Source-linked AI summary
Atom2Vec: learning atoms for materials discovery
Quan Zhou, Peizhe Tang, Shenxiu Liu, Jinbo Pan, Qimin Yan, Shou-Cheng Zhang
TL;DR
Materials machine learning needs useful atom representations, but conventional approaches learn mappings from materials information to labeled properties. Atom2Vec instead learns atom vectors from compound existence and uses them in materials-property models, recovering chemically meaningful organization and achieving accurate predictions. The authors focus on model-free vectors because model-based variants were inferior, likely owing to simplified modeling.
Problem
Materials research has accumulated large datasets, but existing approaches primarily learn materials-to-property mappings rather than atom knowledge itself.
Method
Atom2Vec generates atom–environment pairs from compound compositions and jointly learns high-dimensional atom representations through unsupervised methods.
Results
Atom2Vec vectors recover periodic-table groupings and support materials-property prediction, including formation-energy mean absolute errors as low as 0.15eV/atom and 0.27eV/atom in two tasks.
Takeaways & Limitations
The learned vectors capture atom similarities and properties and are presented as broadly applicable inputs for machine-learning problems in materials science.
Takeaways & Limitations
The workflow uses composition but omits crystal structure as a proof of concept, while model-based variants yield inferior vectors under simplified modeling.
Abstract
from arXiv · showhide
Exciting advances have been made in artificial intelligence (AI) during the past decades. Among them, applications of machine learning (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.
Atom2Vec Workflow
Atom2Vec learns atom representations from atom–environment relations in a materials database using composition-only information, without materials-property labels. Its model-free vectors recover chemically meaningful groupings, including the periodic-table organization of main-group elements.
- Atom2Vec Workflow: Atom-environment pairs are generated for each compound, forming a matrix whose entries count pair occurrences before Atom2Vec encodes atoms as vectors.The environment captures target-atom counts and the composition of the remaining atoms, while crystal structure is ignored.
- Atom2Vec Workflow: Atom2Vec learns atom properties from compound existence alone, using unsupervised data rather than labels for any specific materials property.The authors describe this as a way for machines to learn atom knowledge without bias toward one property aspect.
- Atom2Vec Workflow: Raw atom-environment row vectors are sparse, motivating algorithms that learn atoms and environments simultaneously rather than relying on direct data statistics.The method is motivated by the reciprocal idea that knowing environments can help learn atoms, and vice versa.
- Atom Vectors: The study focuses on model-free vectors because model-based machines produced inferior vectors, probably because of their simplified modeling.Model-based vectors are retained only for comparison when necessary.
- Atom Vectors: Model-free Atom2Vec vectors classify main-group elements into groups matching the periodic table, separating active metals and nonmetals while clustering Groups III–V.The vectors also place heavier-period elements toward more metallic neighbors, such as Tl near alkali metals and Pb near alkaline-earth metals.
Atom Vectors.
Atom vectors capture chemically meaningful structure across main-group elements, transition metals, lanthanoids, actinoids, and functional groups, while improving materials-property prediction over empirical descriptors.
- Atom Vectors: 0.15eV/atom mean absolute error is achieved for elpasolite formation-energy prediction using model-free atom vectors.This error is reported as almost within the error range of first-principles calculations.
- Atom Vectors: Atom-vector clustering recovers chemically meaningful groups, including periodic-table-like organization for main-group elements and distinct transition-metal and lanthanoid clusters.The learned space also reflects trends such as valence and similarities between Fe-like transition metals and Group II atoms.
- Atom Vectors: Atom vectors achieve comparable accuracy to the 18-electron rule for half-Heusler metal/insulator classification.The comparison uses logistic regression with model-based and model-free vectors of dimension d = 20.
- Atom Vectors: Atom2Vec is proposed as unsupervised learning from known-materials databases, with learned vectors intended to be broadly applicable across materials-science machine-learning tasks.The approach learns from compound existence without using specific materials-property labels.
Summary and Outlook.
Atom2Vec learns atom and environment representations from compound data, using model-free and model-based approaches that support clustering and materials-property prediction. The workflow combines normalized atom-environment statistics, compact vector representations, composition models, score functions, and supervised predictors.
- Data and representations: The unsupervised feature-learning dataset uses binary, ternary, and quaternary inorganic compounds, comprising nearly 90% of about 60k Materials Project compounds.Compounds containing more than four element types are screened out for simplicity.
- Model-free methods: For model-free vectors, p = 2 weights environment information and defines distance between normalized atom vectors as dist(u1, u2) = 1 −u1 · u2.The inner product of normalized vectors represents similarity through matching across environments.
- Data and representations: Normalized atom-environment counts provide primitive distribution profiles, which SVD projects into compact feature vectors based on leading singular values.The resulting vectors retain the structure of inner products and support distance-based analysis.
- Model-based methods: Model-based methods represent environments in the atom-vector space, combine constituent atom vectors with a composition model, and optimize vectors using pair-existence scores.The proof-of-concept composition model sums the atom vectors in an environment, while several score functions evaluate atom-environment compatibility.
- Feature evaluation: A one-hidden-layer neural network predicts elpasolite formation energies from concatenated atom vectors, while ridge and logistic regression address half-heusler formation-energy and metal-insulator tasks.The neural-network setup uses a 4d-dimensional input, 10 hidden neurons, and hold-out validation; the regression and classification tasks use the same data splits.