employs density-functional theory (DFT), which can accurately predict atomistic reconstructions and domain formation15,16. However, the high computational cost of these methods severely restricts their application to supercells containing only a few thousand atoms17. Still, these atomistic methods, performed on smaller commensurate lattices, can nonetheless parametrize faster classical and (semi)empirical methods applicable to larger superlattices18,19. A complementary approach, based on continuum elastic theory, integrates out atomic-scale details to operate on larger-scale displacement fields16,20–23. While not strictly atomistic, such elastic-theory methods can successfully predict and interpret several moiré-related phenomena16 and clarify the critical role that local stacking energies and strain distributions play in these systems. A promising generalizable and scalable way to model materials is with machine-learned interatomic potentials (MLIPs). MLIPs combine the computational efficiency of classical force fields with near ab initio accuracy through data-driven training on extensive DFT calculation datasets24. Most MLIPs generally span a spectrum of interpretability. Kernel-based approaches, like Gaussian approximation potentials25, use structured, interpretable descriptors—e.g., the smooth overlap of atomic positions26, atom-centered symmetry functions27, or many-body tensor representation28—to interpolate potential energy surfaces with systematic error control and excellent transferability29,30. Conversely, neural network potentials leverage their flexibility as universal function approximators to model high-dimensional, non-linear potential energy surfaces. Recent developments in E(3)-equivariant and message-passing schemes31,32 explicitly encode Euclidean symmetries, enabling physically correct transformations of scalar, vector, and tensor quantities while requiring significantly fewer DFT calculations for model training31. Parallel efforts explore the augmentation of MLIPs with analytic long-range interaction models and continuum elasticity frameworks to better capture the multiscale physics of moiré and strained systems33–36. Despite these architectural advancements, generating accurate models for moiré materials in practice is still nontrivial37–39. Suitable MLIPs for moiré materials must precisely capture the balance between strong intralayer covalent bonds—about 1 eV per formula unit (f.u.)—and weak interlayer van der Waals (vdW) interactions—about 100 meV/f.u. Beyond the training challenge posed by these disparate energy scales, the validation of such models is equally critical. Conventional validation metrics, such as model energy and force root mean square error (RMSE), inadequately assess moiré relaxation quality as they primarily evaluate local atomic environments while relevant moiré structural effects predominantly emerge at length scales beyond those captured by local descriptors. Additionally, the practical limitations of DFT calculations prevent explicit validation of MLIPs on realistic moiré structures. Here, we introduce a split MLIP approach tailored for multilayer moiré systems to address these limitations. By separately modeling the intralayer and interlayer interactions, our method improves accuracy tenfold for force and energy compared to traditional methods, which we refer to as a unified model. Our approach remains agnostic to the specific MLIP architecture used for the interlayer and intralayer components and accurately predicts complex moiré structures across various twist angles and number of layers. Our work also addresses the validation of MLIPs for moiré materials. We demonstrate the limitations of traditional metrics—energy and force RMSE—in assessing the quality of an MLIP and introduce a metric based on the distribution of stacking registries as a robust alternative. We further show that MLIPs can be explicitly validated against ground-truth calculations on quasi-1D systems, which exhibit large domain formation similar to their 2D counterparts while remaining accessible to DFT calculations. Importantly, this correlation ensures that an MLIP’s fitness in describing 1D structures reliably predicts its fitness in describing 2D moiré systems. As a case study, we apply our framework to bilayer HfS2/GaS. We use quasi-1D structures to validate our split MLIP structural relaxations against DFT calculations and highlight the similarity between the electronic properties obtained at these two geometries. For 2D moiré HfS2/GaS, we predict strong relaxation effects, with triangular domain formation that could support unusually low-energy electronic states. Our split MLIP strategy, combined with a rigorous surrogate-based validation protocol, enables accurate and efficient exploration of structural and emergent phenomena in previously inaccessible multilayer moiré systems beyond TMDs. We present our results in four parts: (1) we detail our split MLIP procedure and demonstrate its improvement over standard approaches; (2) we introduce the mean disregistry error (MDE) metric and establish its relationship with RMSE; (3) we demonstrate how quasi-1D-periodic structures effectively capture the accuracy of 2D moiré reconstructions; and (4) we apply this framework to bilayer HfS2/GaS. Results Need for split MLIPs for layered materials We first motivate and introduce our proposed methodology for the generation of MLIPs for moiré reconstructions of layered materials. We begin by expressing the energy E of a material with N atoms at positions {Ri, i ∈ 1, 2, …, N} in terms of tensor contractions, including correlations of atomic positions of increasingly higher order, where N(i) is the bonding neighborhood of atom i, Vm+1(i,j1,j2,…,jm) is the m + 1 body potential between atoms i, j1, j2, …, jm, and Rij are the relative positions of the atoms (Rij = Ri − Rj). This form is used explicitly within most MLIPs, and implicitly even in MLIPs utilizing message-passing31. An important insight from this paper is that when dealing with layered materials, a similar partition can be formally performed at the layer level. Namely, one can partition the total energy of a layered material into the intralayer energies of atoms in isolated layers plus an interlayer energy contribution from the correlation of atomic positions within two, three, and arbitrarily large numbers of layers. This separation is motivated by the fact that stronger covalent bonds dominate the energy within a single layer, while weaker vdW interactions are responsible, in this framework, for the residual coupling between different layers. Current state-of-the-art MLIPs typically construct bonding neighbor lists N(i) by including atoms within some cutoff radius rcut from atom i. Hence, a practical way to split the total energy into intralayer and interlayer contributions is to define different cutoff radii when constructing these two energy contributions (Fig. 1a) and modify the neighbor list to capture these two types of interactions (see SI). For this work, we choose rcut_intra = 6.0 Å and rcut = rcut_inter = 10.0 Å. Separation of length scales for designing accurate MLIPs for layered materials. a Total energy for a layered system can be separated into an intralayer term involving non-interacting layers, Eintra, and a residual interlayer interaction, Einter, which naively requires models with longer truncation distances (denoted in the figure by rcut_inter and rcut_intra). b Distributions of intralayer and interlayer energies within our generated datasets. The distribution of intralayer energies resembles that in the training of MLIPs with standard approaches, but typical interlayer energies are much smaller—on the order of the error σintra of the intralayer MLIP model (~100 meV/atom)— motivating the split model approach. Figure 1b