WorksComputer Science
MCSA: Multi-Cutoff Spectral Aggregation in SO(3) Equivariant GNNs for Interatomic Potentials
Abstract
Machine Learning Interatomic Potentials (MLIPs) based on Equivariant Graph Neural Networks (EGNNs) have achieved remarkable progress in atomistic modeling. However, accurately representing atomic interactions across different spatial scales remains challenging. In this work, we investigate the radial spectral properties of SO(3)-equivariant message passing and establish that the cutoff radius controls the characteristic spatial-frequency scale of radial interactions. Based on this analysis, we propose Multi-Cutoff Spectral Aggregation (MCSA), a plug-and-play module that aggregates radial kernels with different cutoff radii to enrich multi-scale geometric representations. To achieve efficient implementation, we exploit the compact support of radial kernels to fuse cutoff-dependent radial weights before angular tensor-product operations, avoiding redundant angular computations while preserving rotational equivariance. Extensive experiments across representative EGNN backbones demonstrate consistent improvements in energy, force, and stress prediction, as well as enhanced accuracy in higher-order properties, including phonon dispersion and lattice thermal conductivity. Further evaluations on layered materials with distinct interaction scales demonstrate the effectiveness of physically motivated cutoff selection. Code is available at: https://github.com/JF-Tan/MCSA.
Introduction
The quantitative simulation of atomic systems is fundamental to advancements in materials science and drug discovery[1, 2, 3, 4], yet it has long been constrained by a dichotomy between the accuracy of high-complexity ab initio methods[5, 6, 7] and the efficiency of limited classical potentials[8, 9, 10, 11, 12]. Recently, Machine Learning Interatomic Potentials (MLIPs) have emerged as a pivotal solution[13, 14, 15, 16], effectively bridging this gap by approximating the potential energy surface with neural networks. In particular, Equivariant Graph Neural Networks (EGNNs) [13, 16, 17, 18] have driven a major paradigm shift in atomistic modeling. By preserving physical and geometric symmetries, EGNNs achieve remarkable data efficiency and force prediction fidelity, enabling reliable molecular dynamics simulations across complex, heterogeneous systems.
Despite these remarkable advancements, the fine-grained representation of structural geometry remains an ongoing challenge in the development of robust MLIPs. Significant progress has been made in the angular domain, where the introduction of higher-order spherical harmonics and Clebsch-Gordan tensor products has yielded highly refined geometric representations [18, 15]. Concurrently, in the radial domain, the adoption of radial basis functions (e.g., Gaussian or Bessel functions) has further improved spatial resolution [19, 20, 21, 22]. However, a fundamental limitation in current EGNN architectures is frequently overlooked: the reliance on a single cutoff radius for all message-passing interactions.
In this work, we analyze the radial spectral response of -equivariant message-passing networks and show that the cutoff radius governs the characteristic spatial-frequency scale of radial interactions. Under a scale-consistent radial parameterization, a single cutoff provides a single characteristic spectral scale, potentially limiting the representation of fine-grained local geometries alongside broader atomic interactions. To address this limitation, we propose Multi-Cutoff Spectral Aggregation (MCSA), a plug-and-play module that combines radial kernels with different cutoff radii. By superposing their differently scaled frequency responses, MCSA enriches the radial representation across multiple spatial scales while preserving the underlying equivariant structure.
Furthermore, naively implementing multiple cutoffs may incur redundant evaluations of computationally expensive angular tensor products. We address this challenge through an efficient radial-fusion formulation that exploits the compact support of cutoff radial kernels to unify neighborhood aggregation. Specifically, the cutoff-dependent radial weights are combined before the angular tensor-product operation, allowing each edge and coupling path to share a single angular computation across all cutoffs. This formulation avoids redundant angular tensor products while retaining the original multi-cutoff aggregation, with additional computational cost arising primarily from radial-weight evaluation and fusion.
To comprehensively evaluate MCSA, we integrate it into multiple representative -equivariant GNN backbones. Experiments demonstrate consistent improvements in energy, force, and stress prediction, together with enhanced accuracy in higher-order physical properties, including phonon dispersion and lattice thermal conductivity. We further investigate layered materials, where short-range in-plane covalent bonds coexist with longer-range interlayer interactions, providing a targeted evaluation of MCSA in structurally heterogeneous environments. Finally, we evaluate PES smoothness and long-term molecular dynamics stability, demonstrating that multi-cutoff aggregation maintains smooth energy and force profiles without introducing artificial cutoff-boundary artifacts.
Preliminary: SO(3) Equivariant Message Passing
Modern Machine Learning Interatomic Potentials (MLIPs) typically employ equivariant message-passing mechanisms [13, 14, 15], where node features are composed of irreducible representations (Irreps) of the group, allowing geometric information to be propagated while preserving rotational equivariance. Consider a system of atoms with coordinates . Let , , denote the relative position and corresponding unit direction, respectively. Define as the neighborhood of atom within a cutoff radius . A general equivariant message-passing layer can be expressed as
where denotes the degree- feature of atom , and collects the spherical harmonics of degree . We adopt real orthonormal spherical harmonics throughout the derivation. The radial function determines the distance-dependent interaction strength for each angular coupling pathway and is parametrized with respect to the cutoff radius . It is typically the product of a radial basis function and a smooth truncation function. The operator denotes the Clebsch–Gordan (CG) tensor product, which couples the irreducible representations of degrees and and projects their tensor product onto the output irrep , subject to the angular momentum selection rules . The radial envelope is designed to ensure that the interaction smoothly vanishes at the cutoff boundary, with sufficient differentiability to support the evaluation of energy and force derivatives [19]. Expanding the CG projection in components gives
where denotes a CG coefficient. (2) separates the angular and radial contributions to message passing. The angular dependence is carried by spherical harmonics, whereas the distance-dependent interaction strength is controlled by the radial weight . This angular-radial separation provides a natural basis for analyzing the spectral properties of the radial kernel, particularly how the cutoff radius shapes its represented spatial frequency components.
