Element selection for crystalline inorganic solid discovery guided by unsupervised machine learning of experimentally explored chemistry

Andrij Vasylenko, Jacinthe Gamon, Benjamin B. Duff, Vladimir V. Gusev, Luke M. Daniels, Marco Zanella, J. Felix Shin, Paul M. Sharp, Alexandra Morscher, Ruiyong Chen, Alex R. Neale, Laurence J. Hardwick, John B. Claridge, Frédéric Blanc, Michael W. Gaultois, Matthew S. Dyer, Matthew J. RosseinskyNature Communications 2021 · DOI

TL;DR

This paper introduces a machine learning approach to guide the discovery of new inorganic materials by predicting which combinations of elements are most likely to form stable crystalline compounds. This method, applied to the search for lithium-ion conductors, led to the discovery of a novel material, Li₃.₃SnS₃.₃Cl₀.₇, which exhibits promising ion transport properties.

What Problem Does This Paper Address?

Discovering new materials with desirable properties is a cornerstone of scientific advancement, but the sheer number of possible elemental combinations makes this a daunting task. Researchers typically rely on their intuition and existing knowledge to select promising "phase fields" (sets of elements) for synthesis. However, as the volume of known materials grows, it becomes increasingly difficult for human experts to assimilate this information and identify truly novel, unexplored avenues. This paper tackles the challenge of systematically guiding this initial, crucial step of element selection in materials discovery, moving beyond the analysis of individual materials to focus on the broader elemental combinations that define potential new compounds.

Methods

The core of this work is a Variational Autoencoder (VAE), an unsupervised machine learning model. Here's how it's applied:

  1. Feature Engineering: Each element is described by 37 distinct features (e.g., atomic weight, ionic radius, valence). A phase field (a combination of four elements) is then represented as a 148-dimensional vector (37 features × 4 elements).
  2. VAE Training: The VAE is trained on a dataset of known, synthetically isolable inorganic compounds from the Inorganic Crystal Structure Database (ICSD). The VAE learns to compress these high-dimensional feature vectors into a lower-dimensional "latent space" and then reconstruct them. The "reconstruction error" quantifies how well the VAE can reproduce the original data from its compressed representation.
  3. Phase Field Ranking: For unexplored phase fields, the VAE calculates a reconstruction error. A lower error suggests that the elemental combination is more similar to those that have historically yielded stable compounds, indicating a higher likelihood of synthetic accessibility. This allows for a quantitative ranking of unexplored phase fields.
  4. Collaborative Workflow: The ML model acts as a guide, not a replacement, for human expertise. Researchers use the ranked list to prioritize promising phase fields for further investigation, combining ML predictions with their own chemical intuition and experimental constraints.
  5. Crystal Structure Prediction (CSP): For promising candidate phase fields, computational methods like CSP are used to explore specific compositions and predict potential stable structures.
  6. Experimental Synthesis and Characterization: Promising predicted structures are then synthesized and characterized using techniques like X-ray diffraction (XRD), neutron powder diffraction (NPD), and NMR spectroscopy.

Datasets

  • Inorganic Crystal Structure Database (ICSD): This database of over 200,000 entries of reported crystalline inorganic compounds was used as the primary dataset for training the VAE. The authors focused on quaternary (four-element) compounds with two anions.
  • Newly Discovered Data: The paper reports the synthesis and characterization of a new compound, Li₃.₃SnS₃.₃Cl₀.₇, which was not part of the training data and serves as a validation of the predictive power of the ML model.