displays the distribution of intralayer and interlayer energies in a typical moiré system, bilayer MoS2/MoS2, obtained by fitting dataset energies, detailed in the next section, to Gaussian distributions. Notably, the energy scales associated with interlayer interactions are much smaller than the intralayer energies, and on the order of the error of typical MLIPs fit to DFT calculations (shaded region in Fig. 1b). This makes it difficult for standard approaches to describe intralayer and interlayer interactions accurately, e.g., training unified MLIPs from molecular dynamics simulation, which we explicitly quantify later. In the next section, we address how to efficiently generate datasets and train different MLIPs for the intralayer and interlayer contributions. Dataset generation protocol for split MLIPs A key aspect in generating MLIPs is that the training dataset must sample a diverse set of atomic configurations that capture the bonding environment relevant for subsequent evaluations of the MLIP. A standard approach when training MLIPs is to sample configurations from molecular dynamics trajectories obtained from ab initio or surrogate models. While such an approach could directly be used in training intralayer MLIPs, we choose a more straightforward procedure here, in which we apply random perturbations to the atomic positions and cell parameters starting from a suitable supercell (3 × 3 for the calculations in this paper). The advantage of this procedure is that this dataset can simply be generated from parallel DFT calculations and minimizes the chance of sampling unphysical trajectories that are more likely with two-dimensional materials. We then predict the energies and forces of these structures using DFT with a fully non-local vdW-corrected functional40–42 to ensure the intralayer and interlayer datasets were generated with similar DFT parameters (see “Methods”). While other vdW corrections may be desired (e.g., to improve the predicted energies or provide computational savings), our approach is agnostic to their functional form, and the models can be retrained or fine-tuned on more advanced methods as needed. For the training of interlayer MLIPs for a particular pair of layered materials, we generate a dataset that samples various interlayer distances and local stacking environments. In particular, we first construct small unit cells or supercells appropriate to hold the relevant bilayer structure. This work focuses on relaxations near 0° twist; hence, we consider a bilayer at a 0° twist angle. Our model for the bilayer at 0° is quite accurate on the configurations at 60° as well (interlayer force RMSE: 32.86 meV/Å, interlayer energy RMSE: 6.32 meV/atom), making it generalizable to twist angles quite far away from 0° as well. Such large generalizability to different twist angles was previously reported for classical force fields19, and can be rationalized from steric effects that dominate the interlayer potential energy surface. Of course, one can improve this accuracy by adding other twisted structures into the dataset, including 60° or other intermediate angles, which could be more important for non-TMDC bilayers. Still considering the interlayer dataset, we sample various layer separations d, more finely near equilibrium and coarsely otherwise (see “Methods”). We also sample a set of 12 × 12 stacking configurations, out of which only a 6 × 6 subset is kept in our training and validation sets, with the remaining configurations forming our test set. When computing the interlayer energy and forces for a particular bilayer configuration, we subtract the intralayer contributions from each separate layer. We did not randomize the atomic positions within each layer when constructing the interlayer dataset. A similar dataset generation method has also been used to design MLIPs for hetero-interfaces43. For benchmark purposes, we also consider unified MLIP models not separated into intralayer and interlayer parts. For these models, we create a dataset formed by the union of the intralayer and interlayer datasets. Split MLIPs benchmark: tenfold improvement of energy and force RMSE metrics We now present the first clear benefits of our proposed split MLIP approach for a prototypical bilayer of MoS2/WSe2. We train intralayer MLIPs (WSe2, MoS2) and an interlayer MLIP (MoS2/WSe2) on the split datasets using the Allegro model44, though we find similar results with other MLIPs (Fig. S1 as described in the Supplementary Information). In Fig. 2a, we compare the RMSEs evaluated on our hold-out WSe2 intralayer test datasets for several models: a random, mean-guessing baseline, the classical modified Stillinger-Weber45, a unified MLIP with a cutoff of 6 Å, and the split intralayer MLIP with a cutoff of 6 Å. All MLIPs outperform the classical SW potential and, notably, similar split and unified MLIPs (i.e., constructed with a similar number of trainable parameters) display a similar accuracy on the intralayer dataset. This is expected, as the total energy of a unified MLIP is dominated by the intralayer contribution (Fig. 1b). The accuracy of both unified and split MLIPs in describing intralayer interactions can be improved by making the cutoff radius of the model larger (Fig. S1a as described in the Supplementary Information). Performance of classical interatomic potentials and both unified and split machine-learned interatomic potentials (MLIPs) on our hold-out test datasets. a Absolute residual energy, relative to DFT, for various intralayer models on a hold-out intralayer test set of 494 structures. Error bars mark the interquartile range, and dots mark the median absolute residuals. Horizontal lines are the maximum desired model root mean squared errors (RMSEs). See text for more details. Split and unified models perform similarly on the intralayer predictions when using similar model parameters. b Similar comparison for interlayer models (Kolmogorov-Crespi (KC)) on a hold-out interlayer test set of 3039 structures. Here, even small split MLIPs (6080 neural network (NN) trainable parameters) outperform large unified MLIPs (109824 trainable parameters). Additional details on force residuals and data on the split approach implemented with the MACE48 MLIP are provided in Fig. S1 in the SI. Next, we benchmark the ability of several models to capture interlayer interactions. In Fig. 2b, we evaluate the RMSEs on our hold-out interlayer test datasets for a random-guessing baseline, the classical Kolmogorov-Crespi (KC) potential, a unified MLIP with 109824 trainable parameters, and two split MLIPs with 6080 and 109824 parameters. Notably, the performance of the unified MLIP model is comparable to the classical KC model. However, the split model, even with significantly fewer model parameters, outperforms the unified MLIP. The lower error in these split MLIPs is required for achieving an energy RMSE below a threshold of 1 meV/atom associated with accurate moiré relaxations, which we discuss in the following section. Our results for the split models are not unique to the Allegro MLIP architecture tested here, and we observe similar trends when employing the MACE46 potential (Fig. S1 as described in the Supplementary Information). In fact, one could even use different model architectures for each interaction, e.g., a foundation model47–49 for the intralayer and a system-specific classical potential50 for the interlayer model. Indeed, our benchmarks on two foundation models—Message passing Atomic Cluster Expansion trained on the Materials Project database of trajectories (MACE-MP0)47 and GRaph Atomic Cluster Expansion trained on the Materials Project database of trajectories (GRACE-2L-MPtrj)49—show that, while they perform reasonably well for intralayer interactions, they significantly underperform in generalizing to