Radial Spectral Response in SO(3)-Equivariant Message Passing
To characterize the spatial-frequency response of the radial kernel, we first reformulate the discrete message aggregation as a continuous spatial convolution. Specifically, we define the global atomic feature distribution as
where denotes the Dirac delta distribution. We further define the spatial aggregation kernel as
with compact support within the cutoff radius . The corresponding continuous feature aggregation is then expressed as
where denotes an arbitrary spatial position, and the original node-wise aggregation is recovered by evaluating the continuous feature field at . This formulation separates the atomic feature distribution from the spatial aggregation kernel, providing a convenient basis for analyzing the spectral response of the message-passing operation.
Define the spatial kernel . Under the Fourier transform convention , the plane-wave expansion [23] gives
where and
Here, is the spherical Bessel function of the first kind. Since the aggregation uses , the corresponding Fourier multiplier is evaluated at rather than . This sign reversal introduces only an angular phase change and does not affect the radial response .
We now assume a scale-consistent radial parameterization
where is a cutoff-dependent amplitude, and is a dimensionless radial profile defined on the normalized interval . In particular, the dependence on ensures that the normalized radial profile remains unchanged as the cutoff radius varies, while its spatial support scales with . This parameterization therefore isolates the effect of the cutoff radius on the spatial scale of the radial kernel. Substituting into (7) gives
with the dimensionless response
Thus, for a fixed dimensionless radial profile, changing dilates the frequency axis while scaling the response amplitude by . In particular, a characteristic feature at dimensionless frequency appears at , indicating that smaller cutoffs shift the characteristic frequency scale toward higher frequencies. At low frequencies, the expansion yields as . Nevertheless, oscillations and sidelobes may prevent the response from being a strictly monotonic low-pass filter.
Multi-Cutoff Spectral Aggregation
Radial Spectral Response in MCSA
Motivated by the cutoff-dependent frequency scaling derived above, we introduce Multi-Cutoff Spectral Aggregation (MCSA) to aggregate radial interactions across multiple spatial scales. Given cutoff radii , let the radial kernel for the -th cutoff be
where denotes the radial basis expansion or learned radial filter associated with the -th cutoff, and is a smooth cutoff envelope that ensures the radial kernel vanishes smoothly at . Thus, each kernel defines a radial interaction with its own spatial support and frequency scale. The multi-cutoff message-passing operation is
where denotes the aggregation weight of the -th cutoff, and is the corresponding neighborhood. By the linearity of the Fourier transform, the aggregated radial spectral response satisfies
Since all branches share the same spherical harmonic basis and Clebsch–Gordan coupling rules, the resulting Fourier-domain operator retains the angular structure of the single-cutoff formulation, with its radial response replaced by . To further characterize the resulting spectral response, we consider a scale-consistent parameterization of each radial branch:
where is a cutoff-dependent amplitude and is the dimensionless radial profile of the -th branch, defined on . Substituting Eq. (14) into the Fourier–Bessel transform and applying the change of variables yields
where
denotes the dimensionless radial frequency response of the -th branch. Here, is the spherical Bessel function of the first kind.
Combining Eqs. (13) and (15), we obtain
Algorithm 1 SO(3)-Equivariant Message Passing
Initialize atomic coordinates $\mathit{coords}$
Initialize atomic feature $\mathbf{x}$
Initialize Cutoff Radius $R_c$
Initialize output features $\mathbf{x}' \leftarrow 0$
$E \leftarrow \operatorname{radius\_graph}(\mathit{coords}, R_c)$
for each $(i,j) \in E$ do
$r_{ij} \leftarrow \operatorname{norm}(\mathit{coords}_j-\mathit{coords}_i)$
$Y_{ij} \leftarrow \operatorname{sph\_harm}(\mathit{coords}_j-\mathit{coords}_i)$
$\Phi_{ij} \leftarrow \operatorname{radial\_weight}(r_{ij},R_c)$
$m_{ij} \leftarrow \operatorname{weighted\_TP}(\mathbf{x}_j,Y_{ij},\Phi_{ij})$
$\mathbf{x}'_i \leftarrow \mathbf{x}'_i+m_{ij}$
end for
Algorithm 2 SO(3)-Equivariant Message Passing with MCSA
$E_{\max} \leftarrow \operatorname{radius\_graph}(\mathit{coords},\max(R_c^n))$
for each $(i,j) \in E_{\max}$ do
$r_{ij} \leftarrow \operatorname{norm}(\mathit{coords}_j-\mathit{coords}_i)$
$Y_{ij} \leftarrow \operatorname{sph\_harm}(\mathit{coords}_j-\mathit{coords}_i)$
$\Phi_{ij}^{\mathrm{MC}} \leftarrow 0$
for each $n=1:N$ do
$\Phi_{ij}^{\mathrm{MC}} \leftarrow \Phi_{ij}^{\mathrm{MC}}+\gamma_n\cdot\operatorname{radial\_weight}(r_{ij},R_c^n)$
end for
$m_{ij} \leftarrow \operatorname{weighted\_TP}(\mathbf{x}_j,Y_{ij},\Phi_{ij}^{\mathrm{MC}})$
$\mathbf{x}'_i \leftarrow \mathbf{x}'_i+m_{ij}$
end for
In general, independently parameterized radial branches may have different dimensionless profiles, resulting in distinct responses . When all branches share the same dimensionless radial profile, i.e., for all , their dimensionless responses reduce to a common function , and Eq. (17) simplifies to
Thus, under the scale-consistent parameterization, MCSA constructs a weighted superposition of radially scaled frequency responses, allowing different spatial scales to contribute to a unified equivariant interaction. For interatomic potentials, short-range interactions primarily characterize local bonding and repulsion, whereas larger neighborhoods provide information about coordination and the surrounding atomic environment. By aggregating interactions across these spatial scales, MCSA enriches the radial representation of the potential energy surface and provides a more flexible basis for modeling atomic interactions.
Computational Cost Reduction via Radial Fusion
A direct implementation may independently evaluate the angular tensor products for every cutoff, introducing redundant computation. The aggregated radial kernel is defined as
MCSA avoids this redundancy by exploiting the compact support of the radial kernels and the linearity of message aggregation. Let . Since and outside its cutoff, we have
This identity allows all cutoff-dependent radial contributions to be accumulated before evaluating the CG tensor product. Algorithm 2 summarizes the resulting computation.
