Revealing the Spectrum of Unknown Layered Materials with Super-human Predictive Abilities

< back to research overview

It is a long-standing but as yet unrealized dream of computational material science to elucidate not just the properties of some materials but list all possible materials with a particular property. Here we accomplish this for the first time for two-dimensional layered materials by combining physics with machine learning, and propose a paradigm shift in materials discovery. In this article, we build a machine learning model that can identify every layered material among all possible combinations of binary or ternary compositions unlimited to a particular stoichiometry or a group of elements. Recent data-mining efforts suggest there are over 1000 layered materials that have been synthesized and reported in the literature, but this does not reveal the full extent of these important materials that are possible to synthesize.

In contrast to previous works on layered materials discovery, our model is effective in predicting materials that have yet to be discovered. We discover approximately 1500 additional layered binary and ternary crystalline 2D materials, indicating that civilization to date has synthesized approximately half of all possible layered materials. To our knowledge, such a claim has never been made regarding any materials class.

We further perform density functional theory calculations on 13 of the 1500 candidate layered materials and find that 10 of these materials are mechanically stable layered materials. We find that two of these materials have the potential to exist in two different crystal structures with different band gaps. While phase change materials are crucial to some important engineering applications and two-dimensional materials can offer advantages in size and energy consumption, the number of known potential 2D phase change materials are very limited. We expand the short list of known 2D phase change materials, and see promise in finding more among the 1500 new materials.

We also address in this work one of the biggest challenges in the adoption of machine learning in the physical science: conveying the utility of the model in a fashion that non-expert researchers can immediately appreciate. By pitting our model against over 30 expert humans, we find that the model surpasses the performance of the expert researchers in the field in identifying a layered material based on chemical formula. We also of course find that our model is orders of magnitude faster.

We employ semi-supervised learning techniques for the first time in materials discovery. Semi-supervised learning is a technique in that has proven to be successful in many machine learning applications. It utilizes unlabeled data in addition to labeled data, which is powerful in cases where labels are expensive to obtain or are noisy. We find that semi-supervised learning provides benefits over supervised learning in identifying layered materials. In the field of materials physics, labeled data can be scarce, such as rare materials known to possess certain properties; they can also be noisy, such as property measurements with large errors. Semi-supervised learning may be applicable to a wide range of problems in physics and materials science.


Contact: Gowoon Cheon, gcheon_at_stanford.edu

Publication:

Cheon. G, Cubuk, E.D., Antoniuk, E.R., Blumberg, L., Goldberger, J.E., and Reed, E.J. Revealing the Spectrum of Unknown Layered Materials with Superhuman Predictive Abilities Journal of Physical Chemistry Letters 9(24), 6967-6972 (2018). doi: 10.1021/acs.jpclett.8b03187