interlayer tasks compared to our trained interlayer model (see Fig. S2 as described in the Supplementary Information). The relative success of the KC potential for interlayer interactions lies in its ability to qualitatively capture isotropic attraction together with registry-dependent repulsion. In particular, KC uses the same functional form for the attractive term as a LJ potential, but its repulsive term differs substantially: it captures anisotropic orbital π-overlap effects that are critical for moiré-domain formation in graphene, hBN, and TMDs, and that are not captured by an LJ potential. This repulsive interlayer interaction is also sensitive to local chemical bonding, making it nontrivial to reproduce with simpler models. For instance, in Fig. S1, we show that foundation models, even when augmented with explicit attractive interlayer vdW interactions, cannot adequately describe the registry-dependent interlayer energies, suggesting that their training data (or learned representation) does not adequately capture the repulsive interlayer physics. Related ideas are also being explored: a recent work also separates interlayer and intralayer interactions using a hybrid of monolayer-trained neural potentials and analytic interlayer models, reporting an improvement in model performance51. Yet, the advantage of using an interlayer MLIP, as we propose here, is its broad generalizability to different materials. Finally, we note that separating the intralayer and interlayer interactions also makes the splitting procedure and dataset generation model-agnostic, allowing the split approach to benefit from the rapid pace of MLIP improvements. This approach also simplifies the training and validation of high-quality intralayer MLIP models, given that the number of required intralayer MLIPs grows linearly with the number of layers one is interested in simulating, but the number of pairwise interlayer MLIPs grows quadratically. We also note that we obtain the same conclusions when analyzing the force RMSE across these models (see Fig. S1 as described in the Supplementary Information). While we have demonstrated that splitting the MLIPs achieves better energy and force RMSE on the test datasets, our ultimate goal is to accurately determine the structure of moiré materials. We next show that the energy and force RMSE are, unfortunately, not good predictors of a model’s ability to describe moiré relaxations. For that, we introduce a metric to assess the quality of the structural relaxation. This allows one to determine the maximum permissible energy and force RMSEs (solid lines in Fig. 2), and to create a general approach to designing accurate classical and machine-learned IPs for moiré structure relaxation. Limitations of energy and force errors as validation metrics MLIPs are typically trained by minimizing a loss function based on the energy and force RMSE relative to ground-truth values, often from DFT calculations25,52. This training process relies on feature vectors in the MLIPs that encode the local atomic environment (e.g., Eq. (1))25,30,53,54. So, while the energy and force RMSE is a convenient optimization target for most MLIPs, it does not directly assess the relaxation of moiré materials, which critically takes place at longer, nanometer scales. We illustrate these limitations using MoS2/WSe2 bilayers in Fig. 3. In these heterostructures, small energy asymmetries (~2 meV/atom difference) between the AB and BA stacking configurations fundamentally determine moiré-scale domain formation, domain walls, and the resulting electronic properties17. This energy difference is comparable to the residual interlayer energies of traditional, unified MLIPs (Fig. 2b), suggesting that such models may not describe well the structure of such a moiré bilayer. Hence, we first restrict our discussion to unified MLIPs to qualitatively understand the relationship between model RMSE and large-scale structural features (Fig. 3a–e). Validation metrics for moiré structural effects in twisted MoS2/WSe2 heterostructures. a Atomic configurations at high-symmetry stacking positions (AA, AB, and BA), showing W atoms (colored hexagons), reference Mo atoms, and disregistry vectors (arrows). b Interlayer potential energy landscape per atom versus displacement (δry) for a ground-truth density functional theory (DFT) prediction and two machine-learned interatomic potential (MLIP) models trained on a unified dataset: Unified Model 1 (UM1) with a neighbor cutoff radius of 6 Å, and Unified Model 2 (UM2) with a neighbor cutoff radius of 10 Å. Vertical lines indicate high-symmetry stacking configurations: AA (red), AB (blue), and BA (green). c Voronoi-centered disregistry vectors mapped onto the pristine MoS2 lattice (blue background cell), showing the high-symmetry stacking positions AA, AB, and BA. The color scale indicates the interlayer potential energy relative to AA stacking, and the dashed arrow marks the displacement path along the y-direction. d Atomic relaxation patterns in a 1.1∘ twisted MoS2/WSe2 heterostructure predicted using UM1 and UM2. The color scale indicates local disregistry magnitude at Mo sites, revealing domain formation through regions of constant disregistry. e Distribution of Voronoi-centered disregistry vectors of the same twisted structures in panel d, highlighting domain asymmetry through the density of vector clustering near high-symmetry positions. f Structural validation of moiré geometry using the mean disregistry error (MDE) metric versus model energy prediction root mean squared errors (RMSEs) for MLIPs trained with varying degrees of model corruption. Error bars represent the 25th and 75th quartiles across multiple noise seed initializations. Shaded regions denote the desired target MDE (gray, below 0.01 Å), and the maximum interlayer energy RMSE (green, ≤4 meV/atom) and maximum intralayer energy RMSE (orange, ≤200 meV/atom) to achieve that target MDE. To test the model validity of standard RMSE metrics, we analyzed a 0°-twisted MoS2/WSe2 bilayer with lattice-matched primitive cells. We evaluated total energy across a systematic displacement path of the top W layer relative to the bottom Mo layer (Fig. 3a). This displacement generates a path in the configuration space, which encodes the relative offset between corresponding atoms in adjacent layers (Fig. 3a). We then evaluated the total energy for this set of structures using two similar MLIP models trained on identical unified datasets but with different cutoff radii: 6 Å for the unified model 1 (UM1) and 10 Å for the unified model 2 (UM2). Both models UM1 and UM2 achieved low energy RMSE values relative to the DFT ground truth (11.6 meV/atom for UM1 and 10.6 meV/atom for UM2), well below the typical energy RMSE of 100 meV/atom target for many MLIP training protocols. Despite their relatively low RMSE values, both models incorrectly predict AB stacking as more energetically favorable than BA by about 7 and 11 meV/atom for UM1 and UM2, respectively, which has quantifiable consequences for the models’ abilities to predict the structure of moiré systems. Clearly, a preferred approach for validating MLIPs is to compare the relaxed moiré geometries directly. In principle, traditional MLIP descriptors could be utilized for this, as they often display desired properties such as being invariant under global rotations, translations, and relabeling of atomic species. However, common descriptors typically lack simple physical interpretability, e.g., how their various components translate into the similarity between moiré structures in real-space. To directly obtain a physically interpretable comparison metric, we compare moiré structures at a coarser level in terms of disregistry vectors, which encode local stacking environments in terms of a real-space displacement vector