General Learning of the Electric Response of Inorganic Materials

Bradley A. A. Martin, Keith T. Butler, Venkat Kapil, Alex M. Ganose · 2025

Read General Learning of the Electric Response of Inorganic Materials by Bradley A. A. Martin, Keith T. Butler, Venkat Kapil, Alex M. Ganose (2025) 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-2508-17870.txt.

Introduction

Bradley Martin and co-authors build a machine learning model that predicts how inorganic materials answer an electric field, in a 2025 paper. Such answers, including polarization and dielectric behavior, shape devices from capacitors to memory chips.

Their model, MACE-Field, learns one energy function that depends on both atom positions and the applied field. Exact calculus on that function then yields the electric properties, with physical rules such as symmetry built in by design. The add-on can upgrade existing general models. Tests against standard quantum calculations were strong: the model reproduced polarization across many crystals, and for two test materials it matched reference spectra and switching loops. One trained model now spans much of inorganic chemistry.

Catalog record

CREATORBradley A. A. Martin, Keith T. Butler, Venkat Kapil, Alex M. Ganose
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17870

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

General Learning of the Electric Response of Inorganic Materials Bradley A. A. Martin, Keith T. Butler, Venkat Kapil, Alex M. Ganose Abstract We present MACE-Field , a field-aware O ⁡ ( 3 ) O(3) -equivariant interatomic potential that provides a compact, derivative-consistent route to dielectric properties (such as polarisation 𝐏 \mathbf{P} , Born effective charges Z ∗ Z^{*} and polarisability 𝜶 \bm{\alpha} ) and finite-field simulations across chemistry for inorganic solids. A uniform electric field is injected within each message-passing layer via a Clebsch–Gordan tensor–product which couples the field to latent spherical–tensor features, and perturbs them via an equivariant residual mixing. This plug-in design preserves the standard MACE readout and can inherit existing MACE foundation weights , turning pretrained models into field-aware ones with minimal change. To demonstrate, we train: (i) a cross-chemistry ferroelectric polarisation model (2.5k nonpolar → \!\to polar polarisation branches), (ii) a cross-chemistry BECs/polarisability model ( ∼ \sim 6k Materials Project dielectrics spanning 81 elements), and (iii–iv) single-material molecular dynamics on BaTiO 3 and α \alpha -SiO 2 . The models recover polarisation branches and spontaneous polarisation, predict Z ∗ Z^{*} and 𝜶 \bm{\alpha} (hence ε ∞ \varepsilon_{\infty} ) across diverse chemistries, and reproduce BaTiO 3 hysteresis and the IR/Raman and dielectric spectra of α \alpha -quartz, benchmarking comparatively with Allegro–pol. I Introduction Electric polarisation is a central quantity for understanding the response of insulating crystals to external electric fields and for designing ferroelectrics, piezoelectrics, dielectrics, and nonlinear–optical materials. Unlike finite molecules, however, the polarisation of a periodic solid is not an ordinary dipole per unit volume: within the modern theory of polarisation it is a geometric (Berry) phase and therefore multivalued on a lattice of polarisation quanta. King-Smith and Vanderbilt 1993 ; Resta 1994 In practice, only differences in polarisation are gauge–invariant and measurable, which immediately complicates the learning and prediction of polarisation–related properties. First–principles techniques based on density–functional perturbation theory (DFPT) and finite–field Berry–phase methods provide rigorous access to linear and nonlinear response—including Born effective charges (BECs), electronic polarisabilities, dielectric tensors, and electro–optic coefficients—but at significant computational cost that limits system size, chemical diversity, and time scales Gonze and Lee 1997 ; Nunes and Gonze 2001 ; Souza et al. 2002 . Machine–learned interatomic potentials (MLIPs) promise to bridge this gap by reproducing ab initio accuracy at orders–of–magnitude lower cost Behler and Parrinello 2007 ; Schütt et al. 2017 ; Batzner et al. 2022 . Recent equivariant message–passing architectures (e.g. ACE and MACE ) systematically expand local environments into irreducible spherical components and have delivered state–of–the–art accuracy and transferability for energies, forces, and stresses Drautz 2019 ; Batatia et al. 2022 . Parallel advances have begun to incorporate electrostatics and field response: unified electric–enthalpy learning that differentiates a single scalar functional to obtain \,\mathbf{P}\, , \,Z^{*}\, , and \,\bm{\alpha}\, for select materials Falletta et al. 2025 ; latent long–range treatments that recover polarisation and BECs from learned charge fields Cheng et al. 2025 ; and charge–augmented ACE variants Rinaldi et al. 2024 . Related property–prediction models have also targeted anisotropic dielectric tensors directly from structure, Lou et al. 2024 and molecular/liquid studies have demonstrated direct ML prediction of dipoles and polarisabilities for bulk water Kapil et al. 2024 . Despite progress, two limitations persist. First, most unified electric-response MLIPs are trained as single-material models, which restricts transfer across chemistries and space groups. Second, their architectures