Towards Universal Material Property Prediction with Deep Learning and Single-Descriptor electronic Density

TL;DR

A universal machine learning framework based solely on a physically grounded and theoretically rigorous descriptor -- electronic charge density is developed, which enables accurate prediction of eight different material properties and demonstrates outstanding multi-task learning capability.

Abstract

Owing to its high scalability and computational efficiency, machine learning methods have been increasingly integrated into various scientific research domains, including ab initio-based materials design. It has been demonstrated that, by incorporating modern machine learning algorithms, one can predict material properties with practically acceptable accuracy. However, one of the most significant limitations that restrict the widespread application of machine learning is its lack of transferability, as a given framework is typically applicable only to a specific property. The origin of this limitation is rooted in the fact that a material's properties are determined by multiple degrees of freedom -- and their complex interplay -- associated with nuclei and electrons, such as atomic type, structural symmetry, and the number and quantum states of the valence electrons, among others. The inherent complexity rules out the possibility of a single machine learning framework providing a full description of these critical quantities. In this paper, we develop a universal machine learning framework based solely on a physically grounded and theoretically rigorous descriptor -- electronic charge density. Our framework not only enables accurate prediction of eight different material properties (with R$^2$ values up to 0.94), but also demonstrates outstanding multi-task learning capability, as prediction accuracy improves when more target properties are incorporated into a single training process, thereby indicating excellent transferability. These results represent a significant step toward realizing the long-standing goal of a universal machine learning framework for the unified prediction of all material properties.

Figures (5)

• Figure 1: Schematic illustration of the overall workflow of the MSA-3DCNN model and its main components. (a) Data preparation process. (b) The main body of MSA-3DCNN workflow, comprising two major blocks: multi-scale convolution and multi-head attention mechanisms. (c) Properties of interest considered in this study.
• Figure 2: Charge density data standardization. (a) Illustration of dimensional variations in the dataset, defined as the magnitudes along the $x$-, $y$-, and $z$-direction in the charge density data, respectively. (b) Conversion of charge density data from a 3D matrix into a 3D visualization, using Ti$_4$O$_8$ as an example. (c) Charge density data before and after standardization along the $z$-direction, using Ti$_4$O$_8$ as an example.
• Figure 3: Image representation of 3D charge density data. (a) 3D charge density grid with $x$- and $y$-scales varying randomly. (b) Conversion of 3D charge density data into sixty 2D images, with the x- and y-scales standardized to sixty; each representing a 2D charge density at a specific $z$ value. (c) Combined 3D visualization reconstructed from the sixty 2D images. (d) Grayscale version of the previous RGB visualization.
• Figure 4: Results obtained from single-task DL prediction. (a) Results of the regression tasks. Left panel: Schematic illustration and key evaluation metrics of the regression task. The red dashed diagonal lines represent the true value obtained from DFT calculations. Middle panel: Scatter plots of two representative results obtained from our DL model -- bulk modulus and volume -- for the test set. The scale bar on the right shows the data density of the predicted values. Right panel: overall prediction performance for nine different target properties; white and light-green backgrounds indicate the results are obtained from non-magnetic and magnetic datasets, respectively. (b) Results of the classification tasks. Left panel: Schematic illustration of the classification task. Middle panel: Predicted ROC curves for electronic band gap and dynamic stability classification on both the validation and test sets; the x- and y-axes reprent the false-positive and true-positive rates, respectively. The optimal threshold $\epsilon_{\rm th}$ and the corresponding area under the ROC curve (AUC) are annotated in the plots. Right panel: Confusion matrices obtained based on the optimal thresholds.
• Figure 5: Evaluation of regression and classification performance across different task grouping strategies in a multi-task learning framework. (a) Different task groupings strategies: random grouping (Random group), similarity-based clustering (Similarity clustering) grouping, and training all tasks together (Entire targets). (b) Bar charts depicting the $R^2$ scores and performance gaps for different tasks (such as $\text{B}$, $\upsilon$, $\text{E}_\text{form}$, $\text{E}_\text{total}$, $\text{E}_\text{gap}$, $\text{E}_\text{fermi}$, $\text{W}_\text{band}$, and $\rho$ across single-task, random grouping, similarity-based clustering grouping, and entire-targets grouping training strategies. (c) ROC curves illustrating the classification performance for gap and gapless categories, highlighting True Positive Rate (TPR) against False Positive Rate (FPR) at various thresholds, with AUC values indicated for both validation and test sets. (d) Confusion matrices under the best optimal threshold, showcasing accuracy based on task groupings for multi-task learning, with detailed breakdowns for actual versus predicted classes, including gap and gapless categories.