δr of one layer relative to a reference high-symmetry stacking configuration (Fig. 3a). For our calculations on MoS2/WSe2, we define δr = 0 when the W and Mo atoms are on top of each other (i.e., the AA stacking for a 0°-twisted structure). The vector space associated with the disregistry vectors is the configuration space. Disregistry vectors are not uniquely defined and are often expressed within effective (strained) primitive cells—constructed following a consistent global order to tile any given bilayer material20,21,55,56. We avoid this requirement in our work by instead introducing Voronoi-centered disregistry vectors. In short, we define δr to have the smallest magnitude, which uniquely defines each vector and increases their applicability to moiré materials with larger strain inhomogeneities (see SI for construction details as described in the Supplementary Information). To illustrate the Voronoi-centered disregistry vector approach, we map the disregistry path in Fig. 3b onto a denser path in configuration space in Fig. 3c, where each hexagon at δr represents a structure wherein the W atom on one layer is laterally displaced from the nearest Mo atom in the adjacent layer by δr. When applied to an explicit moiré superlattice, this mapping transforms collections of local atomic stacking environments into a comprehensive fingerprint of the moiré geometry29. As shown in Fig. 3e, relaxed bilayer structures exhibit clustering of disregistry vectors around energetically favorable stacking configurations, with the density of vectors proportional to the corresponding domain area in real-space. The spatial distribution of these vectors directly reflects the balance between interlayer stacking energy minimization and elastic energy penalties from deformation and enables quantitative comparison of structural predictions across different models. We now revisit the differences between UM1 and UM2 by analyzing their relaxations in terms of their disregistry vectors. We first plot the relaxed structure for a 1.1° twisted MoS2/WSe2 predicted by these two MLIPs in Fig. 3d, showing the magnitude ∣δr∣ for each atomic position in the moiré cell. UM2, which significantly (and incorrectly) favors AB stacking, produces larger AB moiré domains compared to UM1. Critically, even though both models have similar and relatively small energy RMSE metrics, they predict domain walls that differ by 40%. We also analyze these trends directly in configuration space in Fig. 3e. For UM2, the increased density of vectors in the AB region reflects the model’s favoring of the AB stacking configuration. To compensate for its energy imbalance, UM2 also displays a sharper transition between stacking configurations, with a large density of disregistry vectors at the AA stacking configuration. Still, while these visual representations provide valuable insights, a quantitative comparison of the stacking disregistry distributions is still needed to rigorously assess the global structural discrepancies between models. Quantifying moiré structural dissimilarity: the mean disregistry error (MDE) We employ an approach rooted in optimal transport theory to quantify structural differences between moiré predictions from competing models. We adapt the concept of the Wasserstein distance—a measure of the cost to transform continuous probability distributions into one another—to our discrete case of disregistry vectors. Since disregistry vectors form a discrete set rather than a continuous distribution, we formulate this transformation as a minimum-cost bipartite matching problem, efficiently solvable via the Hungarian algorithm57. This approach naturally connects to the physical concept of disregistry flow in configuration space during structural relaxation20 and captures moiré structural features that traditional RMSE metrics miss58. We therefore propose a quantitative measure of the similarity between two moiré systems and, hence, the fitness of an MLIP: the mean disregistry error (MDE), defined as the mean distance between two sets of matched disregistry vectors in configuration space. MDE quantifies the mean structural deviation between two moiré systems by comparing their Voronoi-centered disregistry vector distributions in configuration space, which encode the local stacking arrangement between layers. Specifically, where δri and δrπ(i)′ denote the disregistry vectors from two structures, and the minimization is over all bipartite matchings π(i) that respect periodic boundary conditions in the pristine Voronoi cell. Critically, the MDE remains invariant under rigid translations, rotations, or atomic species relabeling, making it a robust metric for comparing moiré structures predicted by different models (see SI for detailed derivation as described in the Supplementary Information). We can now assess whether the MDE correlates with conventional metrics for validating MLIPs, such as the energy RMSE. As training a comprehensive range of models with a uniform distribution of RMSEs is difficult, we instead apply controlled random noise to the weights of a single well-performing split MLIP model. We scale the noise relative to the standard deviation of the model weights within each neural network layer, thereby creating a continuous corruption factor. To minimize bias from the random noise, we average each calculated corruption level over an ensemble of ten initial random seeds. We analyze intralayer and interlayer models separately by fixing one component while systematically perturbing the other, allowing us to independently quantify the error contributions from the intralayer and interlayer MLIPs. Our analysis (Fig. 3f) reveals a strong positive correlation between energy RMSE and moiré structural prediction quality as measured by MDE. However, we observe two important features that highlight the shortcomings of the RMSE as a validation metric for moiré materials. First, the RMSE metric exhibits significantly higher variance compared to the MDE (nearly half an order of magnitude) for equivalent corruption factors. This is consistent with the idea that the MDE is more sensitive to domain symmetry and stacking distributions, making it less sensitive to fluctuations arising from model performance. Second—and more critically—for a fixed corruption factor and resultant MDE, the intralayer model energy RMSE is nearly two orders of magnitude larger than that for the interlayer model. This finding has profound consequences for the design of MLIPs for moiré materials. It shows that one should train MLIPs with a much stricter RMSE for interlayer interactions than intralayer ones. Specifically, achieving an MDE below 0.01 Å (comparable to the in-plane displacement associated with the zero-point motion of atoms in monolayer MoS2) requires interlayer and intralayer energy RMSE values of approximately 4 meV/atom and 200 meV/atom, respectively. Given the impracticality of designing unified MLIPs that simultaneously satisfy such disparate accuracy requirements for moiré systems, these results provide quantitative justification for our split modeling approach that treats interlayer and intralayer interactions separately. Practical validation of MLIPs: surrogate quasi-1D structures While the MDE analysis is effective for assessing the accuracy of MLIPs on small systems, applying this framework to large-scale 2D moiré systems is computationally prohibitive due to the significant cost of generating high-accuracy, ground-truth relaxed structures for comparison. For instance, producing the long-range effects characteristic of moiré systems, such as domain reconstruction and out-of-plane buckling, requires simulation supercells with lattice constants exceeding ~10 nm59. For twist angles below 1°, moiré lattices can surpass 20 nm, resulting in supercells with over 10,000 