Drawing a materials map with an autoencoder for lithium ionic conductors

Yudai Yamaguchi, Taruto Atsumi, Kenta Kanamori, Naoto Tanibata, Hayami Takeda, Masanobu Nakayama, Masayuki Karasuyama, Ichiro TakeuchiScientific 2023 · DOI

TL;DR

This paper introduces a novel "materials map" generated using an autoencoder to visualize the vast landscape of known lithium-ion conductive materials. By compressing complex compositional and structural data into a 2D space, the map aids researchers in intuitively identifying promising new materials and unexplored regions, accelerating the discovery of advanced battery components.

What Problem Does This Paper Address?

The discovery and optimization of new functional materials, particularly for applications like advanced batteries, traditionally rely heavily on the intuition and experience of specialized researchers. While materials informatics (MI) has emerged as a powerful tool to accelerate this process by integrating data and machine learning, it faces challenges in systematically exploring the immense space of undiscovered materials. Existing MI methods often struggle to predict properties of compounds not present in current databases. This paper aims to bridge this gap by creating a visual tool that combines the power of MI with human intuition, making the search for novel materials more efficient and intuitive.

Methods

The core of this work is the development of a materials map using an autoencoder. This neural network architecture is employed to reduce the dimensionality of complex compositional and structural descriptors of materials into a single dimension each. These two encoded dimensions then form the axes of the 2D materials map.

  • Autoencoder for Dimensionality Reduction: The authors used separate autoencoders for compositional (1782 dimensions) and structural (858 dimensions) descriptors. The goal was to compress these high-dimensional inputs into a single, meaningful encoded value ($X_{comp}$ and $X_{struct}$) while retaining as much information as possible.
  • ME Regression Integration: An additional neural network layer was incorporated to learn the relationship between the encoded compositional and structural variables and the material's migration energy (ME), a key indicator of lithium-ion conductivity.
  • Loss Function: A combined loss function was used, balancing the reconstruction accuracy of the original descriptors ($Loss_{comp}$, $Loss_{struct}$) with the accuracy of ME prediction ($Loss_{ME}$). A weight factor ($W$) allowed tuning the trade-off between these objectives.
  • Grid Search and Optimization: Hyperparameters for the autoencoder and regression model were optimized using a grid search, and the Adam optimizer was employed for training.

Datasets

The study utilized a dataset of 708 lithium oxide materials, previously reported by the authors.

  • Source: Materials Project database.
  • Selection Criteria: Compounds containing Li and O, and incorporating one of 25 specified additional elements.
  • Properties Calculated: Migration energy (ME) for lithium-ion diffusion was calculated using the Bond Valence Force Field (BVFF) method.
  • Descriptors:
    • Compositional: 1782 dimensions, including chemical composition ratios and various elemental properties (atomic number, weight, electronegativity, bond radii, electron counts, etc.).
  • Structural: 858 dimensions, derived from Radius Distribution Function (RDF), Angular Distribution Function (ADF), and Voronoi diagrams.

Key Insights and Ideas

  • Visualizing the Unknown: The materials map provides a powerful visual tool that goes beyond simple data tables. It allows researchers to see "empty" regions, representing unexplored material compositions and structures, which are prime candidates for new discoveries.
  • Intuition Meets Data: By mapping materials based on composition and structure, the tool leverages human intuition. Researchers can visually scan the map, identify clusters of known materials, and infer properties of nearby, less-explored regions.
  • ME as a Guiding Color: The color-coding of the map by predicted migration energy (ME) directly highlights areas likely to contain materials with high ionic conductivity, guiding the search for better battery electrolytes.
  • Structure vs. Composition: The study reveals that while both composition and structure are important, the structural axis of the map is particularly sensitive to local atomic arrangements, suggesting that fine-tuning local coordination environments is crucial for optimizing ionic conductivity.
  • Trade-off in Optimization: The ability to tune the weight factor ($W$) in the loss function demonstrates a clear trade-off between accurately reconstructing material descriptors and accurately predicting migration energy, allowing researchers to prioritize based on their specific goals.

Key Results

  • Descriptor Reconstruction: The autoencoder achieved high reconstruction accuracy for both compositional and structural descriptors, with R² scores exceeding 0.8 for both.
  • ME Prediction: The ME prediction performance, while not as high as descriptor reconstruction (R² = 0.29), still showed a general trend that matched measured values and provided a useful basis for the materials map.
  • Materials Map Visualization: The generated 2D materials map successfully clustered known material types (e.g., Garnet, NASICON, Perovskite) and revealed regions with high predicted ionic conductivity (low ME), often corresponding to areas with sparse data points, indicating potential for new discoveries.
  • Correlation with Li/O Ratio: Analysis of the map indicated that high lithium concentration and low oxygen concentration, particularly on the structural axis, are correlated with higher ionic conductivity.

Limitations and Future Work

  • Structure Reconstruction Difficulty: The paper acknowledges that while descriptors can be decoded from the map, directly reconstructing the crystal structure from these encoded values remains technically challenging due to the irreversibility of the descriptor-to-structure mapping.
  • ME Prediction Accuracy: The authors note that the predictive performance for ME is not perfect, and in some areas, predicted values may not precisely match measured values.
  • Future Work: The authors suggest that integrating researcher experience with the materials map is crucial for overcoming the limitations of direct structure reconstruction. They anticipate that the 2D visualization will inspire researchers to explore specific regions of the map, leading to the discovery of new materials with optimized ionic conductivity. Further work could focus on improving the accuracy of ME prediction and developing methods for more direct structure prediction from the encoded space.

Summary updated

Abstract

Efforts to optimize known materials and enhance their performance are ongoing, driven by the advancements resulting from the discovery of novel functional materials. Traditionally, the search for and optimization of functional materials has relied on the experience and intuition of specialized researchers. However, materials informatics (MI), which integrates materials data and machine learning, has frequently been used to realize systematic and efficient materials exploration without depending on manual tasks. Nonetheless, the discovery of new materials using MI remains challenging. In this study, we propose a method for the discovery of materials outside the scope of existing databases by combining MI with the experience and intuition of researchers. Specifically, we designed a two-dimensional map that plots known materials data based on their composition and structure, facilitating researchers' intuitive search for new materials. The materials map was implemented using an autoencoder-based neural network. We focused on the conductivity of 708 lithium oxide materials and considered the correlation with migration energy (ME), an index of lithium-ion conductivity. The distribution of existing data reflected in the materials map can contribute to the development of new lithium-ion conductive materials by enhancing the experience and intuition of material researchers.
BibTeX
@inproceedings{yamaguchi2023drawing,
  author = {Yudai Yamaguchi and Taruto Atsumi and Kenta Kanamori and Naoto Tanibata and Hayami Takeda and Masanobu Nakayama and Masayuki Karasuyama and Ichiro Takeuchi},
  title = {{Drawing a materials map with an autoencoder for lithium ionic conductors}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-43921-1},
}
Figure