As shown in Algorithm 2, each cutoff contributes only to the scalar radial weight, while the CG tensor product is evaluated once per edge and coupling path, independently of . For a direct implementation that independently evaluates the CG tensor products for all cutoff branches, the total number of angular contractions scales as . By exploiting the shared angular basis and the linearity of tensor-product operations, radial fusion reduces this to , requiring only one angular contraction per edge in the maximum neighborhood. The cutoff-dependent radial filters must still be evaluated within their respective interaction ranges, so their computational cost remains dependent on . Consequently, MCSA reduces redundant angular computations, while the overall runtime savings depend on the relative costs of radial evaluations and angular contractions, as well as the neighborhood sizes.
Plug-and-Play Integration
MCSA modifies only the invariant scalar radial weights, leaving the spherical harmonics and CG coupling structure unchanged. It can therefore be incorporated into existing SO(3)-equivariant message-passing layers whose edge interactions admit a separable radial-weighted form and linear aggregation, including compatible radial interaction blocks in architectures such as NequIP, MACE, EquiformerV2, and eSCN. For attention-based layers, exact fusion additionally requires preserving any cutoff-dependent masks and nonlinear attention normalization.
The mechanism requires no change to the angular representation or its coupling rules. Moreover, when and , MCSA recovers the corresponding single-cutoff radial kernel and hence the original message-passing formulation. Since a finite weighted sum of sufficiently smooth, compactly supported radial kernels retains the required cutoff smoothness, MCSA also preserves the differentiability conditions needed for energy and force prediction.
Experiments
We evaluate MCSA on multiple benchmarks to demonstrate its effectiveness in improving geometric representation fidelity, higher-order derivative accuracy, and long-term simulation stability. Unless otherwise specified, the EGNN+MCSA configuration uses with cutoff radii Å, incorporating the Radial Fusion optimization.
Generality Across Backbones
To verify the universality of the Multi-Cutoff Spectral Aggregation (MCSA) module, we integrate it into representative SO(3)-equivariant GNNs: Nequix [24], Eqnorm [25], and EquiformerV3 [26]. We evaluate three representative SO(3)-equivariant backbones on the MPtrj dataset [27, 28] using matched training settings for their baseline and MCSA variants. As shown in Table 1, we evaluate performance on both the in-distribution (ID) MPtrj test split and the out-of-distribution (OOD) Alexandria dataset[29, 30].
| Method | MPtrj (ID) | MPtrj (ID) | MPtrj (ID) | Alexandria (OOD) | Alexandria (OOD) | Alexandria (OOD) |
| Energy | Force | Stress | Energy | Force | Stress | |
| Nequix | 14.2 | 40.4 | 0.273 | 41.8 | 75.5 | 0.366 |
| Nequix+MCSA | 13.4 | 38.7 | 0.267 | 39.9 | 74.3 | 0.354 |
| Eqnorm | 11.5 | 34.2 | 0.211 | 36.7 | 68.3 | 0.331 |
| Eqnorm+MCSA | 10.8 | 32.5 | 0.203 | 35.6 | 67.1 | 0.323 |
| EquiformerV3 | 6.3 | 18.1 | 0.151 | 22.1 | 43.5 | 0.217 |
| EquiformerV3+MCSA | 6.2 | 17.2 | 0.147 | 20.3 | 41.1 | 0.206 |
Table 1. Performance improvements of EGNNs with MCSA module on two datasets. Mean Absolute Error (MAE): energy (meV/atom), force (meV/Å), stress (meV/atom/Å).
As shown in Table 1, MCSA consistently improves energy, force, and stress prediction across all three architectures on both ID and OOD datasets. Notably, the improvements extend beyond the training distribution, indicating better generalization to unseen atomic structures. For EquiformerV3, MCSA reduces OOD energy, force, and stress MAEs by 8.1%, 5.5%, and 5.1%, respectively. Similar improvements are observed in Nequix and Eqnorm, demonstrating that the benefits of multi-cutoff radial aggregation are not restricted to a specific backbone. These results highlight the effectiveness and transferability of MCSA as a lightweight enhancement for improving the radial representation of equivariant interatomic potentials.
Performance on Higher-Order Derivatives
While fitting energy and forces (first-order derivatives) is the primary objective of MLIPs, the true reliability of a learned potential energy surface is governed by its higher-order derivatives.
Phonon Dispersion Analysis
We evaluate the fidelity of the learned potential energy surfaces (PESs) through phonon dispersion relations and phonon density of states (PDOS), which provide a stringent assessment of their local curvature. Specifically, squared phonon frequencies correspond to the eigenvalues of the mass-weighted dynamical matrix, which is constructed from the interatomic force constants [31]. It makes phonon properties particularly sensitive to errors in the second-order derivatives of the PES.
Figure 1 compares the predicted phonon spectra of three representative materials: BAs, BN, and CaSe. The baseline Nequix exhibits noticeable deviations from DFT references, particularly in the high-frequency optical branches. This discrepancy is especially pronounced in strongly covalent BN, whose phonon modes extend to approximately 38 THz. The baseline underestimates the highest optical frequencies by nearly 10 THz and produces a substantially distorted PDOS, indicating insufficient accuracy in capturing the local curvature associated with stiff atomic bonds.
Figure 1. Comparison of phonon spectra and phonon density of states (PDOS) inferred by different MLIPs against DFT for three materials. (a) BAs. (b) BN. (c) CaSe.
In contrast, Nequix+MCSA achieves closer agreement with DFT across all three materials, improving the prediction of both acoustic and optical branches. In particular, the upper optical branches are shifted toward the reference frequencies, leading to better alignment with the corresponding PDOS peaks. By combining radial interactions across multiple cutoff scales, MCSA enriches the representation of local atomic geometries while retaining information from broader atomic environments. These improvements suggest that multi-cutoff radial aggregation provides a more expressive representation for capturing the fine-scale structural variations underlying higher-order physical properties.
Lattice Thermal Conductivity
Accurate lattice thermal conductivity prediction requires reliable second-order harmonic and third-order anharmonic interatomic force constants, while fourth-order force constants are additionally required when four-phonon scattering is considered.[32, 33, 34]
As shown in Table 2, we evaluate the prediction performance on the PhononDB-103 thermal conductivity benchmark [35]. We report the Symmetric Relative Mean Error () [35] to account for the logarithmic distribution of across different materials. To directly validate the efficiency of MCSA, we also record the average computational time required to process each material.
| Model | Avg. Time (min/mat) | (↓) |
| Nequix | 1.107 | 0.487 |
| Nequix+MCSA | 1.136 | 0.420 |
Table 2. Lattice thermal conductivity () prediction performance and computational efficiency on the PhononDB-103 dataset. A lower indicates better agreement with ab initio references. The inference time is measured in minutes per material.