atoms17,20,56 – impractical for DFT calculations. To address these challenges, we propose quasi-one-dimensional (Q1D) moiré structures as surrogate systems for validating MLIPs in 2D moiré materials. Q1D structures can be physically realized by uniaxially straining a bilayer structure and creating a supercell with a long moiré lattice parameter ℓm but a small width w. These structures retain the critical features of long-range moiré reconstruction along ℓm while reducing the number of atoms across their width60,61. This dimensionality reduction allows the number of atoms to scale linearly with ℓm, in contrast to the quadratic scaling with moiré lattice constants for twisted 2D moiré systems. This scaling improvement enables one to directly perform ground-truth DFT calculations to assess the performance of MLIPs in capturing long-range moiré phenomena with the previously introduced MDE metric. We present an illustrative Q1D structure comprising a stretched strip of 24 WSe2 unit cells on top of a compressed strip of 25 MoS2 unit cells (Fig. 4a), generating a 1D moiré pattern associated with a ~0.07% uniaxial strain. We note that Q1D structures can also host non-zero interfacial twists, and hence serve as surrogate systems for validating the MLIP performance on arbitrarily twisted 2D moiré materials (see the SI for the detailed construction scheme as described in the Supplementary Information). Initially, the W and Mo atoms are aligned along the x-axis (Fig. 4b). Upon relaxation, the top layer shifts along the y-axis, producing a periodic modulation of stacking configurations along the length (Fig. 4c). This shift reflects the balance between stacking energies and elastic strain, as the system minimizes its total energy by favoring lower-energy stacking configurations (AB and BA) while limiting elastic deformation20. In the initial, unrelaxed configuration, the Voronoi-centered disregistry vectors align along a linear path parallel to the x-axis. After relaxation, these vectors trace a sinusoid-like path near the Voronoi cell edges, reflecting the stacking modulation along the Q1D structure’s length. This relaxation behavior illustrates how the atoms migrate from a configuration path that passes through the high-energy near-AA stacking regions towards a path through the lower-energy AB and BA regions. We note that these Q1D systems could be generated along a diverse set of paths in configuration space to strengthen the validation procedure further. One-dimensional periodic moiré bilayers as a surrogate system for two-dimensional twisted structures. a Quasi-1D (Q1D) heterostructure: 24-unit-cell strip of WSe2 (a = 3.319 Å) aligned with 25-unit-cell strip of MoS2 (a = 3.184 Å). The lattice mismatch requires ~0.07% strain in the WSe2 layer to achieve periodicity. The initial, unrelaxed structure (transparency) is overlaid onto the relaxed structure. Shifting of local disregistry and out-of-plane buckling is observed. b Top view comparing unrelaxed (top) and DFT-relaxed (bottom) structures. Insets show local stacking configurations at three positions, with hexagons marking top-layer W atoms and indicating relative disregistry from bottom-layer Mo atoms (colored by indexed position along the strip). c Voronoi-centered disregistry analysis of unrelaxed (top) and relaxed (bottom) structures. Colored hexagons represent disregistry vectors of W atoms, demonstrating relaxation-induced displacement along the short axis that reduces stacking energy while increasing strain. d Validation of structural predictions, using the mean disregistry error (MDE), comparing Q1D and 2D moiré structures. These machine-learned interatomic potentials (MLIPs) were trained on our intralayer and interlayer DFT datasets, and we use the uncorrupted model as a reference (see text). Models are corrupted following the same scheme as in Fig. 3, and error bars represent the 25th and 75th quartiles across multiple noise seed initializations. Notably, the performance of the split MLIP on the 1D structure is a good predictor of the quality of the MLIP on the 2D moiré structure. To demonstrate the effectiveness of Q1D structures as conservative surrogates for capturing moiré relaxation in large 2D twisted moiré systems, we evaluate the MDE predicted by MLIPs on both Q1D and 2D structures in Fig. 4d. Specifically, we consider a 2D MoS2/WSe2 bilayer with a 3.17° twist (~1300 atoms) and a Q1D heterostructure (~147 atoms) constructed to match the same moiré length scale (~5 nm). Since direct DFT relaxation of the 2D structure is computationally prohibitive, we take the uncorrupted split MLIP (trained on DFT forces and energies) prediction as a ground truth reference for both geometries, enabling the construction of a tractable validation dataset. To assess model robustness, we apply a controlled corruption procedure to the trained MLIPs, similar to that in Fig. 3f. Across a range of corruption levels, the MDEs between corrupted and reference relaxations remain strongly correlated between Q1D and 2D structures. Notably, Q1D systems consistently exhibit greater sensitivity to model corruption, expressed in larger MDE values, which we attribute to their increased freedom for transverse relaxation (Fig. 4d). This suggests that MLIPs performing well on Q1D structures will perform at least as well on the corresponding 2D ones, providing a conservative and reliable validation approach for MLIPs instead of the traditional, computationally intractable 2D moiré systems. To summarize, we have presented an approach to systematically train and validate MLIPs for moiré reconstruction of large arbitrary bilayers. Further technical details on the practical usage of this approach are included in the SI as described in the Supplementary Information. Explicit validation: medium-sized and twisted 2D moiré structures To directly validate the predictive accuracy of our split MLIP on large-twist angle and medium-sized twisted 2D structures, we evaluate two representative cases: a) a 1317-atom MoS2/WSe2 bilayer at 3.17° twist, and b) a 120-atom MoS2/WSe2 bilayer at ~12° twist. These systems allow us to assess the model’s performance in both the low- and intermediate-angle regimes, where moiré-scale reconstruction is more and less prominent, respectively. For the larger 1317-atom structure, a full DFT geometry relaxation is computationally prohibitive. Instead, we select 10 snapshots with equally spaced atomic displacements along the relaxation trajectory generated using the Allegro-based split MLIP. For each snapshot, we perform single-point, self-consistent DFT calculations to evaluate the total energy and compare it with the corresponding MLIP prediction (relative to the initial configuration). As shown in Fig. 5a, the MLIP-predicted energies closely follow the DFT energies throughout the trajectory, indicating that the model faithfully descends the true DFT energy landscape. Explicit validation of the split machine-learned interatomic potential (MLIP) against DFT for twisted MoS2/WSe2 moiré structures. a Split MLIP relaxation trajectory of a 1317-atom, 3.17∘ twisted MoS2/WSe2 bilayer. Energy per atom is plotted at equal-displacement steps along the Allegro split MLIP and DFT trajectories. After the final MLIP step, the structure is further relaxed for 6 steps using DFT (gray shaded region). The final structures differ by a mean disregistry error (MDE) of 0.007 Å. b Validation of the split MLIP energy calculated along a DFT relaxation trajectory for a 120-atom MoS2/WSe2 bilayer at 12∘ twist. Relative energy per atom at selected snapshots along the DFT relaxation trajectory, calculated from both the Allegro46 split MLIP and DFT. The MLIP energies closely follow the DFT trajectory, with an overall root mean squared error (RMSE) of just 0.1 meV/atom. To quantify structural