Abstract:Searching for technologically promising crystalline materials with desired thermal transport properties requires an electronic level comprehension of interatomic interactions and chemical intuition to uncover the hidden structure-property relationship. Here, we propose two chemical bonding descriptors, namely negative normalized integrated crystal orbital Hamilton population (normalized -ICOHP) and normalized integrated crystal orbital bond index (normalized ICOBI) and unravel their strong correlation to both lattice thermal conductivity (LTC) and rattling effect characterized by mean squared displacement (MSD). Our new descriptors outperform empirical models and the sole -ICOHP quantity in closely relating to extreme LTCs by testing on a first-principles dataset of over 4,500 materials with 62 distinct species. The Pearson correlation of both descriptors with LTC are significantly higher in magnitude compared with the traditional simple rule of average mass. We further develop crystal attention graph neural networks (CATGNN) model and predict our proposed descriptors of ~200,000 materials from existing databases to screen potentially ultralow and high LTC materials. We select 367 (533) with low (high) normalized -ICOHP and ICOBI for first-principles validation. The validation shows that 106 dynamically stable materials with low normalized -ICOHP and ICOBI have LTC less than 5 W/mK, among which 68% are less than 2 W/mK, while 13 stable materials with high normalized -ICOHP and ICOBI possess LTC higher than 100 W/mK. The proposed normalized -ICOHP and normalized ICOBI descriptors offer deep insights into LTC and MSD from chemical bonding principles. Considering the cheap computational cost, these descriptors offer a new reliable and fast route for high-throughput screening of novel crystalline materials with extreme LTCs for applications such as thermoelectrics and electronic cooling.

Accelerating Discovery of Extreme Lattice Thermal Conductivity by Crystal Attention Graph Neural Network (CATGNN) Using Chemical Bonding Intuitive Descriptors

High-throughput computational discovery of 3218 ultralow thermal conductivity and dynamically stable materials by dual machine learning models

Lattice Thermal Conductivity: An Accelerated Discovery Guided by Machine Learning

Accelerated Discovery and Design of Ultralow Lattice Thermal Conductivity Materials Using Chemical Bonding Principles

Exploring low lattice thermal conductivity materials using chemical bonding principles

CrysGraphFormer: an equivariant graph transformer for prediction of lattice thermal conductivity with interpretability

Predicting lattice thermal conductivity of semiconductors from atomic-information-enhanced CGCNN combined with transfer learning

Efficient Calculation of the Lattice Thermal Conductivity by Atomistic Simulations with Ab Initio Accuracy

An interpretable formula for lattice thermal conductivity of crystals

Prediction of lattice thermal conductivity with two-stage interpretable machine learning

Identification of Crystalline Materials with Ultra-Low Thermal Conductivity Based on Machine Learning Study

High-throughput Study and Machine Learning on MAX and MAB Phases: New Materials and Fingerprints of Superior Lattice Thermal Conductivities

A High-Throughput Descriptor for Prediction of Lattice Thermal Conductivity of Half-Heusler Compounds

High-throughput study and machine learning on MAX and MAB phases: new materials and fingerprints of superior lattice thermal conductivity

Charting Lattice Thermal Conductivity of Inorganic Crystals

Predicting lattice thermal conductivity from fundamental material properties using machine learning techniques

Lattice Thermal Conductivity of 2D Nanomaterials: A Simple Semi-Empirical Approach

Discovery of low thermal conductivity compounds with first-principles anharmonic lattice dynamics calculations and Bayesian optimization

Large-scale machine-learning-assisted exploration of the whole materials space

Searching High Temperature Superconductors with the assistance of Graph Neural Networks

Scale-invariant machine-learning model accelerates the discovery of quaternary chalcogenides with ultralow lattice thermal conductivity