Atom2Vec: learning atoms for materials discovery (1807.05617v1)

Published 15 Jul 2018 in physics.comp-ph and cond-mat.mtrl-sci

Abstract: Exciting advances have been made in AI during the past decades. Among them, applications of ML and deep learning techniques brought human-competitive performances in various tasks of fields, including image recognition, speech recognition and natural language understanding. Even in Go, the ancient game of profound complexity, the AI player already beat human world champions convincingly with and without learning from human. In this work, we show that our unsupervised machines (Atom2Vec) can learn the basic properties of atoms by themselves from the extensive database of known compounds and materials. These learned properties are represented in terms of high dimensional vectors, and clustering of atoms in vector space classifies them into meaningful groups in consistent with human knowledge. We use the atom vectors as basic input units for neural networks and other ML models designed and trained to predict materials properties, which demonstrate significant accuracy.

Citations (169)

View on Semantic Scholar

Summary

Atom2Vec: Learning Atomic Representations for Enhanced Materials Discovery

The paper "Atom2Vec: learning atoms for materials discovery" by Zhou et al. outlines an innovative approach to atomic-level representation learning, utilizing machine learning techniques derived from linguistics to forecast material properties. This work highlights how unsupervised learning can be leveraged to decipher atomic properties directly from data, presenting a shift towards fully data-driven materials discovery.

Atom2Vec is designed to learn atomic properties from materials databases in an unsupervised manner, relying solely on the existence of chemical compounds rather than specific material properties. This method bridges the gap between atomic-level data representation and practical materials discovery, aiming to overcome limitations imposed by previous approaches that depended heavily on manually curated atomic descriptors. Atom2Vec represents atoms as high-dimensional vectors, abstracting their physical and chemical characteristics, which can subsequently be used as inputs in predictive models for material properties.

Methodology Overview

The Atom2Vec workflow begins by creating atom-environment pairs derived from chemical compound datasets. The environment for an atom in a compound is described by the presence and quantities of other atoms, focusing initially on chemical composition while disregarding structural information. An atom-environment matrix is constructed, capturing the frequency of pairwise occurrences between atoms and their respective environments. Singular value decomposition (SVD) is then applied to derive atom vectors from this matrix, encoding atomic properties in a more compact form than raw input data.

Evaluation and Results

The efficacy of Atom2Vec was validated through its ability to rediscover known atomic organizations such as the periodic table. Grouping algorithms applied to the learned atom vectors reliably classify elements into their respective categories, demonstrating coherence with recognized chemical behavior. For example, cluster analysis accurately separates alkali metals, alkaline earth metals, and nonmetals, with elemental similarities illustrated via vector clustering and projection in principal component spaces.

Furthermore, the paper applies Atom2Vec vectors in predictive tasks, including formation energy prediction for elpasolite compounds and classification problems within half-heusler alloys. The learned atomic representations enhance prediction accuracy compared to traditional empirical descriptors, with notable improvements in model-based tasks. These results underscore the robustness of Atom2Vec vectors in capturing relevant material details and supporting ML-driven materials exploration.

Implications and Future Directions

Atom2Vec represents a significant step towards automated materials discovery by providing a generalizable, machine-learned representation of atomic properties. The implications extend beyond improving prediction precision; Atom2Vec vectors have the potential to be foundational building blocks in broader ML frameworks within materials science. As computational models for materials exploration become increasingly intricate, integrating sophisticated representations like Atom2Vec ensures enhanced model reliability and predictive efficiency.

Future research could explore advanced environment descriptions, potentially involving structural insights and higher-order data representations. The unsupervised learning approach's scalability to varied datasets and integration into recursive and graph-based neural architectures remains a promising direction. Continuing advancements could refine the Atom2Vec methodology further, achieving comprehensive atomic feature learning while remaining agnostic to the specific nature of material properties.