The results demonstrate that the integration of the MCSA module reduces the from 0.487 to 0.420, representing a significant 13.7% relative reduction in error. Remarkably, this enhancement in capturing high-order anharmonicity comes at an almost negligible computational cost. The average calculation time per material increases by merely 1.7 seconds (from 1.107 to 1.136 minutes), an overhead of less than 3%. These results demonstrate that MCSA improves lattice thermal conductivity prediction with only a modest increase in computational time, consistent with the efficiency of the proposed radial fusion formulation.
PES Smoothness and Molecular Dynamics Stability
Beyond static prediction accuracy, a reliable MLIP should maintain a smooth potential energy surface (PES) and stable long-term molecular dynamics (MD) trajectories. We evaluate these properties through a Si-Si dissociation process and long-time MD simulations.
Figure 2(a) and (Figure 2b) show that Nequix+MCSA maintains smooth energy and force profiles across the intermediate cutoff boundaries at 2.0 and 4.0 Å, without visible discontinuities or abrupt transitions. This behavior follows from the compactly supported smooth envelopes: their weighted sum preserves the required cutoff smoothness of the individual radial kernels, avoiding artificial artifacts introduced by multi-cutoff aggregation.
Figure 2. PES smoothness and MD stability. (a) Predicted potential energy and (b) interatomic force during Si-Si dissociation using Nequix+MCSA (, Å). (c) Energy evolution during a 100 ps NVE simulation on TM23 at 300 K using Nequix+MCSA.
We further evaluate dynamical stability through 100 ps NVE simulations on the TM23 dataset [36] at 300 K with a 0.5 fs time step. As shown in Figure 2(c), Nequix+MCSA completes the entire trajectory without structural collapse, demonstrating stable MD behavior under the tested conditions.
Physically Motivated Multi-Cutoff Design for Layered Materials
To evaluate MCSA on layered materials with distinct interaction length scales, we consider three representative Mo–S structures from JARVIS-DFT [37]: JVASP-100856, JVASP-28733, and JVASP-34141. These structures are treated as held-out test cases. Before model training, all ALIGNN-FF-DB configurations corresponding to these three JARVIS entries were excluded from the training and validation splits. Structural analysis reveals two characteristic interaction scales: in-plane Mo–S covalent bonds with distances around 2.4 to 3.0 Å, and interlayer S–S separations ranging from 3.61 to 6.51 Å. Based on these structural characteristics, we adopt a physically motivated dual-cutoff configuration of Å. The shorter cutoff emphasizes local in-plane environments, while the larger cutoff expands the interaction range to incorporate neighboring layers.
We train two models on the ALIGNN-FF-DB dataset [38] using identical architectures and training configurations, differing only in their cutoff settings: a single-cutoff baseline with Å, and MCSA with Å. Figure 3 reports their energy and force prediction errors on the three representative structures. MCSA consistently reduces both energy and force MAEs across all three systems, with particularly notable improvements in force prediction. These results demonstrate the potential of physically motivated multi-cutoff aggregation for modeling materials with distinct short-range and interlayer interaction scales.

Figure 3. (a) Mo–S crystal structure illustrating the multi-cutoff design for intra- and interlayer interactions. (b) Energy MAE (meV/atom) and (c) Force MAE (meV/Å) in log scale on three representative layered Mo–S structures, comparing single-cutoff and MCSA on Nequix.
Ablation on MCSA Configurations
We conduct comprehensive ablation studies to investigate the effects of three key MCSA configurations: the number and placement of cutoff radii (), the radial basis function (RBF) allocation, and the polynomial envelope degree (). We additionally report the average inference latency per material, measured in milliseconds on a single NVIDIA A800 GPU, to assess the accuracy-efficiency trade-off. All models are trained on MPtrj and evaluated on its test split. The results are summarized in Table 3.
| Variant | Cutoff Radii (Å) | RBF Size | Poly Degree | Inference Time | Params | Energy | Force | Stress | |
| Baseline | 1 | [6.0] | [8] | [6] | 47 ms | 0.7M | 14.15 | 40.41 | 0.273 |
| Part I: Number and Placement of Cutoff Radii | |||||||||
| Two-scale | 2 | [2.0, 6.0] | [8, 8] | [6, 6] | 50 ms | 0.9M | 13.70 | 39.25 | 0.268 |
| Three-scale (Ours) | 3 | [2.0, 4.0, 6.0] | [8, 8, 8] | [6, 6, 6] | 51 ms | 1.1M | 13.44 | 38.54 | 0.267 |
| Four-scale (A) | 4 | [3.0, 4.0, 5.0, 6.0] | [8 ×4] | [6 ×4] | 54 ms | 1.3M | 13.57 | 38.87 | 0.269 |
| Four-scale (B) | 4 | [1.0, 2.0, 3.0, 6.0] | [8 ×4] | [6 ×4] | 54 ms | 1.3M | 13.79 | 39.34 | 0.268 |
| Six-scale | 6 | [1.0, 2.0, 3.0, 4.0, 5.0, 6.0] | [8 ×6] | [6 ×6] | 56 ms | 1.6M | 14.26 | 40.18 | 0.276 |
| Part II: Radial Basis and Envelope Configurations | |||||||||
| Outer-Heavy RBF | 2 | [3.0, 6.0] | [8, 16] | [6, 6] | 52 ms | 0.9M | 13.76 | 39.28 | 0.267 |
| Inner-Heavy RBF | 2 | [3.0, 6.0] | [16, 8] | [6, 6] | 52 ms | 0.9M | 13.68 | 39.15 | 0.272 |
| Envelope Mix (A) | 2 | [3.0, 6.0] | [8, 8] | [6, 8] | 50 ms | 0.9M | 13.83 | 39.71 | 0.270 |
| Envelope Mix (B) | 2 | [3.0, 6.0] | [8, 8] | [8, 6] | 49 ms | 0.9M | 13.88 | 39.74 | 0.269 |
Table 3. Comprehensive ablation study on radial configurations using the Nequix baseline. MAE is reported in meV/atom (Energy), meV/Å (Force), and meV/atom/Å (Stress). The inference time demonstrates the marginal overhead of the MCSA mechanism.
