Materials research from the perspective of computing: data, model, and more

Опубликовано 19 мая 2022, 14:06

The way people search for a new material is undergoing a transition from "experimental trial and error" to "rational design". While the current computational screening scheme based on existing databases has guided the synthesis of a number of key materials, the intrinsic limitation on chemical and structural space of the databases gradually impedes the further discovery of novel materials with desired functionality. This talk will briefly cover the current research paradigms, commonly used tools, and challenges in computational materials science. On that basis, the effort at Microsoft Research will be introduced, including Graphormer (a Transformer-based molecular modeling foundation model), Local Similarity kernel Optimization (LOSIKO, a data-driven atomic structure optimizer for unlabeled data), MoLeR (a graph-based model for molecule generation), and KD-DTI (a database for literature mining). Future perspectives will be discussed in the end.

Speaker: Ziheng Lu, Microsoft Research Asia
microsoft.com/en-us/research/p...