agreement, we further relaxed the final MLIP-relaxed structure using DFT for six additional Broyden-Fletcher-Goldfarb-Shanno (BFGS) steps, converging the forces below the 0.03 eV/Å force threshold of our training set. The resulting DFT-relaxed structure exhibited only a minor deviation from the MLIP prediction, with an MDE of 0.007 Å, within our 0.01 Å accuracy criterion for moiré-scale fidelity. Residual force magnitudes along the trajectory also remain within our desired threshold (see Fig. S3b as described in the Supplementary Information), reinforcing the physical consistency of the MLIP-generated relaxation. To further test the generalizability at higher twist angles, we consider a 120-atom bilayer at ~12° twist, small enough to permit a full DFT relaxation. Starting from a rigidly twisted initial configuration, we performed a full structural relaxation using DFT, with forces converged below 0.03 eV/Å. From this trajectory, we selected 12 frames of equal atomic displacement and calculated the corresponding energies using an Allegro-based split MLIP model. As shown in Fig. 5b, the MLIP energies closely track the DFT energy descent path, achieving an energy RMSE of less than 0.1 meV/atom. As before, residual force magnitudes are also within our desired threshold (see Fig. S4b as described in the Supplementary Information). Notably, the energy RMSE observed for this bilayer is significantly smaller than the (already small) RMSE observed for the larger 1317-atom structure, even though the twist angle of this 120-atom bilayer (~12°) is farther from that used in our training dataset (0°). The higher accuracy of our split MLIP here can be attributed primarily to minimal structural reconstruction at this twist angle. In TMDC bilayers with twist angles far from ~0° and ~60°, there is minimal in-plane reconstruction, and relaxation is dominated solely by changes in interlayer spacing. Additionally, for this 120-atom bilayer, we independently relaxed the structure using the split MLIP and the FIRE optimization algorithm62 instead of DFT. Despite employing a different relaxation algorithm from the BFGS algorithm used by Quantum ESPRESSO, the final MLIP-relaxed structure shows excellent structural agreement with the DFT results, yielding an MDE of only 0.001 Å. This close match, irrespective of the relaxation method, strongly confirms the MLIP’s capability to accurately predict the detailed registry and atomic positions associated with the true DFT energy minimum. Example: emergent triangular lattice in twisted bilayer HfS2/GaS To showcase the applicability of the surrogate approach to understudied 2D moiré bilayers, we apply it to predict the relaxations of the HfS2/GaS bilayer. HfS2, known for its higher predicted mobility compared to traditional TMDs63, has been studied as a promising material for 2D transistors64 and ultra-sensitive phototransistors65. Similarly, monolayer GaS, with its wider band gap, has been actively studied for its optical properties66,67. The comparable lattice parameters of these materials allow for the formation of large moiré lattices with potentially significant reconstruction. Recent first-principles calculations of HfS2/GaS bilayers68, along with related gallium chalcogenide heterostructures69–71, reveal their potential for photocatalytic water splitting, attributed to a type-II band alignment and efficient charge separation, tunable via strain engineering. However, the relaxation of large moiré bilayers in this system has not yet been studied. We first generate an interlayer dataset and train an interlayer model specifically for the HfS2/GaS bilayer, while we model the intralayer interactions using a MACE foundation model. Using a generic foundation model here is consistent with our finding that one can use a less accurate intralayer MLIP compared to the interlayer one, and is sufficient here to capture the atomic and electronic structure for this bilayer. The intralayer interactions may be modeled using either (i) a pretrained foundation model, (ii) a fine-tuned foundation model, or (iii) a system-specific intralayer MLIP trained from scratch. In general, we expect the latter two to perform better than a foundation model without any tuning. The Q1D validation tests introduced here and performed for HfS2/GaS provide a practical way to determine whether a foundation model is sufficiently accurate for the intralayer component in a given bilayer system. Following our Q1D approach, we construct a 1D commensurate cell (0.99° twist, 301 atoms, ℓm = 15.6 nm), shown in Fig. 6a, and relax it using both DFT and our split MLIP framework. Comparing the two relaxed structures, we find a small MDE of 0.02 Å—very close to our previously established, and somewhat arbitrary, desired threshold of 0.01 Å (Fig. 3f), despite employing a simplified intralayer MLIP from a foundation model. In contrast, the MDE between the initial structure and final DFT-relaxed structure is much larger, 0.48 Å, highlighting that the composite MLIP captures the significant relaxation predicted from DFT (Table 1). Additionally, the energy of the MLIP-relaxed structure, when evaluated within DFT, is only 2.5 meV/atom higher than the DFT-relaxed structure, within the desired accuracy range. We also compute the cleavage energies of this Q1D bilayer and find that the MLIP-predicted energy is in remarkable agreement with DFT predictions (Table 2). The cleavage energies Eclv were computed as Eclv=(Ebl−NGaSEstrGaS−NHfS2EstrHfS2)/nbl, where Ni is the number of formula units of structure i, Ebl is the energy of a bilayer structure, nbl is the number of atoms in the bilayer, and Estri is the total energy of material i in its strained unit cell. Applicability of split MLIP approach to HfS2/GaS bilayers. a Quasi-1D (Q1D) structure of bilayer HfS2/GaS before (transparent circles) and after (solid circles) relaxation by our split machine-learned interatomic potential (MLIP) model. b Band structure of the moiré system unfolded onto a strained GaS unit cell Brillouin zone (BZ) (see Fig. S5 as described in the Supplementary Information), computed within density functional theory (DFT) at the initial (yellow), MLIP-relaxed (red), and DFT-relaxed (blue) structures. Continuous curves denote the band structure of a single strained GaS monolayer. Inset highlights the band offsets between the different structural configurations. c Similar to (b), but unfolded onto a strained HfS2 cell. d MLIP-relaxed 2D moiré cell of GaS/HfS2, along with the magnitude of the displacement vectors, showcasing the significant reconstruction and formation of triangular domains. Table 1Validation of the split machine-learned interatomic potential (MLIP) model for bilayer HfS2/GaS on the quasi-1D structureStructureDFT total energy error (meV/atom)DFT Hamiltonian error (%)MDE (Å)MLIP relaxed2.521.030.02Initial13.9847.000.48We compute several properties from DFT at either the MLIP-relaxed or the initial, rigidly strained structure and compare their deviation from the ground-truth values obtained at the DFT-relaxed structure. Comparison metrics include the DFT total energy error, the relative error in the Kohn-Sham DFT Hamiltonian, and the mean disregistry error (MDE). Validation of the split machine-learned interatomic potential (MLIP) model for bilayer HfS2/GaS on the quasi-1D structure We compute several properties from DFT at either the MLIP-relaxed or the initial, rigidly strained structure and compare their deviation from the ground-truth values obtained at the DFT-relaxed structure. Comparison metrics include the DFT total energy error, the relative error in the Kohn-Sham DFT Hamiltonian, and the mean disregistry error (MDE). Table 2Comparison of cleavage energies, in meV/atom, between split machine-learned interatomic potential (MLIP) and DFT for bilayer HfS2/GaS