Part I of Table 3 shows that three cutoffs with Å achieve the best overall performance, reducing Energy, Force, and Stress MAEs to 13.44, 38.54, and 0.267, respectively. Increasing beyond three brings no further gains, indicating that selecting complementary interaction scales is more important than simply adding cutoffs.
MCSA introduces only modest inference overhead, increasing latency from 47 ms to 51 ms for (8.5%). This demonstrates the efficiency of radial fusion, which avoids redundant angular tensor-product computations across cutoff branches.
Part II shows that the Inner-Heavy RBF allocation ([16, 8]) improves Energy and Force accuracy over the Outer-Heavy configuration ([8, 16]), while slightly increasing Stress error. The envelope configurations yield comparable results, suggesting that cutoff selection has a greater impact than the tested envelope-degree variations.
Related Works
Machine Learning Interatomic Potentials (MLIPs). The evolution of MLIPs has transitioned from hand-crafted descriptors like Behler-Parrinello NNs [20], GAP [39], and ACE [40] to end-to-end Graph Neural Networks (GNNs). Early GNNs, such as SchNet [21] and DimeNet [22], incorporated geometric information through continuous-filter convolutions and directional message passing, primarily using invariant scalar features. The emergence of -equivariant architectures [41] marked a shift toward strictly enforcing symmetry on tensor features. NequIP [15] established the high data efficiency of equivariant spherical harmonics, leading to diverse architectures: Allegro [16] prioritized strict locality via many-body tensor products, while MACE [13] unified ACE and GNNs through body-ordered symmetric contractions. Recent state-of-the-art models like EquiformerV3 [26], eSCN [42] and eSEN [19] further optimize these concepts via attention mechanisms and efficient tensor operations.
Radial Representations, Cutoff Design, and Multi-Scale Interactions. Radial basis representations are fundamental to MLIPs, ranging from atom-centered symmetry functions [20] and SOAP [39, 43] to MTP [44] and ACE [40]. Subsequent studies have explored more efficient radial basis parameterizations and expansions [45, 46]. In neural architectures, SchNet [21] and DimeNet [22] employ continuous radial filters, while equivariant models such as NequIP [15], MACE [13], and Allegro [16] integrate learnable radial functions with smooth cutoff envelopes. Recent efforts further investigate equivariant architecture design and interaction ranges through many-body representations, nonlocal interactions, and adaptive cutoff strategies [47, 48, 49, 50, 51, 52]. Unlike approaches that primarily modify radial parameterization or neighborhood selection, MCSA focuses on the spectral properties of radial kernels, combining differently scaled cutoff responses to enrich multi-scale representations while retaining the original equivariant tensor-product structure.
Conclusion
In this work, we proposed Multi-Cutoff Spectral Aggregation (MCSA), a plug-and-play mechanism that enriches the radial representations of -equivariant GNNs through the superposition of differently scaled cutoff responses. By fusing cutoff-dependent radial weights before angular tensor-product operations, MCSA enables efficient multi-scale interactions while preserving rotational equivariance. Extensive experiments demonstrate improvements in energy, force, and stress prediction, as well as higher-order properties including phonon dispersion and lattice thermal conductivity. Further evaluations on layered materials and long-term molecular dynamics demonstrate the effectiveness and stability of the proposed approach.
Limitations
Despite its efficacy, MCSA currently relies on predefined radial scales, which may not be universally optimal for all chemical environments. Exploring adaptive cutoff scheduling remains a promising direction. Additionally, while our associative optimization ensures efficiency by fusing scales linearly in the scalar domain, future work could investigate lightweight cross-scale interaction modules to capture potential non-linear couplings between different spatial resolutions.
Implementation Details
We implemented the Multi-Cutoff Spectral Aggregation (MCSA) strategy on three -equivariant architectures: Nequix[24], EquiformerV3[26], and Eqnorm[25]. All models were trained using multi-cutoff radii Å.
Nequix-MCSA
We train the Nequix-MCSA, following the training protocol of Nequix [24], jointly optimizing the Mean Absolute Error (MAE) for energy, forces, and macroscopic stress. To efficiently accelerate convergence, we employ the Muon optimizer. Unlike the EquiformerV3 [26] setup, no auxiliary positional denoising tasks were utilized during the training of this backbone. The complete set of training hyper-parameters is detailed in Table 4. The model was trained across 120 NVIDIA A800 GPUs, and the entire training process takes approximately 1617 GPU-hours.
| Hyper-parameters | Value |
| Optimizer | Muon |
| Learning rate scheduling | Cosine decay + linear warmup |
| Warmup epochs | 0.1 |
| Maximum learning rate | |
| Batch size (per GPU) | 64 |
| Weight decay | |
| Energy coefficient | 20 |
| Force coefficient | 20 |
| Stress coefficient | 5 |
| Gradient clipping norm threshold | 100 |
| Model EMA decay | 0.999 |
| Cutoffs (Å) | [2, 4, 6] |
| Radial basis function | Bessel |
| Number of radial bases | [8, 8, 8] |
| Radial polynomial envelope degree | [6.0, 6.0, 6.0] |
| Radial MLP hidden dimension | 64 |
| Radial MLP layers | 2 |
| Maximum degree | 3 |
| Number of message passing layers | 4 |
| Hidden Irreps |
Table 4. Hyper-parameters for Nequix-MCSA.