Key Insights and Ideas

  • Unsupervised Learning for Element Selection: The VAE effectively learns complex patterns of chemical similarity from existing data without needing explicit labels for "good" or "bad" element combinations. This is crucial because negative data (combinations that don't form stable compounds) is scarce and unreliable.
  • Bridging the Gap Between Elements and Compounds: The model focuses on ranking phase fields (elemental combinations) rather than individual compositions, directly addressing the initial decision-making process in materials discovery.
  • Quantitative Guidance for Intuition: The ML model provides a numerical ranking that complements expert intuition, helping researchers prioritize their efforts and resources more effectively in the vast chemical space.
  • Discovery of a Novel Lithium-Ion Conductor: The approach successfully guided the discovery of Li₃.₃SnS₃.₃Cl₀.₇, a new mixed-anion material with a defect-stuffed wurtzite structure that facilitates efficient lithium-ion transport.
  • Mixed-Anion Materials as a Promising Frontier: The discovery highlights the potential of multi-anion systems, which are often under-represented in research but can offer unique properties and improved stability compared to single-anion counterparts.

Key Results

  • The VAE model achieved a high accuracy in reconstructing known phase fields, with 79.5% having a normalized reconstruction error below 0.5.
  • The model successfully guided the selection of the Li-Sn-S-Cl phase field, which was ranked #5 among unexplored quaternary two-anion phase fields for Li-ion conductors.
  • Experimental synthesis within this phase field led to the discovery of a new compound, Li₃.₃SnS₃.₃Cl₀.₇.
  • This new material exhibits a defect-stuffed wurtzite structure and demonstrates a bulk conductivity of 3.2(3) × 10⁻⁵ S cm⁻¹ at room temperature, with an activation energy of 0.38(3) eV.
  • The material showed good chemical compatibility with lithium metal and stable cycling performance over 400 hours.
  • Analysis revealed a 1D lithium-ion diffusion pathway along the c-axis, enabled by the specific interstitial site occupancy in the defect wurtzite structure.

Limitations and Future Work

  • Focus on Positive Data: The VAE is trained only on positive examples (synthetically isolable compounds). While this is a strength in materials discovery, it means the model ranks based on likelihood of success rather than definitively ruling out failures.
  • Experimental Constraints: The selection of candidate phase fields in Stage II (e.g., avoiding transition metals, toxic elements) is based on human expertise and may limit the scope of discoveries.
  • Future Work: The authors suggest that the VAE model can be extended to other types of materials and that the collaborative workflow can be integrated with other high-throughput experimental methods to further accelerate materials discovery. The discovery of Li₃.₃SnS₃.₃Cl₀.₇ opens avenues for further exploration of mixed-anion wurtzite structures for ion transport applications.

Summary updated

Abstract

The selection of the elements to combine delimits the possible outcomes of synthetic chemistry because it determines the range of compositions and structures, and thus properties, that can arise. For example, in the solid state, the elemental components of a phase field will determine the likelihood of finding a new crystalline material. Researchers make these choices based on their understanding of chemical structure and bonding. Extensive data are available on those element combinations that produce synthetically isolable materials, but it is difficult to assimilate the scale of this information to guide selection from the diversity of potential new chemistries. Here, we show that unsupervised machine learning captures the complex patterns of similarity between element combinations that afford reported crystalline inorganic materials. This model guides prioritisation of quaternary phase fields containing two anions for synthetic exploration to identify lithium solid electrolytes in a collaborative workflow that leads to the discovery of Li<sub>3.3</sub>SnS<sub>3.3</sub>Cl<sub>0.7.</sub> The interstitial site occupancy combination in this defect stuffed wurtzite enables a low-barrier ion transport pathway in hexagonal close-packing.
BibTeX
@inproceedings{vasylenko2021element,
  author = {Andrij Vasylenko and Jacinthe Gamon and Benjamin B. Duff and Vladimir V. Gusev and Luke M. Daniels and Marco Zanella and J. Felix Shin and Paul M. Sharp and Alexandra Morscher and Ruiyong Chen and Alex R. Neale and Laurence J. Hardwick and John B. Claridge and Frédéric Blanc and Michael W. Gaultois and Matthew S. Dyer and Matthew J. Rosseinsky},
  title = {{Element selection for crystalline inorganic solid discovery guided by unsupervised machine learning of experimentally explored chemistry}},
  booktitle = {Nature Communications},
  year = {2021},
  doi = {10.1038/s41467-021-25343-7},
}
Figure