Differentiable Particle-Mesh Ewald with Cartesian Tensor Message Passing for Learning Long-Range Electrostatics and Dipole Response
Zhiyue Guo 1 , Junjie Wang 1* , Haoting Zhang 1 , Zhixin Liang 1 Ziyang Yang 1 , Yujian Pan 1 , Jian Sun 1* 1 National Laboratory of Solid State Microstructures, School of Physics and Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing 210093, China · 2026
Read Differentiable Particle-Mesh Ewald with Cartesian Tensor Message Passing for Learning Long-Range Electrostatics and Dipole Response by Zhiyue Guo 1 , Junjie Wang 1* , Haoting Zhang 1 , Zhixin Liang 1 Ziyang Yang 1 , Yujian Pan 1 , Jian Sun 1* 1 National Laboratory of Solid State Microstructures, School of Physics and Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing 210093, China (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-01598.txt.
Introduction
Guo and co-authors make particle-mesh Ewald differentiable for learned potentials. Machine learning potentials can near quantum accuracy for short-range chemistry. Most designs stay local in their receptive field. They miss long-range electrostatics and polarization effects. Those effects are essential for ionic, polar, and interfacial systems. Ewald-based designs recover some long-range physics from local variables. But many compute reciprocal terms by direct summation. That leaves a gap with production molecular dynamics practice. Their framework embeds PME inside a tensor message passing network.
Charges come from scalar local features. Dipoles come from equivariant vector features. Both enter the same particle-mesh solver as charge density. Forces train end to end through FFT-space electrostatics. On molten NaCl, their channel gives the lowest force error.
Catalog record
| CREATOR | Zhiyue Guo 1 , Junjie Wang 1* , Haoting Zhang 1 , Zhixin Liang 1 Ziyang Yang 1 , Yujian Pan 1 , Jian Sun 1* 1 National Laboratory of Solid State Microstructures, School of Physics and Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing 210093, China |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.01598
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Differentiable Particle-Mesh Ewald with Cartesian Tensor Message Passing for Learning Long-Range Electrostatics and Dipole Response Zhiyue Guo 1 , Junjie Wang 1* , Haoting Zhang 1 , Zhixin Liang 1 Ziyang Yang 1 , Yujian Pan 1 , Jian Sun 1* 1 National Laboratory of Solid State Microstructures, School of Physics and Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing 210093, China Abstract Machine learning interatomic potentials (MLIPs) can approach quantum accuracy for short-range chemistry, yet most architectures remain fundamentally local and fail to describe the long-range electrostatic and polarization interactions essential for ionic, polar, and interfacial systems. Recent and contemporaneous Ewald-based MLIPs have shown that locally predicted electrostatic variables can recover important long-range physics, including multipolar response. However, many energy-based implementations still evaluate the reciprocal term by direct sums over k k vectors. This leaves a practical gap between long-range MLIP development and production molecular dynamics, where particle-mesh Ewald (PME) with O ( N log N ) O(N\log N) scaling is the standard. Here we introduce a fully differentiable PME framework for learned charges and, centrally, learned atomic dipoles within an E( n n )-equivariant Cartesian tensor message passing network. Charges are predicted from scalar local features, while dipole vectors are predicted from equivariant vector features and enter the same particle-mesh solver as an effective bound charge density. This dipolar density is constructed by analytic real-space gradients of the Hockney–Eastwood spline assignment weights, allowing charge–dipole and dipole–dipole long-range forces to be trained end-to-end through FFT-space electrostatics without direct charge or dipole supervision. On a charged-dimer test case, the differentiable PME module reproduces explicit Ewald energies and forces to numerical precision when the assignment-kernel deconvolution is enabled. On molten NaCl, the charge+dipole long-range channel gives the lowest force RMSE among the tested models while all energy RMSE values remain in the sub-meV/atom regime. Timing tests show the expected crossover from explicit Ewald summation to particle-mesh scaling. These results identify differentiable dipole PME as a scalable route toward polarization-aware MLIPs for condensed-phase and interfacial systems. 1 Introduction Machine learning interatomic potentials (MLIPs) have transformed atomistic materials modeling over the past decade, achieving first-principles accuracy at a fraction of the computational cost behler2007 ; bartok2010 ; thompson2015 ; shapeev2016 ; smith2017ani ; zhang2018dp ; schuett2018 ; batzner2022 ; batatia2022 ; wang2024 ; unke2021 . By learning the Born–Oppenheimer potential energy surface directly from electronic structure reference data, these methods have enabled simulations at time and length scales far beyond the reach of ab initio molecular dynamics. This progress has followed a clear architectural trajectory. Early neural-network and kernel potentials demonstrated that atom-centered descriptors could learn accurate local energy decompositions behler2007 ; bartok2010 ; smith2017ani , while moment and spectral descriptor families made this idea systematically improvable for materials thompson2015 ; shapeev2016 . Deep message-passing and equivariant architectures then replaced fixed