Source-linked AI summary
Universal Fragment Descriptors for Predicting Electronic Properties of Inorganic Crystals
Olexandr Isayev, Corey Oses, Cormac Toher, Eric Gossett, Stefano Curtarolo, Alexander Tropsha
TL;DR
Materials discovery is constrained by a vast, expensive search space and limited characterization. This paper develops PLMF descriptors with QMSPR machine-learning models to predict eight properties of stoichiometric inorganic crystals. The models achieve strong reported predictive performance across classification and regression tasks and are intended to accelerate targeted materials screening.
Problem
The enormous diversity of possible inorganic materials and the cost of conventional characterization make trial-and-error and brute-force exploration difficult.
Method
The paper combines geometry-derived Property-Labeled Materials Fragments with machine-learning QMSPR models trained on AFLOW data to predict eight electronic and thermomechanical properties.
Results
The metal/insulator classifier achieves AUC 0.98, sensitivity 0.95, specificity 0.92, and CCR 0.93 across 26,674 materials.
Takeaways & Limitations
The framework supports rapid screening and materials design using atomic species and positions as input, including searches constrained by desired property combinations.
Abstract
from arXiv · showhide
Historically, materials discovery has been driven by a laborious trial-and-error process. The growth of materials databases and emerging informatics approaches finally offer the opportunity to transform this practice into data- and knowledge-driven rational design. By using data from the AFLOW repository for high-throughput ab-initio calculations, we have generated Quantitative Materials Structure-Property Relationship (QMSPR) models to predict eight critical electronic and thermomechanical materials properties, such as the metal/insulator classification, band gap energy, bulk and shear moduli, Debye temperature, and heat capacity. The prediction accuracy obtained with these QMSPR models approaches training data for virtually any stoichiometric inorganic crystalline material. The success and universality of these models is attributed to the construction of new materials descriptors---referred to as the universal Property-Labeled Materials Fragments (PLMF). The representation requires only minimal structural input and affords straightforward model interpretation in terms of simple heuristic design rules that guide rational materials design. This study demonstrates the power of materials informatics to dramatically accelerate the search for new materials.
I. INTRODUCTION
Materials discovery faces an enormous, expensive search space that limits trial-and-error and brute-force computational exploration. The paper introduces universal PLMF descriptors combined with QMSPR models to predict key properties across stoichiometric inorganic crystals.
- The potential search space includes roughly 3×10^11 stoichiometric quaternary compounds, 10^13 quinary combinations, and as many as 10^100 theoretical materials.
- Materials characterization is difficult because calculations such as band structures can require expensive finite-size, charge-correction, and beyond-DFT methods.
- Prior structure-property relationships range from empirical rules to specialized ML models for selected material classes, rather than a universal framework.
- The paper introduces fragment descriptors and QMSPR models to predict eight electronic and thermomechanical properties for virtually any stoichiometric inorganic crystalline material.
- Unlike approaches requiring computationally obtained input quantities, the proposed models use tabulated or geometry-derived inputs and require no further ab-initio calculations after training.
II. METHODS
The methods construct PLMF descriptors from crystal geometry, atomic properties, and crystal-wide features, then use gradient boosting models to predict one classification and seven regression targets. The workflow combines graph-based fragment construction, descriptor filtering, validation, and sequential property prediction.
- Data preparation: Training data from AFLOW cover electronic and thermomechanical properties for diverse compounds, while thermomechanical models predict previously uncharacterized AFLOW compounds.
- Descriptors: PLMFs label graph vertices with tabulated atomic and physical properties, while crystal-wide descriptors encode lattice geometry, density, composition, symmetry, and cell size.
- PLMF construction: Voronoi-Dirichlet partitioning identifies atomic connectivity, producing a three-dimensional graph whose edges require shared Voronoi faces and interatomic distances within covalent-radius sums.
- PLMF construction: The graph adjacency matrix is partitioned into path fragments encoding up to four atoms and circular fragments encoding nearest-neighbor shells, with fragment length restricted to control descriptor complexity.
- Descriptors: After low-variance filtering, the concatenated representation contains 2,494 descriptors.
- QMSPR modeling: Gradient boosting decision trees are trained without hand tuning or variable selection, validated by label scrambling, and assessed with five-fold cross-validation.
- Integrated modeling workflow: The integrated workflow classifies materials as metals or insulators, predicts band gaps only for insulators, and independently predicts six thermomechanical properties.
III. RESULTS
The models accurately predict electronic and thermomechanical properties across broad materials datasets, while descriptor analyses reveal interpretable structural and chemical trends. These trends support simple heuristics for classification and potential strategies for tuning band gaps, stiffness, and Debye temperatures.
- Electronic-property models: AUC 0.98 and overall classification rate 0.93 demonstrate strong external performance for distinguishing metals from insulators across 26,674 materials.The model achieved 0.95 sensitivity and 0.92 specificity, with 2,103 misclassified materials.
- Thermomechanical-property models: At least 90% of the full training set is predicted within 25% of calculated values for the thermomechanical regression models.For both bulk and shear moduli, over 85% of materials are predicted within 20 GPa; predictions for GVRH, θD, and αV are slightly underestimated at higher values.
- Electronic-property models: 86% classification accuracy is achieved by assigning quadrant-I materials as insulators and materials outside quadrant I as metals.The heuristic misclassifies 3,621 materials: 2,414 as insulators and 1,207 as metals.
- Descriptor interpretation: Band-gap dependence on average bond ionization-potential differences is small but increases monotonically, consistent with larger gaps in polar materials such as oxides and fluorides.Band-gap response also depends on interactions with descriptor variability and density, making tuning a highly non-convex multidimensional problem.
- Thermomechanical-property models: The volume per atom is especially important for thermomechanical models, with larger interatomic distances generally associated with lower bulk modulus.The relationship is attributed to the connection between tight atomic binding, stronger bonds, and higher stiffness.
- Descriptor interpretation: Debye-temperature trends suggest tuning opportunities through bond-affinity descriptors, cell dimensions, and anisotropy: uniform expansion lowers θD, whereas strong elongation can increase it.The reported dependence links larger cell size to weaker bonding and elongated or layered systems to phonon-related interactions.
IV. DISCUSSION
The universal QMSPR framework is intended to improve high-throughput materials screening by predicting key properties across stoichiometric inorganic crystals. Its design also supports constraint-based materials selection, while theoretical structures remain an important source of prediction outliers.
- IV. DISCUSSION: High-throughput DFT screening is costly because medium-sized structures require substantial computation, motivating faster predictive models.An average calculation for a structure of about 50 atoms per unit cell is described as time-consuming even at high-throughput rates.
- IV. DISCUSSION: Interaction diagrams support designing materials that satisfy multiple constraints, including matched thermal expansion coefficients for optoelectronic applications.Differences in thermal expansion coefficients can cause bending and cracking during growth, so matching them is a practical design consideration.
- IV. DISCUSSION: The models show minor deviations, but theoretical structures are among the strongest outliers because their true stability conditions may remain undetermined.The passage notes that some theoretical structures may require high-pressure or high-temperature conditions, if they are stable at all.
- IV. DISCUSSION: The framework predicts eight electronic and thermomechanical properties across all 230 space groups and most periodic-table elements.Targets include metal/insulator classification, band gap, bulk and shear moduli, Debye temperature, heat capacities, and thermal expansion coefficient.