[Paper Review] Predicting thermoelectric properties from crystal graphs and material descriptors - first application for functional materials

Léo Laugier, Daniil Bash+6 more|arXiv (Cornell University)|Nov 15, 2018

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This study proposes a machine learning framework to predict thermoelectric power factor using crystal graphs (CGCNN), fully connected neural networks (FCNN), and XGBoost, leveraging crystal and atomic descriptors. The FCNN outperforms CGCNN and XGBoost by using DFT-informed material descriptors, demonstrating that physics-informed features are critical for accurate prediction of functional material properties.

ABSTRACT

We introduce the use of Crystal Graph Convolutional Neural Networks (CGCNN), Fully Connected Neural Networks (FCNN) and XGBoost to predict thermoelectric properties. The dataset for the CGCNN is independent of Density Functional Theory (DFT) and only relies on the crystal and atomic information, while that for the FCNN is based on a rich attribute list mined from Materialsproject.org. The results show that the optimized FCNN is three layer deep and is able to predict the scattering-time independent thermoelectric powerfactor much better than the CGCNN (or XGBoost), suggesting that bonding and density of states descriptors informed from materials science knowledge obtained partially from DFT are vital to predict functional properties.

Motivation & Objective

• To accelerate the discovery of functional thermoelectric materials by replacing computationally expensive DFT and Boltzmann transport calculations with machine learning.
• To investigate whether crystal graph-based models (CGCNN) can predict thermoelectric power factor without relying on DFT inputs.
• To evaluate the performance of FCNN and XGBoost using a rich set of material descriptors derived from the Materials Project and Matminer.
• To determine whether DFT-informed descriptors significantly improve prediction accuracy over basic crystal and atomic features.
• To establish a foundation for future data-driven design of thermoelectric materials using ML models trained on theoretical datasets.

Proposed method

• Trained a Crystal Graph Convolutional Neural Network (CGCNN) using only crystal structure and atomic information, without DFT inputs.
• Constructed a Fully Connected Neural Network (FCNN) using 28 material descriptors—such as bonding, atomic, and electronic properties—extracted from the Materials Project via Matminer.
• Applied XGBoost as a gradient-boosting baseline model for comparison with deep learning approaches.
• Used DFT + Boltzmann Transport Equation (BTE) computed power factors as ground truth labels for model training and evaluation.
• Optimized hyperparameters for FCNN and XGBoost using Bayesian hyperparameter search, with loss function MSE and optimizer RMSProp.
• Evaluated model performance using mean absolute error (MAE) on a dataset of 2,819,682 datapoints derived from 7,230 compounds with varying doping, temperature, and crystal direction.

Experimental results

Research questions

• RQ1 Can CGCNN models trained solely on crystal and atomic structure predict thermoelectric power factor accurately without DFT inputs?
• RQ2 How does the performance of FCNN with rich material descriptors compare to CGCNN and XGBoost in predicting DFT+BTE power factor?
• RQ3 To what extent do DFT-informed descriptors (e.g., bonding, density of states) improve prediction accuracy compared to basic structural features?
• RQ4 Does the inclusion of physics-informed descriptors significantly enhance model capability in predicting functional material properties?
• RQ5 Can machine learning models trained on theoretical datasets serve as reliable surrogates for expensive DFT and BTE calculations in thermoelectric materials screening?

Key findings

• The FCNN model, with three fully connected layers and optimized hyperparameters, achieved the highest prediction accuracy for thermoelectric power factor.
• The FCNN outperformed both CGCNN and XGBoost, indicating that DFT-informed descriptors such as bonding and density of states are essential for accurate prediction.
• CGCNN, which relies only on crystal graph information without DFT inputs, performed worse than both FCNN and XGBoost, suggesting that structural features alone are insufficient.
• The best-performing FCNN used ReLU, tanh, and sigmoid activations across three layers with 50, 41, and 41 neurons respectively, and employed a Robust scaler and RMSProp optimizer.
• The model achieved lower mean absolute error (MAE) than prior work predicting only the Seebeck coefficient, with improvements ranging from 30% to 46% across different temperature ranges.
• Despite limitations in the dataset—such as unstable polymorphs and unrealistic compounds—the FCNN demonstrated strong generalization, highlighting the value of enriched descriptors in ML-driven materials discovery.