Graph-based Descriptors for Condensed Matter

An Wang,Gabriele C. Sosso

2024-08-12

Abstract:Computational scientists have long been developing a diverse portfolio of methodologies to characterise condensed matter systems. Most of the descriptors resulting from these efforts are ultimately based on the spatial configurations of particles, atoms, or molecules within these systems. Noteworthy examples include symmetry functions and the smooth overlap of atomic positions (SOAP) descriptors, which have significantly advanced the performance of predictive machine learning models for both condensed matter and small molecules. However, while graph-based descriptors are frequently employed in machine learning models to predict the functional properties of small molecules, their application in the context of condensed matter has been limited. In this paper, we put forward a number of graph-based descriptors (such as node centrality and clustering coefficients) traditionally utilised in network science, as alternative representations for condensed matter systems. We apply this graph-based formalism to investigate the dynamical properties and phase transitions of the prototypical Lennard-Jones system. We find that our graph-based formalism outperforms symmetry function descriptors in predicting the dynamical properties and phase transitions of this system. These results demonstrate the broad applicability of graph-based features in representing condensed matter systems, paving the way for exciting advancements in the field of condensed matter through the integration of network science concepts.

Computational Physics,Disordered Systems and Neural Networks

What problem does this paper attempt to address?

The paper attempts to address the problem of developing a new graph-theoretic descriptor in condensed matter systems to better characterize the structural and dynamic properties of the system. Specifically, the authors argue that traditional Euclidean geometry-based methods (such as symmetry functions and Smooth Overlap of Atomic Positions (SOAP) descriptors) perform limitedly under different thermodynamic conditions and cannot fully overcome the non-independent and identically distributed (non-i.i.d.) problem when dealing with many-body systems. The main contribution of the paper is the proposal of several graph-theoretic descriptors (such as node centrality and clustering coefficient) and the application of these descriptors to the study of classical Lennard-Jones systems. By comparing with previous methods using symmetry functions, the paper demonstrates the superiority of graph-theoretic descriptors in predicting the dynamic properties and phase transitions of the system. These results indicate that graph-theoretic features can effectively capture the structural invariants of condensed matter systems under different thermodynamic conditions, thereby making the models more robust and transferable. Overall, the goal of the paper is to improve the characterization methods of particle structure and dynamic properties in condensed matter systems by introducing graph-theoretic descriptors. Particularly when dealing with complex disordered materials, this approach can provide more intuitive and accurate analytical tools.

Graph-based Descriptors for Condensed Matter

Graph-Based Feature Engineering to Predict the Dynamical Properties of Condensed Matter

Advancing descriptor search in materials science: feature engineering and selection strategies

Persistent homology-based descriptor for machine-learning potential

Atomistic graph networks for experimental materials property prediction

Graph Dynamical Networks for Unsupervised Learning of Atomic Scale Dynamics in Materials

Graph attention neural networks for mapping materials and molecules beyond short-range interatomic correlations

Universal and Interpretable Classification of Atomistic Structural Transitions Via Unsupervised Graph Learning

A String-Graph Approach to Molecular Geometry

Chemistry and Local Environment Adaptive Representation Graphs as Material Descriptors

Exploiting graph theory in MD simulations for extracting chemical and physical properties of materials

Learning physical descriptors for materials science by compressed sensing

Inherent structural descriptors via machine learning

Density-of-states similarity descriptor for unsupervised learning from materials data

Topological Analysis of Molecular Dynamics Simulations using the Euler Characteristic

Persistent homology-based descriptor for machine-learning potential of amorphous structures

A data driven approach to classify descriptors based on their efficiency in translating noisy trajectories into physically-relevant information

Graphic Characterization and Clustering Configuration Descriptors of Determinant Space for Molecules

Modified reverse degree descriptors for combined topological and entropy characterizations of 2D metal organic frameworks: applications in graph energy prediction

Predicting Material Properties Using a 3D Graph Neural Network with Invariant Local Descriptors

Self-adaptable Materials Structure Descriptor Based on Graph Attention Network for Machine Learning