Electron transfer rules of minerals under pressure informed by machine learning

Yanzhang Li, Hongyu Wang, Huan Ye, Yanan Zhang, Rongzhang Yin, Haoning Jia, Bingxu Hou, Changqiu Wang, Hongrui Ding, Xiangzhi Bai, Anhuai LuNature Communications 2023 · DOI

TL;DR

This paper introduces a deep learning model to predict element electronegativity and mineral work functions under extreme pressures, revealing new rules for electron transfer in Earth's deep interior. These insights explain observed high-conductivity anomalies and predict novel redox reactions, offering a powerful new tool for understanding planetary interiors.

What Problem Does This Paper Address?

Understanding electron transfer is fundamental to chemistry, physics, and geoscience, but existing rules for predicting it break down under extreme conditions like those found deep within the Earth. High pressures can dramatically alter electronic behavior, leading to phenomena like insulator-metal transitions and unusual reactivity. Traditional experimental and computational methods struggle to capture these effects due to the immense computational burden and experimental challenges at high pressures. This leaves a critical gap in our understanding of elemental and mineral behavior in deep planetary interiors, hindering our ability to explain phenomena like Earth's high-conductivity anomalies and predict deep geochemical processes.

Methods

The researchers developed a two-pronged approach:

  1. Deep Learning for Electronegativity Prediction: They trained a Dense Attention Network (DAN) model on a dataset of element electronegativities calculated via Density Functional Theory (DFT) at four specific pressure points (0, 50, 200, and 500 GPa). To overcome the challenge of a relatively small dataset (384 points), they employed feature augmentation (using operations like square, log, reciprocal) and symbolic transformers to create a richer feature set. The DAN model was designed with dense connections and self-attention mechanisms to improve performance and prevent overfitting.
  2. Symbolic Regression for Unified Formula: Once the DAN model could predict electronegativity across a continuous pressure range, symbolic regression was used to derive a unified, analytical formula. This formula captures the relationship between electronegativity, pressure, and an element's electronic configuration. The elements were grouped into blocks (s, p, d, f, etc.) based on their valence electrons, and separate formulas were fitted for each block, ultimately leading to a general expression.

This combined approach allowed them to predict not only element electronegativity but also the relative work function of minerals, which is crucial for understanding electron transfer in condensed matter.

Datasets

  • Element Electronegativity Dataset: This dataset, comprising electronegativity values for 96 elements at 0, 50, 200, and 500 GPa, was derived from prior DFT calculations (Dong et al.). This dataset served as the training and evaluation data for the deep learning model.
  • Mineral Work Function Data: The model's predictions for mineral work functions were then used to analyze a dataset of 5828 minerals, drawing compositions and classifications from the Mindat database.

Key Insights and Ideas

  • Pressure-Induced Electronegativity Decrease: A fundamental finding is that element electronegativity generally decreases with increasing pressure. This is attributed to the repulsive forces between nuclei under compression, making it easier to remove electrons.
  • Unified Formula for Electronegativity: The development of a single, analytical formula that describes electronegativity as a function of both pressure and electronic configuration is a significant breakthrough. This formula provides a generalizable and interpretable way to understand elemental behavior under pressure.
  • Work Function as a Proxy for Electron Transfer: The paper demonstrates that the pressure-modulated relative work function (PR-WF) of minerals is a powerful indicator of electron transfer tendency. A decreasing PR-WF with pressure suggests increased electron delocalization and a more reduced state for minerals.
  • Explaining Earth's Conductivity Anomalies: The model's predictions of decreasing PR-WF with depth and sharp drops at phase interfaces (like the 660 km discontinuity) directly correlate with observed high-conductivity anomalies in Earth's interior, providing a physical explanation for these phenomena.
  • Predicting Novel Redox Reactivity: The framework allows for predictions of redox reactions, such as the potential for Fe(II)-bearing minerals to produce H₂ with water under high pressure, offering new insights into deep Earth geochemistry and potential H₂ production pathways.

Key Results

  • High Prediction Accuracy: The deep learning model achieved high accuracy in predicting element electronegativity, with an R² of 0.987 and an MAE of 0.283 on the testing set.
  • Pressure-Dependent Work Function: The study found that mineral work functions decrease significantly (up to ~4 eV) with increasing pressure, particularly for metal alloys.
  • Correlation with Conductivity: The predicted PR-WF values showed a strong correlation with observed electrical conductivity profiles in Earth's interior, particularly at phase boundaries. For instance, a ~1.2 eV PR-WF difference at the majorite-ferropericlase interface correlated with a three-fold increase in conductivity.
  • Redox Reactivity Predictions: The paper predicts that olivine and its high-pressure polymorphs could exhibit high reactivity with H₂O, leading to H₂ production up to the 660 km discontinuity.

Limitations and Future Work

  • Reference System Dependence: The authors acknowledge that the predicted electronegativity and work function values are based on a specific reference system (homogeneous electron gas model with enthalpy) and may differ from other systems, though the relative trends should remain comparable.
  • Temperature Effects: While the study focuses on pressure, the authors note that temperature also significantly influences chemical reactions and material properties, suggesting future work should incorporate its effects.
  • Experimental Validation: While the predictions align with existing geophysical observations, direct experimental validation of electronegativity and work function under extreme pressures remains challenging.
  • Broader Applications: The authors suggest that the developed framework for predicting pressure- and configuration-dependent electronegativity could be extended to understand other thermodynamic properties and applied to fields like batteries, electrocatalysis, and sensors.

Summary updated

Abstract

Electron transfer is the most elementary process in nature, but the existing electron transfer rules are seldom applied to high-pressure situations, such as in the deep Earth. Here we show a deep learning model to obtain the electronegativity of 96 elements under arbitrary pressure, and a regressed unified formula to quantify its relationship with pressure and electronic configuration. The relative work function of minerals is further predicted by electronegativity, presenting a decreasing trend with pressure because of pressure-induced electron delocalization. Using the work function as the case study of electronegativity, it reveals that the driving force behind directional electron transfer results from the enlarged work function difference between compounds with pressure. This well explains the deep high-conductivity anomalies, and helps discover the redox reactivity between widespread Fe(II)-bearing minerals and water during ongoing subduction. Our results give an insight into the fundamental physicochemical properties of elements and their compounds under pressure.
BibTeX
@inproceedings{li2023electron,
  author = {Yanzhang Li and Hongyu Wang and Yan Li and Huan Ye and Yanan Zhang and Rongzhang Yin and Haoning Jia and Bingxu Hou and Changqiu Wang and Hongrui Ding and Xiangzhi Bai and Anhuai Lu},
  title = {{Electron transfer rules of minerals under pressure informed by machine learning}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-37384-1},
}
Figure