EquiformerV3-MCSA
We follow the practice of direct pre-training and gradient fine-tuning [26]. Auxiliary task of denoising based on DeNS [53] was used in direct pre-training. For DeNS, after sampling which atoms to corrupt, we ensure that the fraction of corrupted atoms does not exceed 0.75, preventing denoising from degenerating into the case where noise is added for all atoms. The hyper-parameters are in Table 5. 120 A800 GPUs are used for training. Direct pre-training takes about 1680 GPU-hours, and gradient fine-tuning takes about 961 GPU-hours.
| Hyper-parameters | Direct pre-training | Gradient fine-tuning |
| Optimizer | AdamW | AdamW |
| Learning rate scheduling | Cosine decay + linear warmup | Cosine decay + linear warmup |
| Warmup epochs | 0.1 | 0.1 |
| Maximum learning rate | ||
| Batch size (per GPU) | 16 | 16 |
| Number of epochs | 70 | 10 |
| Weight decay | ||
| Dropout rate | 0.1 | 0.0 |
| Stochastic depth | 0.05 | 0.0 |
| Energy coefficient | 20 | 20 |
| Force coefficient | 20 | 20 |
| Stress coefficient | 5 | 5 |
| Gradient clipping norm threshold | 100 | 100 |
| Model EMA decay | 0.999 | 0.999 |
| Cutoffs (Å) | [2, 4, 6] | [2, 4, 6] |
| Maximum number of neighbors | 300 | 300 |
| Radial basis function | Gaussian | Gaussian |
| Number of radial bases | 10 | 10 |
| Dimension of hidden scalar features in radial functions | (0, 128) | (0, 128) |
| Maximum degree | 4 | 4 |
| Maximum order | 2 | 2 |
| Number of Transformer blocks | 7 | 7 |
| Embedding dimension | (4, 128) | (4, 128) |
| dimension after SwiGLU- | (4, 32) | (4, 32) |
| Number of attention heads | 8 | 8 |
| dimension | (0, 64) | (0, 64) |
| Value dimension | (4, 16) | (4, 16) |
| Hidden dimension in feedforward networks after SwiGLU- | (4, 512) | (4, 512) |
| Resolution of grid points in attention | (8, 14) | (8, 14) |
| Resolution of grid points in feedforward networks | (14, 14) | (14, 14) |
| Probability of optimizing DeNS | 0.5 | 0.0 |
| DeNS coefficient | 10 | 0.0 |
| Standard deviation of Gaussian noise | 0.025 | 0.0 |
| Corruption ratio | 0.5 | 0.0 |
Table 5. Hyper-parameters for EquiformerV3-MCSA.
Eqnorm-MCSA
Similar to the Nequix backbone, we train the Eqnorm-MCSA model end-to-end to jointly optimize the Mean Absolute Error (MAE) for energy, forces, and macroscopic stress. To efficiently process the stress tensor derivations, the model utilizes an edge-based gradient computation mode. We employ the Muon optimizer to accelerate convergence without relying on auxiliary denoising tasks. The comprehensive hyper-parameters are summarized in Table 6. The model was trained across 120 NVIDIA A800 GPUs, and the entire training process takes approximately about 2341 GPU-hours.
| Hyper-parameters | Value |
| Optimizer | Muon |
| Learning rate scheduling | Cosine decay + linear warmup |
| Warmup epochs | 0.1 |
| Maximum learning rate | |
| Batch size (per GPU) | 8 |
| Weight decay | |
| Energy coefficient | 20 |
| Force coefficient | 20 |
| Stress coefficient | 5 |
| Gradient clipping norm threshold | 100 |
| Model EMA decay | 0.999 |
| Cutoffs (Å) | [2, 4, 6] |
| Radial basis function | Bessel |
| Number of radial bases | [8, 8, 8] |
| Radial polynomial envelope degree | [6.0, 6.0, 6.0] |
| Maximum degree | 3 |
| Number of convolutional layers | 5 |
| Hidden Irreps | |
| Number of invariant layers | 2 |
| Number of invariant neurons | 128 |
| Stress gradient mode | Edge-based |
Table 6. Hyper-parameters for Eqnorm-MCSA.
All models adopt independent radial MLPs and envelope functions for each cutoff, with learnable scale-mixing coefficients . Implementation was built upon the official codebases with necessary extensions for multi-cutoff support.
NeurIPS Paper Checklist
The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit.
Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist:
You should answer [Yes], [No], or [N/A].
[N/A] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available.
Please provide a short (1–2 sentence) justification right after your answer (even for [N/A]).
The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will also be asked to include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper.
The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While [Yes] is generally preferable to [No], it is perfectly acceptable to answer [No] provided a proper justification is given (e.g., error bars are not reported because it would be too computationally expensive” or “we were unable to find the license for the dataset we used”). In general, answering [No] or [N/A] is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found.
IMPORTANT, please:
Delete this instruction block, but keep the section heading “NeurIPS Paper Checklist”,
Keep the checklist subsection headings, questions/answers and guidelines below.
Do not modify the questions and only use the provided macros for your answers.
Claims
Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?
Answer: [Yes]
Justification: We assure that the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope.
Guidelines:
The answer [N/A] means that the abstract and introduction do not include the claims made in the paper.
The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers.
The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.
It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
Limitations
Question: Does the paper discuss the limitations of the work performed by the authors?
Answer: [Yes] Justification: The limitations are in Appendix A.
Guidelines:
The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper.
The authors are encouraged to create a separate “Limitations” section in their paper.
The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.
The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.
The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.
The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.
If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.
While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
Theory assumptions and proofs
Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
Answer: [Yes]
Justification: Proofs and Derivations are in Appendix.
Guidelines:
The answer [N/A] means that the paper does not include theoretical results.
All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.
All assumptions should be clearly stated or referenced in the statement of any theorems.
The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.
Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.
Theorems and Lemmas that the proof relies upon should be properly referenced.
Experimental result reproducibility
Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
Answer: [Yes]
Justification: We provide details for reproduction in Section 5 and Appendix B.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.
If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.
Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.
While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
Open access to data and code
Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
Answer: [Yes]
Justification: The code is available at the anonymous repository: https://anonymous.4open.science/r/MCSSD-C854.
Guidelines:
The answer [N/A] means that paper does not include experiments requiring code.
Please see the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.
While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).
The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.
The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.
The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.
At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).
Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
Experimental setting/details
Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?
Answer: [Yes]
Justification: The details are provided in Section 5 and Appendix § 9.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.
The full details can be provided either with the code, in appendix, or as supplemental material.
Experiment statistical significance
Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
Answer: [No]
Justification: Our experiments are mostly computational expensive. Therefore, we run only one group of experiment for each setting, following previous works.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).
The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)
The assumptions made should be given (e.g., Normally distributed errors).
It should be clear whether the error bar is the standard deviation or the standard error of the mean.
It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified.
For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates).
If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
Experiments compute resources
Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
Answer: [Yes]
Justification: The details are provided in Section 5 and Appendix § 9.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.
The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.
The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper).
