Discovery of Crystalline Inorganic Solids in the Digital Age

Dmytro Antypov, Andrij Vasylenko, Christopher M. Collins, Luke M. Daniels, George R. Darling, Matthew S. Dyer, John B. Claridge, Matthew J. RosseinskyAccounts 2025 · DOI

TL;DR

This paper presents a computational workflow that combines machine learning and crystal structure prediction to accelerate the discovery of novel inorganic materials. By focusing on hypothesis-driven exploration rather than brute-force screening, this approach has successfully identified new materials, such as the Li-ion electrolyte Li7Si2S7I, and offers a more efficient path to understanding and optimizing material properties.

What Problem Does This Paper Address?

The discovery of new crystalline inorganic materials is crucial for technological advancement, but the sheer vastness of "chemical space" – the combination of all possible compositions and structures – makes exhaustive exploration impossible. Traditional methods often rely on modifying known structures (optimization by analogy), which is less risky but limits the discovery of truly novel materials. The challenge lies in efficiently navigating this immense space to identify and synthesize entirely new structural motifs that can unlock new properties and functionalities. This paper tackles the problem of how to computationally guide experimentalists to promising, unexplored regions of chemical space, moving beyond simple analogy to discover genuinely new materials.

Methods

The authors describe a two-pronged computational workflow:

  1. Identifying promising chemistries: They use unsupervised machine learning (specifically, a variational autoencoder) trained on existing experimental data (from databases like ICSD) to identify "phase fields" (sets of elements) that are likely to yield experimentally isolatable phases. This prioritizes chemistries that are likely to lead to new materials, rather than just variations on known ones.
  2. Targeting compositions within a phase field: Once a promising chemistry is identified, they employ crystal structure prediction (CSP) methods. Instead of aiming for the absolute ground state (which is computationally expensive and often unnecessary for initial discovery), they use "probe structures." These are computationally derived structures that represent accessible stability at a given composition. The energies of these probe structures help map out regions of energetic stability, guiding experimentalists to specific compositions to target for synthesis.

This approach contrasts with broader screening studies by being hypothesis-driven and focusing on identifying novel structural motifs.

Datasets

The primary dataset used for training the machine learning model is the Inorganic Crystal Structure Database (ICSD), which contains experimentally realized crystal structures. The authors mention that this dataset has been expanded to a more comprehensive set. They also refer to a database of experimentally measured lithium solid electrolyte conductivities.

Key Insights and Ideas

  • Discovery vs. Optimization: The paper strongly emphasizes the distinction between discovering entirely new structures and optimizing properties of existing ones through compositional changes. Both are valuable, but the latter relies on the former.
  • Hypothesis-Driven Exploration: The authors advocate for a hypothesis-driven approach, where computational tools support expert-formulated ideas about where to look for new materials, rather than purely data-driven, broad screening.
  • Probe Structures for Guidance: Using "probe structures" in CSP is a practical way to explore stability landscapes without necessarily finding the absolute lowest energy structure, making the process more computationally tractable and directly guiding experimental synthesis.
  • Novelty Beyond Analogy: The discovery of Li7Si2S7I, which has a unique structure not easily predicted by analogy, highlights the importance of exploring beyond known structural families.
  • Context is Key for Novelty: Claims of novelty should be carefully assessed, considering the size of chemical space and the limitations of computational predictions, especially regarding disorder. Expert evaluation remains crucial.

Key Results

  • Discovery of Li7Si2S7I: The workflow led to the discovery of the solid lithium electrolyte Li7Si2S7I, which exhibits a novel structure with 15 distinct Li environments, enabling superionic transport.
  • Enhanced Properties via Substitution: Substitution of Ge for Si in Li7Si2S7I (creating Li7Si2-xGexS7I) further improved its low-temperature conductivity and stabilized Li+ site disorder over a wider temperature range.
  • Successful Workflow Validation: The paper demonstrates the utility of their combined ML and CSP workflow in identifying promising regions of chemical space and leading to experimentally realized new materials.

Limitations and Future Work

  • Computational Limitations: CSP methods are limited by the number of atoms that can be treated computationally, meaning they cannot always perfectly capture large unit cells or complex disorder.
  • Experimental Realization is Paramount: Computational predictions are candidates, not discoveries. Experimental validation is essential.
  • Expert Evaluation: Assessing the true novelty and impact of a discovered material requires expert human judgment.
  • Future Work: The authors suggest bringing together experts from chemistry and computer science to design hypothesis-based discovery routes. They also highlight the need for improved metrics to assess novelty and the potential of quantum computing for CSP. The ongoing development and expansion of datasets for machine learning are also implied.

Summary updated

Abstract

ConspectusThis Account considers how the discovery of crystalline inorganic materials, defined as their experimental realization in the laboratory, can benefit from computation: computational predictions afford candidates for laboratory exploration, not discoveries themselves. The discussion distinguishes between the novelty of a material in terms of its composition and in terms of its structure. The stepwise modification of the composition of a parent material with retention of its crystal structure can reduce the risk in seeking new materials and offers the ability to fine-tune properties which has demonstrated value in optimizing materials performance. However, the parent structures first need to be identified, thus emphasizing the importance of materials discovery beyond simple analogy as a key complementary activity. We describe a workflow we have developed to accelerate discovery of such new structures by addressing many of the challenges, in particular the identification of chemistries that are likely to afford materials and the targeting of reactions within their compositional spaces. Data on experimentally isolated phases are used to prioritise candidate chemistries with machine learning, and crystal structure prediction is used to target compositions within those chemistries for synthesis by computationally constructing probe structures whose energies are indicative of the accessible stability at a given composition. We show how this workflow usefully identifies the parts of chemical space offering new materials and has afforded new structures in practice. The discovery of the solid lithium electrolyte Li<sub>7</sub>Si<sub>2</sub>S<sub>7</sub>I illustrates the role of the workflow in exploring design hypotheses constructed by synthesis researchers and the role of new materials in increasing understanding, in this case by expanding the design paths available for superionic transport. Substitution into Li<sub>7</sub>Si<sub>2</sub>S<sub>7</sub>I affords a structurally related material with superior low temperature transport properties, emphasizing the role of new structures in enabling subsequent materials optimization by compositional modification founded on that structural scaffold.We contrast our focused hypothesis-driven approach with the recent screening studies that cover a much broader range of chemistries and do not target novel structural motifs. These approaches are good at interpolation and identifying the low hanging fruit for substitutional chemistry, but they struggle to deliver new chemistry knowledge, new understanding and new experimentally observed crystal structures. We comment on reporting the large number of proposed hypothetical structures when considering advances in prediction and the importance of context of the size of the chemical space including continuous composition variation and disorder. An example is the difference between predicting superstructures of known parent structures and experimentally realizing these in the face of competition from structural disorder. Given the scope for prediction of candidates, discussion of structural novelty can usefully be restricted to realized experimental examples based on expert interrogation of their structures. We advocate for bringing experts from chemistry and computer science together to design hypothesis-based routes to materials discovery that incorporate appropriate assessment of novelty.
BibTeX
@inproceedings{antypov2025discovery,
  author = {Dmytro Antypov and Andrij Vasylenko and Christopher M. Collins and Luke M. Daniels and George R. Darling and Matthew S. Dyer and John B. Claridge and Matthew J. Rosseinsky},
  title = {{Discovery of Crystalline Inorganic Solids in the Digital Age}},
  booktitle = {Accounts},
  year = {2025},
  doi = {10.1021/acs.accounts.4c00694},
}
Figure