Scaling Up Computational Materials Discovery Via Deep Learning
The application of artificial intelligence has started having a substantial impact on the field of materials science. In this presentation, I will discuss the potential of foundation models in materials simulations, particularly highlighting our efforts in the development of GNoME. I will demonstrate how the scaling up of data and computation facilitates the creation of deep learning models capable of achieving unprecedented levels of accuracy in predicting material properties. Additionally, these models possess the ability to generalize to previously unseen simulation tasks. By realizing highly accurate energy predictions, GNoME has enabled us to significantly expand the number of known stable inorganic compounds, representing an order of magnitude increase. I will further present remarkable examples of GNoME's capabilities in addressing complex simulation tasks in materials research, such as predicting ionic conductivities for in silico design of battery materials and crystallization products of amorphous inorganic materials, at scales that were previously impractical using conventional models.
BIOGRAPHY
Dr. Muratahan Aykol is a senior researcher at Google DeepMind, specializing in the intersection of materials science, chemistry, and machine learning. Previously, he led a battery data science team at Rivian Automotive and was a founding member of the Accelerated Materials Design & Discovery division at Toyota Research Institute. He holds a Ph.D. in Computational Materials Science from Northwestern University and was a postdoctoral scholar at the Berkeley Lab. Murat has authored 90+ papers and patents in machine learning, materials discovery, materials synthesis, and energy materials.