Code of ethics
Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
Answer: [Yes]
Justification: We respect the NeurIPS code of ethics.
Guidelines:
The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics.
If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics.
The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
Broader impacts
Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
Answer: [N/A]
Justification: This work is about physics/materials science, which does not involve societal impact.
Guidelines:
The answer [N/A] means that there is no societal impact of the work performed.
If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact.
Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.
The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.
If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
Safeguards
Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?
Answer: [N/A]
Justification: Our research is about physics/materials science that dose not have such risks.
Guidelines:
The answer [N/A] means that the paper poses no such risks.
Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.
Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.
We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
12. Licenses for existing assets
Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
Answer: [Yes]
Justification: The discussion is in Appendix B.
Guidelines:
The answer [N/A] means that the paper does not use existing assets.
The authors should cite the original paper that produced the code package or dataset.
The authors should state which version of the asset is used and, if possible, include a URL.
The name of the license (e.g., CC-BY 4.0) should be included for each asset.
For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.
If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets,
paperswithcode.com/datasetshas curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.
If this information is not available online, the authors are encouraged to reach out to the asset’s creators.
13. New assets
Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
Answer: [Yes]
Justification: The code is available at the anonymous repository: https://anonymous.4open.science/r/MCSD-C854 with a README as the documentation.
Guidelines:
The answer [N/A] means that the paper does not release new assets.
Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.
The paper should discuss whether and how consent was obtained from people whose asset is used.
At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
14. Crowdsourcing and research with human subjects
Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A]
Justification: We do not include these resources.
Guidelines:
The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.
Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.
According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
15. Institutional review board (IRB) approvals or equivalent for research with human subjects
Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
Answer: [N/A]
Justification: We do not include such risks.
Guidelines:
The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.
Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.
We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.
For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
16. Declaration of LLM usage
Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.
Answer: [N/A]
Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components.
Guidelines:
The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.
Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.
References
- [1]Christel AS Bergström and Per Larsson. Computational prediction of drug solubility in water-based systems: Qualitative and quantitative approaches used in the current drug discovery and development setting. The International Journal of Pharmaceutics, 540(1-2):185–193, 2018.DOI
- [2]David W Borhani and David E Shaw. The future of molecular dynamics simulations in drug discovery. Journal of Computer-Aided Molecular Design, 26(1):15–26, 2012.DOI
- [3]Martin Karplus and Gregory A Petsko. Molecular dynamics simulations in biology. Nature, 347(6294):631–639, 1990.pubmed.ncbi.nlm.nih.gov/2215695
- [4]Wei-Qiao Deng, Lei Sun, Jin-Dou Huang, Shuo Chai, Shu-Hao Wen, and Ke-Li Han. Quantitative prediction of charge mobilities of π-stacked systems by first-principles simulation. Nature Protocols, 10(4):632–642, 2015.DOI
- [5]W. Kohn and L. J. Sham. Self-consistent equations including exchange and correlation effects. Physical Review, 140:A1133–A1138, Nov 1965.DOI
- [6]Paul Geerlings, Frank De Proft, and Wilfried Langenaeker. Conceptual density functional theory. Chemical Reviews, 103(5):1793–1874, 2003.
- [7]David S Sholl and Janice A Steckel. Density functional theory: a practical introduction. John Wiley & Sons, 2022.
- [8]Murray S. Daw and M. I. Baskes. Embedded-atom method: Derivation and application to impurities, surfaces, and other defects in metals. Physical Review B, 29:6443–6453, Jun 1984.DOI
- [9]J E Lennard-Jones. Cohesion. Proceedings of the Physical Society, 43(5):461, sep 1931.
- [10]Martin H Müser, Sergey V Sukhomlinov, and Lars Pastewka. Interatomic potentials: Achievements and challenges. Advances in Physics: X, 8(1):2093129, 2023.DOI
- [11]H. Balamane, T. Halicioglu, and W. A. Tiller. Comparative study of silicon empirical interatomic potentials. Physical Review B, 46(4):2250, 1992.pubmed.ncbi.nlm.nih.gov/10003901
- [12]Tomoyuki Hata, Giacomo Giorgi, Koichi Yamashita, Claudia Caddeo, and Alessandro Mattoni. Development of a classical interatomic potential for MAPbBr3. The Journal of Physical Chemistry C, 121(7):3724–3733, 2017.
- [13]Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Alice H. Oh, Alekh Agrawal, Danielle Belgrave, and Kyunghyun Cho, editors, Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022.
- [14]Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. In Proceedings of the International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2306.12059
- [15]Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13(1):2453, 2022.DOI
- [16]Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14(1):579, 2023.DOI
- [17]Guillem Simeon and Gianni De Fabritiis. Tensornet: Cartesian tensor representations for efficient learning of molecular potentials. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022.arxiv.org/abs/2306.06482
- [18]Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks. arXiv preprint arXiv:2207.09453, 2022.arxiv.org/abs/2207.09453
- [19]Xiang Fu, Brandon M Wood, Luis Barroso-Luque, Daniel S. Levine, Meng Gao, Misko Dzamba, and C. Lawrence Zitnick. Learning smooth and expressive interatomic potentials for physical property prediction. In Proceedings of the 42nd International Conference on Machine Learning (ICML), pages 17875–17893, 2025.arxiv.org/abs/2502.12147
- [20]Jörg Behler and Michele Parrinello. Generalized neural-network representation of high-dimensional potential-energy surfaces. Physical Review Letters, 98:146401, Apr 2007.pubmed.ncbi.nlm.nih.gov/17501293
- [21]K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. Müller. Schnet – a deep learning architecture for molecules and materials. The Journal of Chemical Physics, 148(24):241722, 03 2018.DOI
- [22]Johannes Gasteiger, Janek Groß, and Stephan Günnemann. Directional message passing for molecular graphs. In Proceedings of the International Conference on Learning Representations (ICLR), 2020.arxiv.org/abs/2003.03123
- [23]John David Jackson. Classical Electrodynamics. Wiley, 3 edition, 1998.
- [24]Teddy Koker, Mit Kotak, and Tess Smidt. Training a foundation model for materials on a budget. In AI for Accelerated Materials Design - NeurIPS, 2025.arxiv.org/abs/2508.16067
- [25]Yuzhuo Chen, Lyuwen Fu, Shuxiang Yang, and Lipeng Chen. Eqnorn: E(3)-equivariant graph neural network for material energy, force and stress predictions. https://github.com/yzchen08/eqnorm, 2025.