descriptors with learned geometric representations. These models respect the fundamental symmetries of atomistic systems: the total energy is invariant to global translations and permutations of equivalent atoms, atom-wise features are permutation equivariant with respect to atom indexing, and scalar, vector, and tensor channels transform consistently under rotations. This symmetry-aware structure improves data efficiency for energies, forces, and tensorial properties schuett2018 ; schuett2021painn ; batzner2022 ; batatia2022 ; musaelian2023 ; wang2024 . Despite these advances, most MLIPs still rely on a locality ansatz : the total energy is decomposed into atom-centered contributions that depend only on the atomic environment within a finite cutoff radius r_{c} . The assumption is physically motivated by the nearsightedness of electronic matter and gives the linear scaling that makes large-scale simulations possible. It is also the source of a sharp limitation. Coulomb interactions decay as r^{-1} , charge–dipole interactions as r^{-2} , and polarization response can depend on the dielectric environment over length scales far beyond a typical 4–6 Å local cutoff. Short-range MLIPs can therefore interpolate accurately within many covalent and metallic training domains while still failing in ionic materials, polar molecular liquids, solid electrolytes you2024llzo , solid–liquid interfaces, and any system where charge redistribution or long-range screening controls the forces prodan2005 ; anstine2023longrange . Existing long-range MLIP strategies can be organized by how they relax this locality constraint. The most direct route is to enlarge the receptive field, either by increasing the cutoff, stacking more message-passing layers, or using multiscale short/long-cutoff modules. This keeps the model architecture close to standard local MLIPs, but the cost and the number of neighbors grow rapidly, while distant electrostatic information is still only represented implicitly. A second route adds physically motivated energy terms. PhysNet and SpookyNet, for example, combine neural short-range energies with learned charges, dispersion corrections, total-charge or spin information, and damped electrostatic terms physnet2019 ; spookynet2021 . Such decompositions impose useful asymptotic structure, but the learned electrostatic variables may depend on a chosen charge convention and are not always sufficient to describe non-local charge redistribution. Global electrostatic-variable models address part of this problem by constraining the charges themselves. Charge-equilibration approaches predict local electronegativities and hardnesses and then solve a global minimization problem to obtain charges mortier1986 ; rappe1991 ; streitz1994 ; behler2015 ; ko2021 . These methods can capture non-local charge transfer, but they introduce an additional global solve and inherit the well-known ambiguities of partial-charge definitions. Deep potential long-range (DPLR) models avoid direct atomic-charge fitting by using maximally localized Wannier centers as physically motivated charge carriers zhang2022dplr ; this provides a smoother description of electrostatics but requires Wannier information in the reference data. Classical polarizable force fields offer a complementary lesson: induced dipoles and atomic multipoles are often essential for dielectric response beyond fixed charges thole1981 ; ren2003amoeba , but in conventional force fields these variables are usually parameterized rather than learned end-to-end from quantum-mechanical energy-force data. A third family uses global descriptors, reciprocal operators, or non-local message passing to carry information beyond the local cutoff. Reciprocal-space neural networks construct structure-factor-like descriptors or learned reciprocal potentials from the full simulation cell, allowing local MLIPs to access full-cell information without assigning explicit atomic charges yu2022rsnn . LODE-type descriptors augment local representations with fields generated by long-ranged density kernels grisafi2019lode ; huguenin2023lode . Ewald message passing instead uses reciprocal-space filters as a non-local message-passing operation kosmala2023ewaldmp . Along the same line, LOREM treats inverse-power-law potential evaluation as equivariant message passing: scalar and higher-order equivariant charge channels transmit orientation-dependent geometric information beyond the cutoff rumiantsev2025lorem . These latent channels are not identical to physical atomic multipoles, but they are closely related in spirit to using high-order tensor representations as carriers of long-range information. A complementary energy-based route is to predict electrostatic variables and evaluate a separate long-range energy. The Latent Ewald Summation (LES) framework uses local descriptors to predict latent scalar charges and trains them only through total energies and forces cheng2023latent ; king2025charges , avoiding direct supervision on convention-dependent partial charges. Kim et al. kim2026 recently extended this strategy to a polarizable multipole hierarchy, including monopoles, dipoles, quadrupoles, and non-self-consistent linear response. Together, these developments show a broad movement toward combining local equivariant representations with global electrostatic or reciprocal-space operators. They differ in what is made non-local—descriptors, messages, latent charges, or multipoles—and should therefore be viewed as complementary rather than as a single settled formulation.