Speaker
Description
The discovery of new functional materials is one of the holy grails of computational chemistry, among other reasons because it has the potential to accelerate the adoption of renewable energy sources and reduce the energy consumption of chemical industry. Unfortunately, computational materials discovery is a daunting task, due to the enormous space of possible candidates and the challenges associated with accurately predicting the functionality, stability and synthesizability of even a single material.
Currently, the most common approach used to this end is the virtual screening of materials databases according to simple bulk properties (so called descriptors), which are expected to correlate with the real functional properties of the material. While this has produced some important insights, it is ultimately inefficient (because many unsuitable materials need to be screened), biased (due to the use of predefined databases) and only of qualitative accuracy (due to the limitations of the used descriptors). More sophisticated approaches are thus clearly warranted. To address this demand, we have explored several routes in recent years.
On one hand, we recently proposed the use of formal generative grammars for materials discovery [1]. Such grammars can in principle prune the search space of potential materials by enforcing certain rules that are either based on physical insights or derived from databases of known materials. This ensures that all candidates have reasonable chemical compositions, without overly constraining the search space to already known materials. Beyond this, the use of grammars opens the door to tackle materials discovery with methods of natural language processing. This has already been found to be a powerful approach for designing organic molecules, where grammars are more commonplace.
On the other hand, we have assessed the capabilities of deep generative models for materials discovery [2]. Generative models learn building rules or probability distributions that underlie a given dataset and are thus able to extend it with similar materials. Importantly, these models also allow for the conditional generation of data, leading to the focused discovery of materials with desired properties. In our work, we compared several different generative approaches (Variational Autoencoders, Generative Adversarial Networks and Reinforcement Learning), quantitatively comparing the precision and diversity of the proposed materials. We find all approaches to be highly promising, though the question of how to effectively represent realistic materials remains open.
[1] J.T. Margraf, Z. Ulissi, Y. Jung, and K. Reuter, J. Phys. Chem. C 126, 2931 (2022).
[2] H. Türk, E. Landini, C. Kunkel, J.T. Margraf, and K. Reuter, DOI:10.26434/chemrxiv-2022-glxvs (2022).
| Abstract Number (department-wise) | TH 06 |
|---|---|
| Department | TH (Reuter) |
