An Invertible Crystallographic Representation for <b>General</b> Inverse Design of Inorganic Crystals with Targeted Properties

Zekun Ren, Siyu Tian, Juhwan Noh, Felipe Oviedo, Guangzong Xing, Qiaohao Liang, Ruiming Zhu, Armin G. Aberle, Shijing Sun, Xiaonan Wang, Yi Liu, Qianxiao Li, J. Senthilnath, Kedar Hippalgaonkar, Yousung Jung, Tonio BuonassisiSSRN 2021 · DOI

TL;DR

Researchers have developed a new framework, FTCP, that can design novel inorganic crystals with specific properties, going beyond limitations of previous methods that were restricted to certain elements or structures. This advancement promises to accelerate the discovery of new materials by enabling "general inverse design" – predicting a material's structure based on desired properties.

What Problem Does This Paper Address?

Discovering new materials with specific, user-defined properties is a long-standing goal in materials science. Traditionally, this involved intuition-driven heuristics or extensive computational screening. While machine learning has shown promise, existing generative models for inverse design of crystalline solids are often limited to specific compositions (e.g., only oxides) or crystal structures (e.g., only cubic). This severely restricts their ability to explore the vast chemical and structural space of inorganic materials. The authors aim to overcome this by developing a "general inverse design" framework that can predict both the chemistry and structure of a material based on desired properties, without prior constraints on elements or crystal structures.

Methods

The core of the FTCP framework lies in its novel invertible crystallographic representation. This representation encodes crystals using both:

  • Real-space features: These are akin to Crystallographic Information File (CIF) data, including element types, lattice parameters, atomic positions, and site occupancies. This ensures the representation is invertible, meaning a crystal can be reconstructed from it.
  • Reciprocal-space features: These capture crystal properties by projecting elemental descriptors into the frequency domain, similar to Fourier transforms used in X-ray diffraction. This featurizer is inspired by diffraction fingerprinting and is designed to be compact and informative.

This representation is then fed into a Variational Autoencoder (VAE). The VAE is augmented with a target-learning branch that maps latent space representations to user-specified properties. This "property-structured latent space" organizes the latent space such that similar properties are clustered together, enabling targeted sampling for inverse design. The framework involves a four-stage workflow: defining target properties, sampling from the trained model, structural relaxation (using DFT), and first-principles validation of properties.

Datasets

  • Materials Project database: Used for training and testing the model. For specific design cases, subsets were selected based on criteria like energy above hull (<0.08 eV/atom) and number of sites (e.g., ≤20 or ≤40).
  • Reference database (for power factor): A database from a specific reference was used for thermoelectric (TE) power factor calculations, as this property is not as comprehensively available in Materials Project.

Key Insights and Ideas

  • General Inverse Design is Achievable: The FTCP framework demonstrates that it's possible to design novel inorganic crystals with user-defined properties across a wide range of chemistries and structures, breaking free from previous limitations.
  • The Power of Combined Representations: The combination of real-space (guaranteeing invertibility) and reciprocal-space (capturing crystal physics) features is crucial for the framework's success.
  • Property-Structured Latent Space: Organizing the latent space based on desired properties is key to efficiently sampling for materials with specific characteristics. This allows for a "property gradient" that guides the design process.
  • Excited-State Properties are Accessible: The framework successfully designed materials with targeted thermoelectric power factors, an excited-state property that is challenging to compute and predict. This was achieved through semi-supervised learning.
  • Uniqueness is Key: The designed crystals are unique, meaning they do not exist in the training database, highlighting the generative power of the model.

Key Results

  • High Success Rates: The FTCP framework achieved success rates ranging from 7.1% to 38.9% across three design cases.
  • Significant Improvement Over Random: The framework demonstrated substantial improvement over random chance, with success rates showing 38.8% to 560% higher than random sampling.
  • Design Cases:
    • Formation Energy (Ef): Achieved success rates between 14.3% and 38.9% for various target Ef values.
  • Bandgap (Eg) with Ef Constraint: Achieved a success rate of 36.8% for a target Eg of 1.5 eV and Ef < -1.5 eV/atom.
  • Thermoelectric Power Factor: Designed two unique crystals with power factors comparable to state-of-the-art materials like GeTe, achieving a 7.1% success rate.
  • Structural Uniqueness: The designed crystals exhibited a median dissimilarity value of 0.37 to 0.67 compared to existing crystals, indicating significant structural novelty.

Limitations and Future Work

  • Synthesizability Challenge: While the framework designs promising materials, experimental validation and synthesis remain a significant hurdle. The authors explored a naive "ICSD score" as a proxy for synthesizability, suggesting future work could integrate more sophisticated synthesizability metrics.
  • Invariance: The current FTCP representation does not fully satisfy all crystallographic invariances (e.g., rotational, translational). While this is an open research area, the authors acknowledge its potential impact on property prediction.
  • Computational Cost: DFT calculations for validation are still computationally expensive, limiting the throughput of the design process. Future work could involve ML surrogate models for faster structural relaxation.
  • Hyperparameter Tuning: The authors note that optimal hyperparameter tuning might be dataset-dependent, encouraging users to perform their own tuning.
  • Error Sources: The authors identified decompression and interpolation errors in the latent space as sources of error affecting validity rates.

Summary updated

BibTeX
@misc{ren2021invertible,
  author = {Zekun Ren and Siyu Tian and Juhwan Noh and Felipe Oviedo and Guangzong Xing and Qiaohao Liang and Ruiming Zhu and Armin G. Aberle and Shijing Sun and Xiaonan Wang and Yi Liu and Qianxiao Li and J. Senthilnath and Kedar Hippalgaonkar and Yousung Jung and Tonio Buonassisi},
  title = {{An Invertible Crystallographic Representation for \&lt;b\&gt;General\&lt;/b\&gt; Inverse Design of Inorganic Crystals with Targeted Properties}},
  year = {2021},
  doi = {10.2139/ssrn.3862821},
  howpublished = {SSRN},
}
Figure