- [26]Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. EquiformerV3: Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers. arXiv preprint arXiv:2604.09130, 2026.arxiv.org/abs/2604.09130
- [27]Anubhav Jain, Shyue Ping Ong, Geoffrey Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, et al. Commentary: The materials project: A materials genome approach to accelerating materials innovation. APL Materials, 1(1):011002, 2013.DOI
- [28]Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J Bartel, and Gerbrand Ceder. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modeling. Nature Machine Intelligence, 5(9):1031–1041, 2023.DOI
- [29]Jonathan Schmidt, Noah Hoffmann, Hai-Chen Wang, Pedro Borlido, Pedro JMA Carriço, Tiago FT Cerqueira, Silvana Botti, and Miguel AL Marques. Machine-learning-assisted determination of the global zero-temperature phase diagram of materials. Advanced Materials, 35(22):2210788, 2023.DOI
- [30]Jonathan Schmidt, Love Pettersson, Claudio Verdozzi, Silvana Botti, and Miguel AL Marques. Crystal graph attention networks for the prediction of stable materials. Science Advances, 7(49):eabi7948, 2021.DOI
- [31]Atsushi Togo and Isao Tanaka. First principles phonon calculations in materials science. Scr. Mater., 108:1–5, 2015.arxiv.org/abs/1506.08498
- [32]Olle Hellman and David A. Broido. Phonon thermal transport in Bi₂Te₃ from first principles. Physical Review B, 90(13):134309, 2014.DOI
- [33]Lucas Lindsay, Ankita Katre, Andrea Cepellotti, and Natalio Mingo. Perspective on ab initio phonon thermal transport. Journal of Applied Physics, 126(5), 2019.DOI
- [34]Wu Li, Jesús Carrete, Nebil A Katcho, and Natalio Mingo. Shengbte: A solver of the boltzmann transport equation for phonons. Computer Physics Communications, 185(6):1747–1758, 2014.DOI
- [35]Balázs Póta, Paramvir Ahlawat, Gábor Csányi, and Michele Simoncelli. Thermal conductivity predictions with foundation atomistic models. Nature Communications, 17(1):10566, 2026.DOI
- [36]Cameron J Owen, Steven B Torrisi, Yu Xie, Simon Batzner, Kyle Bystrom, Jennifer Coulter, Albert Musialian, Lixin Sun, and Boris Kozinsky. Complexity of many-body interactions in transition metals via machine-learned force fields from the tm23 data set. npj Computational Materials, 10(1):92, 2024.arxiv.org/abs/2302.12993
- [37]Kamal Choudhary, Kevin F. Garrity, Andrew C. E. Reid, et al. The joint automated repository for various integrated simulations (jarvis) for data-driven materials design. npj Computational Materials, 6:173, 2020.DOI
- [38]Kamal Choudhary, Brian DeCost, Lily Major, Keith Butler, Jeyan Thiyagalingam, and Francesca Tavazza. Unified graph neural network force-field for the periodic table: solid state applications. Digital Discovery, 2(2):346–355, 04 2023.
- [39]Albert P. Bartók, Mike C. Payne, Risi Kondor, and Gábor Csányi. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. Physical Review Letters, 104:136403, Apr 2010.pubmed.ncbi.nlm.nih.gov/20481899
- [40]Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99:014104, Jan 2019.DOI
- [41]Kristof Schütt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Research, pages 9377–9388. PMLR, 18–24 Jul 2021.arxiv.org/abs/2102.03150
- [42]Saro Passaro and C Lawrence Zitnick. Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs. In International Conference on Machine Learning (ICML), pages 27420–27438, 2023.arxiv.org/abs/2302.03655
- [43]Albert P. Bartok, Risi Kondor, and Gabor Csanyi. On representing chemical environments. Physical Review B, 87:184115, 2013.DOI
- [44]Alexander V. Shapeev. Moment tensor potentials: A class of systematically improvable interatomic potentials. Multiscale Modeling and Simulation, 14(3):1153–1173, 2016.DOI
- [45]Alexander Goscinski, Felix Musil, Sergey Pozdnyakov, Jigyas Nigam, and Michele Ceriotti. Optimal radial basis for density-based atomic representations. The Journal of Chemical Physics, 155(10):104106, 2021.
- [46]Anton Bochkarev, Yury Lysogorskiy, Sarath Menon, Minaam Qamar, Matous Mrovec, and Ralf Drautz. Efficient parametrization of the atomic cluster expansion. Physical Review Materials, 6:013804, 2022.DOI
- [47]Ilyes Batatia, Simon Batzner, David Peter Kovacs, Albert Musaelian, Gregor N. C. Simm, Ralf Drautz, Christoph Ortner, Boris Kozinsky, and Gabor Csanyi. The design space of e(3)-equivariant atom-centred interatomic potentials. Nature Machine Intelligence, 7(1):56–67, 2025.DOI
- [48]Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science, 2:718–728, 2022.DOI
- [49]Oliver T. Unke, Stefan Chmiela, Michael Gastegger, Kristof T. Schutt, Huziel E. Sauceda, and Klaus-Robert Muller. Spookynet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nature Communications, 12:7273, 2021.DOI
- [50]Yuchao Lin, Keqiang Yan, Youzhi Luo, Yi Liu, Xiaoning Qian, and Shuiwang Ji. Efficient approximations of complete interatomic potentials for crystal property prediction. In International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 21260–21287, 2023.arxiv.org/abs/2306.10045
- [51]Yajie Ji, Jiuyang Jiang, and Zhenli Xu. Machine-learning interatomic potentials for long-range systems. Physical Review Letters, 135(17):178001, 2025.DOI
- [52]Kevin Han, Haolin Cong, Bowen Deng, and Amir Barati Farimani. Smooth dynamic cutoffs for machine learning interatomic potentials. In International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pages 39672–39690, 2026.arxiv.org/abs/2601.21147
- [53]Yi-Lun Liao, Tess Smidt, Muhammed Shuaibi, and Abhishek Das. Generalizing denoising to non-equilibrium structures improves equivariant force fields. Transactions on Machine Learning Research, 2024.arxiv.org/abs/2403.09549