on the quasi-1D structureStructureCleavage energy from MLIPCleavage energy from DFTError in cleavage energyInitial31.74730.859+0.888MLIP relaxed32.34534.742−2.397DFT relaxed34.86434.020+0.844 Comparison of cleavage energies, in meV/atom, between split machine-learned interatomic potential (MLIP) and DFT for bilayer HfS2/GaS on the quasi-1D structure Next, we assess the accuracy of the MLIP in predicting not only the moiré structure but also the electronic structure. We compute the electronic band structure, within DFT, of the Q1D structures relaxed with both DFT and the composite MLIP model. To make the interpretation easier, we unfold the computed band structures onto the primitive cell of either monolayer HfS2 (Fig. 6b) or GaS (Fig. 6c), revealing that the electronic structure evaluated at the MLIP-relaxed structure is very close to that evaluated at the DFT-relaxed structure. Finally, we quantify the similarity of the two electronic structures by computing the relative change of the occupied Kohn-Sham Hamiltonian, ∥H − Href∥F/∥Href∥F, where Href is the occupied subspace of the Kohn-Sham Hamiltonian obtained at the reference, DFT-relaxed structure, H is the Hamiltonian evaluated at either the MLIP-relaxed or initial structure, and ∥ ⋅ ∥F is the Frobenius norm (see SI section on computing error in the Hamiltonian as described in the Supplementary Information). Notably, this difference is only 1% for the MLIP-relaxed structure and 47% for the initial structure. Altogether, this validation protocol using Q1D structures demonstrates the ability of the split MLIP for HfS2/GaS to accurately capturing atomistic relaxations necessary to describe moiré domains and the electronic structure in the system. Finally, we apply the MLIP to a relaxation of a 2D moiré superlattice (2.86° twist, 5163 atoms, ℓm = 6.8 nm), revealing the formation of clear triangular-like domains in the relaxed structure (Fig. 6d). Such domains are qualitatively different from those obtained in twisted TMDs close to either 0° or 60° twist angles and hint at the possibility of realizing a Kagome lattice in such a system. Future DFT calculations utilizing large-scale, first-principles approaches, or machine-learned Hamiltonians can resolve the emergent electronic properties of such bilayer systems now that the relaxed moiré structure—which is challenging to obtain from DFT—is accessible. Discussion In this work, we established a comprehensive framework for constructing and rigorously validating MLIPs specifically designed for multilayer 2D materials. By explicitly separating the total energy and atomic forces into distinct intralayer and interlayer components, we demonstrated a tenfold improvement in predictive accuracy compared to conventional unified models, particularly highlighted in the MoS2/WSe2 material system. Our comparative analysis revealed significant limitations of traditional validation metrics, such as RMSE, which fail to capture the structural accuracy in large-scale moiré reconstructions. To address these limitations, we introduced a physically intuitive and computationally efficient quantitative validation metric based on the Wasserstein distance, the mean disregistry error (MDE). This generalizable metric quantifies structural accuracy by comparing Voronoi-centered stacking disregistry vector distributions, thereby explicitly assessing the nanometer-scale distribution of local stacking disregistries crucial for moiré phenomena. By systematically evaluating models with controlled corruption, we established a clear correlation between energy RMSE and the MDE metric, providing actionable accuracy thresholds for practical IP design: specifically, a stringent requirement of ~4 meV/atom for interlayer interactions and a more tolerant threshold of 200 meV/atom for intralayer interactions to achieve MDE values below 0.01 Å. These findings provide a crucial, physically motivated justification for our split modeling approach. Furthermore, recognizing the computational impracticality of validating large two-dimensional moiré structures directly with first-principles methods, we proposed quasi-one-dimensional (Q1D) surrogate systems. These Q1D surrogates successfully preserve essential long-range structural and energetic features, making first-principles ground-truth validations computationally feasible. By demonstrating that MDE values in these surrogate systems correlate strongly with those in full 2D moiré structures, we confirmed Q1D systems as reliable validators for MLIPs. This allowed us to confidently extend our validation to previously uninvestigated materials systems, exemplified by our analysis of the twisted HfS2/GaS bilayer. Our method accurately captured subtle relaxation effects (achieving an MDE of 0.02 Å), predicted the formation of triangular domain patterns at small twist angles, and reproduced electronic structures consistent with high-fidelity DFT calculations, validating the robustness of our approach. Overall, the split interatomic potential approach developed here, combined with the quasi-1D surrogate validation strategy and the introduction of the MDE metric, provides a robust, scalable, and rigorous validation pathway for accurately predicting relaxation-driven emergent phenomena in these large-scale multilayer moiré systems. Importantly, this pipeline is generalizable and model-agnostic, readily adaptable to other layered material systems, and establishes a foundation for extending layered materials design beyond conventional twisted bilayer systems to more complex multilayer architectures with predictive reliability and computational efficiency. Methods First-principles calculations Dataset generation All DFT calculations for MoS2 and WSe2 were run using Quantum ESPRESSO (v.7.2)72. The self-consistent field (SCF) calculations were run with a plane-wave cutoff of 60 Ry with a self-consistent field (SCF) convergence threshold of 10−8 Ry. An unshifted Monkhorst-Pack k-point sampling of 12 × 12 × 1 was used in the primitive cells and the structures in the interlayer dataset. For the intralayer dataset (constructed from 3 × 3 × 1 supercells of MoS2 or WSe2 primitive cells), an unshifted Monkhorst-Pack k-point sampling of 4 × 4 × 1 was used. For the calculations involving MoS2 or WSe2, we used the scalar-relativistic norm-conserving pseudopotentials from the PseudoDojo database73. For bilayer HfS2/GaS, we used a plane-wave cutoff of 80 Ry with an SCF convergence threshold of 10−6 Ry. An unshifted Monkhorst-Pack k-point sampling of 12 × 12 × 1 was used in the primitive cells and the structures in the interlayer dataset. For these calculations involving HfS2 and GaS, we used the scalar-relativistic norm-conserving pseudopotentials from the SG15 database74. To include vdW interactions, all calculations used the vdw-df-c09 functional42. For the intralayer dataset, we generate 1,000 random perturbations to the atomic positions and cell parameters starting from a 3 × 3 supercell. Specifically, each Cartesian component of every atom is perturbed (rattled) according to a Gaussian distribution with a standard deviation σ = 0.1 Å. Similarly, we apply a random fractional change to each in-plane lattice constant (σ = 5%) and the angle γ between the two in-plane lattice vectors (σ = 1%). These structures are then split randomly for training, validation, and test sets. For the interlayer dataset, we first determined an effective interlayer bonding distance, l0, by summing the covalent radii75 of the closest two atoms on different layers. We sample various interlayer separations d in steps of 0.1 Å between the bond length of the two closest atoms l0 and 2l0. We also sample d in steps of 0.5 Å between 2l0 and 3l0 to capture the decay of the long-range interaction. At every d, we sample at different interlayer stacking registries by