[Paper Review] Descriptors for Machine Learning Model of Generalized Force Field in Condensed Matter Systems

Puhan Zhang, Sheng Zhang+1 more|arXiv (Cornell University)|Jan 3, 2022

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes a symmetry-invariant descriptor framework for machine learning models to simulate the adiabatic dynamics of classical fields—such as lattice distortions, spins, and order parameters—in correlated electron systems. By leveraging group-theoretical methods, particularly bispectrum coefficients with reference irreducible representations, the approach enables accurate, symmetry-preserving energy predictions from local field configurations, enabling large-scale, first-principles-inspired simulations of complex condensed matter dynamics beyond empirical models.

ABSTRACT

We outline the general framework of machine learning (ML) methods for multi-scale dynamical modeling of condensed matter systems, and in particular of strongly correlated electron models. Complex spatial temporal behaviors in these systems often arise from the interplay between quasi-particles and the emergent dynamical classical degrees of freedom, such as local lattice distortions, spins, and order-parameters. Central to the proposed framework is the ML energy model that, by successfully emulating the time-consuming electronic structure calculation, can accurately predict a local energy based on the classical field in the intermediate neighborhood. In order to properly include the symmetry of the electron Hamiltonian, a crucial component of the ML energy model is the descriptor that transforms the neighborhood configuration into invariant feature variables, which are input to the learning model. A general theory of the descriptor for the classical fields is formulated, and two types of models are distinguished depending on the presence or absence of an internal symmetry for the classical field. Several specific approaches to the descriptor of the classical fields are presented. Our focus is on the group-theoretical method that offers a systematic and rigorous approach to compute invariants based on the bispectrum coefficients. We propose an efficient implementation of the bispectrum method based on the concept of reference irreducible representations. Finally, the implementations of the various descriptors are demonstrated on well-known electronic lattice models.

Motivation & Objective

• To develop a general, symmetry-preserving descriptor framework for machine learning models applied to classical fields in condensed matter systems.
• To address the lack of a systematic theory for descriptors in multi-scale dynamical modeling of strongly correlated electron systems.
• To enable accurate, large-scale simulations of classical field dynamics under the influence of quasi-equilibrium electrons, going beyond empirical energy models.
• To provide a systematic, group-theoretically grounded method for constructing invariants from local classical field configurations, especially for systems with internal symmetries.
• To demonstrate the feasibility and accuracy of the descriptor on benchmark models such as the Holstein, Jahn-Teller, and s-d Hamiltonians.

Proposed method

• Formulates a general theory of descriptors for classical fields in condensed matter systems, distinguishing models with and without internal symmetries.
• Proposes a group-theoretical approach based on decomposing local classical fields into irreducible representations (IRs) of the site-symmetry point group.
• Uses bispectrum coefficients of three IRs as fundamental invariants, analogous to scalar triple products, to construct symmetry-invariant features.
• Introduces a simplification strategy using reference irreducible representations to reduce over-completeness of bispectrum invariants while preserving amplitude and relative phase information.
• Employs deep neural networks as the learning model, with symmetry-invariant descriptors as input features to predict local energy from classical field configurations.
• Demonstrates the method on well-known models: scalar-field Holstein and Jahn-Teller models, and vector-field s-d model for itinerant magnets.

Experimental results

Research questions

• RQ1 How can machine learning models be systematically designed to preserve the symmetry of the electron Hamiltonian when modeling classical fields in correlated electron systems?
• RQ2 What is a general, mathematically rigorous framework for constructing symmetry-invariant descriptors for classical fields with or without internal symmetries?
• RQ3 How can the over-completeness of bispectrum invariants be mitigated while retaining essential physical information for learning?
• RQ4 Can the proposed descriptor enable accurate, large-scale simulations of classical field dynamics that go beyond empirical or phenomenological models?
• RQ5 How do different descriptor approaches—eigenvalue-based, symmetry-function-like, and group-theoretical—compare in performance and physical consistency?

Key findings

• The group-theoretical approach using bispectrum coefficients of irreducible representations provides a systematic and controlled method for constructing symmetry-invariant descriptors for classical fields in condensed matter systems.
• The proposed reference IR-based simplification reduces the number of bispectrum invariants while preserving the amplitude and relative phase of each irreducible representation, ensuring physical fidelity.
• The descriptor framework successfully captures the essential physics of classical fields in the Holstein model (scalar field) and the s-d model (vector field with internal symmetry), demonstrating its versatility.
• The method enables accurate prediction of local energy from classical field configurations, paving the way for large-scale, first-principles-inspired simulations of complex spatio-temporal dynamics in correlated materials.
• The framework overcomes limitations of empirical energy models by integrating the electronic structure response on-the-fly, allowing for dynamical coupling between electrons and classical fields.
• The implementation on benchmark models confirms the descriptor's ability to encode local symmetry and field structure with high accuracy, essential for simulating topological defects and inhomogeneous electronic states.