often diverge from widely used backbones, hindering reuse of foundation models pretrained on large energy/force/stress datasets. Given the relative scarcity of high-quality dielectric labels, the ability to inherit and fine-tune such foundations is especially valuable. Therefore we introduce MACE-Field Martin, B A. A., Butler, K T. 2025 , a symmetry–aware extension of MACE that learns a single electric enthalpy \mathcal{F}(\{\mathbf{R}\},\mathbf{E}) across diverse inorganic crystals and obtains all dielectric observables by exact differentiation : \mathbf{P}=-\Omega^{-1}\frac{\partial\mathcal{F}}{\partial\mathbf{E}},\qquad Z^{*}_{\kappa,ij}=\Omega\,\frac{\partial P_{i}}{\partial u_{\kappa j}},\qquad\alpha_{ij}=\frac{\partial P_{i}}{\partial E_{j}}. A uniform field couples inside each message-passing layer to latent equivariant features via Clebsch–Gordan tensor products with equivariant residual mixing; the readout remains scalar ( L{=}0 ). Berry–phase multivaluedness is handled by a branch–invariant loss. Because all quantities derive from one scalar, Maxwell reciprocity and the acoustic sum rule follow by construction. Our two key advances. 1. Foundation-model inheritance. Field coupling is a plug-in at the latent irrep level that leaves the standard MACE readout/baselines unchanged, so existing MACE foundation weights can be reused and fine-tuned to become field-aware. 2. Cross-chemistry training. We train models across broad chemistries rather than per-material: a ferroelectric path model ( \sim 2.5k structures, 61 elements), a dielectric/BEC model ( \sim 6k MP DFPT entries, 81 elements), and benchmark single-material MD models (BaTiO 3 , \alpha -SiO 2 ). Together these address the two prevailing limitations—per-material training and non-reusable architectures—by providing a derivative-consistent, foundation-compatible framework that scales across chemistry and predicts \mathbf{P} , Z^{*} , and \bm{\alpha} for diverse inorganic solids. . The remainder of the paper details the relevant background theory (§ II ), architecture and design choices (§ III ), datasets and training protocols (§ IV ), quantitative results (§ V ), and limitations and outlook (§ VI ). II Theory II.1 Modern theory of polarisation (Berry phase) In periodic insulators the macroscopic polarisation is not the naive dipole per unit volume, because the position operator is ill-defined under periodic boundary conditions. The modern theory of polarisation resolves this by expressing the electronic contribution as a Berry phase accumulated by the occupied Bloch bands along loops in the Brillouin zone (BZ) King-Smith and Vanderbilt 1993 ; Resta 1994 . Decomposing the total polarisation into ionic and electronic parts, \mathbf{P}=\mathbf{P}_{\text{ion}}+\mathbf{P}_{\text{el}},\qquad\mathbf{P}_{\text{ion}}=\frac{-e}{\Omega}\sum_{\kappa}Z_{\kappa}\,\mathbf{R}_{\kappa}, (1) where e is the electron charge, \Omega the cell volume, Z_{\kappa} the nuclear charge for site \kappa , and \mathbf{R}_{\kappa} the corresponding the position vector. The electronic part of the polarisation is \mathbf{P}_{\text{el}}=-\frac{e}{(2\pi)^{3}}\sum_{n\in\text{occ}}\int_{\text{BZ}}\!d\mathbf{k}\ \langle u_{n\mathbf{k}}\,|\,i\nabla_{\mathbf{k}}\,|\,u_{n\mathbf{k}}\rangle, (2) where |u_{n\mathbf{k}}\rangle are cell-periodic Bloch functions. Only changes of \mathbf{P} between adiabatically connected states are gauge-invariant; absolute values are defined modulo the polarisation lattice generated by the quanta \mathbf{Q}_{k}=\frac{e\,\mathbf{a}_{k}}{\Omega},\qquad k=1,2,3, (3) with \{\mathbf{a}_{k}\} the direct lattice vectors King-Smith and Vanderbilt 1993 ; Resta 1994 . This multivaluedness must be handled explicitly in learning problems (Sec. III ). II.2 Finite electric fields and DFPT/finite-field formalisms A uniform static electric field, \mathbf{E} , couples to polarisation through the electric enthalpy (fixed- \mathbf{E} ) \mathcal{F}(\{\mathbf{R}\},\mathbf{E})=E_{0}(\{\mathbf{R}\})-\Omega\,\mathbf{E}\!\cdot\!\mathbf{P}(\{\mathbf{R}\}), (4) where E_{0} is the zero-field energy and \Omega the cell volume Nunes and Gonze 2001 ; Souza et al. 2002 . This functional permits direct optimisation of field-polarised states and yields forces and stresses at fixed \mathbf{E} . Within density-functional perturbation theory (DFPT), linear-response equations for first-order wavefunctions under atomic displacements or homogeneous fields provide access to the complete set of dielectric, piezoelectric, and vibrational properties at zero field without performing finite differences Gonze and Lee 1997 . II.3 Macroscopic response tensors as energy derivatives In SI units, the primary response functions used here are defined by exact derivatives of the electric enthalpy: \displaystyle\mathbf{P}(\{\mathbf{R}\}) \displaystyle=-\frac{1}{\Omega}\,\frac{\partial\mathcal{F}}{\partial\mathbf{E}}, (5) \displaystyle Z^{*}_{\kappa,ij}(\{\mathbf{R}\}) \displaystyle=\Omega\,\frac{\partial P_{i}}{\partial u_{\kappa j}}=\frac{\partial F_{\kappa j}}{\partial E_{i}}, (6) \displaystyle\alpha_{ij}(\{\mathbf{R}\}) \displaystyle=\frac{\partial P_{i}}{\partial E_{j}}. (7) Here u_{\kappa j} denotes the Cartesian displacement of sublattice \kappa , F_{\kappa j}=-\partial\mathcal{F}/\partial u_{\kappa j} the force component, Z^{*} the Born effective charge tensor (dynamical charge), and \bm{\alpha} the electronic polarisability per unit

Read the complete text