considering all possible in-plane layer displacements on a 12 × 12 grid in fractional coordinates. Our training and validation set is formed by the 6 × 6 subgrid of the 12 × 12 set of stacking configurations. The remaining stacking configurations form the test set, separate from the remaining 6 × 6 training and validation set. Overall, we generate 4320 calculations for the interlayer dataset. Q1D relaxation and band structure of HfS2/GaS DFT calculations for the quasi-one-dimensional (Q1D) HfS2/GaS bilayer were run using Quantum ESPRESSO (v.7.2)72. The relaxation was performed using a BFGS optimizer from an initial rigidly twisted Q1D structure with a plane-wave cutoff of 60 Ry, force convergence threshold of 10−3 Ry/Bohr, and a total energy convergence threshold of 10−4 Ry. An unshifted Monkhorst-Pack k-point sampling of 1 × 12 × 1 was used, accounting for the supercell in the first lattice direction (a1) and the periodic in-plane direction with the smallest lattice constant along the second direction (a2) of the Q1D structure. The SCF calculations for the initial, DFT-relaxed, and MLIP-relaxed structures were performed with a plane-wave cutoff of 60 Ry with an SCF convergence threshold of 10−10 Ry on an unshifted Monkhorst-Pack k-point sampling of 1 × 12 × 1. The non-self-consistent field (NSCF) calculations for each structure were calculated at 20 k-points along a path in the supercell Brillouin zone (BZ), which are equivalent to the high-symmetry k-point paths of the individual strained HfS2 and GaS monolayer BZ, up to multiples of the supercell reciprocal lattice vectors. We used the scalar-relativistic norm-conserving pseudopotentials from the SG15 database74. To analyze the electronic structure in terms of the constituent primitive cells, we employed the primitive band unfolding method developed by Popescu and Zunger76. Additionally, we resolve each Kohn-Sham (KS) state according to the two-layer projections. Specifically, we multiply each KS orbital in the moiré system, in real-space, by a Heaviside function that selects the region closer to either the HfS2 or GaS layers. Interatomic potential parameterization and usage Allegro IPs All Allegro MLIP models were trained using standard hyperparameters and termination conditions (see Data Availability Statement). For the interlayer models, we customize the Allegro model to include only interlayer interactions. Specifically, in the Allegro model44, the total energy of the system is expressed as where Zi is the atomic number of atom i, Eij is the pairwise energy and σZi,Zj is the per-species-pair scaling factor. We decompose the energy into its intralayer (Eintra) and interlayer (Einter) contributions, By using a distinct Z for atoms in each layer, we can set σZi,Zj to 0 for the intralayer pairs. This modification procedure could also be applied to other areas, such as defects, interfaces, or molecules on surfaces. Classical interatomic potentials All classical interatomic potentials were calculated using LAMMPS code77 with the pair styles available as sw/mod and kolmogorov/crespi/z, and parameters from Table 1 of Naik et al.19. Structure relaxation was performed using the FIRE optimization up to a force tolerance of 10−4 eV/Å. Corruption analyses were performed by introducing controlled random perturbations (corruption) into a single accurate split MLIP model trained on classical interatomic potentials. Gaussian random noise was added to the weights of the neural networks constituting the intralayer or interlayer MLIPs independently. The noise magnitude was scaled relative to the standard deviation of the original model parameters within each neural network layer, providing a controlled corruption factor. For each corruption level, we averaged results across ten different random seed initializations to increase statistical significance. Supplementary information Supplementary Information Transparent Peer Review file Author contributions F.H.J. conceived the approach, led the project, and acquired funding for the research. J.D.G. and F.H.J. designed the structural similarity metric. J.D.G. performed the model and surrogate system validation. A.R. and F.H.J. developed the split model approach. A.R. wrote the training and dataset generation code and trained the models. J.D.G., A.R., and F.H.J. wrote the manuscript. C.H.S. and E.H. helped refine the dataset generation and model training processes. J.D.G. and T.L. refined the surrogate system relaxation process. A.R. and B.B. extended the split MLIP approach to the MACE model. Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by the Office of Naval Research (ONR) through the Multi-University Research Initiative (MURI) on Twist-Optics, Grant # N00014-23-1-2567 (validation metric and surrogate system development), and by the National Science Foundation (NSF) CAREER award through Grant No. DMR-2238328 (split MLIP and electronic structure). This research used resources of the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility located at Lawrence Berkeley National Laboratory, operated under Contract No. DE-AC02-05CH11231 using NERSC award BES-ERCAP ERCAP0024305 to generate the DFT datasets for MoS2/WSe2 and HfS2/GaS. The 1D DFT relaxations and band-structure calculations for quasi-1D HfS2/GaS used computational resources from the Texas Advanced Computing Center (TACC) at The University of Texas at Austin, funded by the National Science Foundation (NSF) award 1818253, through allocation DMR21077. The geometry relaxation calculations of bilayer MoS2/WSe2 used resources of the Argonne Leadership Computing Facility (ALCF), a U.S. Department of Energy (DOE) Office of Science user facility at Argonne National Laboratory, based on research supported by the U.S. DOE Office of Science-Advanced Scientific Computing Research Program, under Contract No. DE-AC02-06CH11357. J.D.G. acknowledges support from the Natural Science and Engineering Research Council (NSERC) of Canada through the Post-Graduate Scholarship PGS D-568202-2022. Data availability The data generated in this study and scripts to reproduce the figures in this paper can be found at https://github.com/jornada-group/mlips_split_data/. We also share the associated raw log information for our MLIP training in the wandb format78, available at https://wandb.ai/twist-anything-ml/MLIPS_Paper?nw=mentxo4d3si. Code availability The Allegro code used to train the machine-learned interatomic potentials in this work is publicly available at https://github.com/mir-group/allegro and archived on Zenodo under https://doi.org/10.5281/zenodo.18856902. The Allegro methods are described in Ref. 46. The scripts used to generate the data and reproduce the figures are available at https://github.com/jornada-group/mlips_split_data/ and archived on Zenodo under https://doi.org/10.5281/zenodo.20128126. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Johnathan D. Georgaras, Akash Ramdas. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-74482-2. References Associated Data Data Availability Statement The data generated in this study and scripts to reproduce the figures in this paper can be found at https://github.com/jornada-group/mlips_split_data/. We also share the associated raw log information for our MLIP training in the wandb format78, available at https://wandb.ai/twist-anything-ml/MLIPS_Paper?nw=mentxo4d3si. The Allegro code used to train the machine-learned interatomic potentials in this work is publicly available at https://github.com/mir-group/allegro and archived on Zenodo under https://doi.org/10.5281/zenodo.18856902. The Allegro methods are described in Ref. 46. The scripts used to generate the data and reproduce the figures are available at https://github.com/jornada-group/mlips_split_data/ and archived on Zenodo under https://doi.org/10.5281/zenodo.20128126.