A Machine Learning Framework for Knowledge Discovery
by Prof Abhishek Singh, IISC Bangalore, India
Date: 21/11/2023 |Time: 11:00 AM to 12:00 PM | Venue: K-G14-256, Webster 256
Bio: Prof. Abhishek K. Singh is currently a Professor in the Materials Research Centre at IISc, Bangalore. He did his PhD from the Institute of Materials Research, Tohoku University, Japan. He was a JSPS Postdoctoral fellow. He has also worked as a postdoctoral research associate at University of California Santa Barbara, and Rice University, Houston, USA. His group IISC Bangalore is leading an effort in designing materials for target applications using data-driven methods. His group has established India’s first computational materials database aNANt. He is currently leading the materials informatics initiative of IISc (MI3).
Prof. Singh has published ~190 papers. His work has received more than 10,000 citations, and his current h-index is 55. He is a recipient of the Materials Research Society of India medal in 2014, the distinguished lectureship award of the Chemical Society of Japan in 2020, JSPS invitation fellowship in 2020. He is the current Chair of the Office of International Relations IISc.
Abstract : Data driven machine learning methods in materials science are emerging as one of the promising tools for expanding the discovery domain of materials to unravel useful knowledge. In this talk, the power of these methods will be illustrated by covering two major aspects, namely, development of prediction models and establishment of hidden connections. For the first aspect, we have developed accurate prediction models for various computationally expensive physical properties such as band gap, band edges and lattice thermal conductivity. The prediction model for band gap and band edges are developed on 2D family of materials -MXene, which are very promising for a wide range of electronic to energy applications, which rely on accurate estimation of band gap and band edges. These models are developed with GW level accuracy and hence can accelerate the screening of desired materials by estimating the band gaps and band edges in a matter of minutes. For the lattice thermal conductivity prediction model, an exhaustive database of bulk materials is prepared. By employing the high-throughput approach, several ultra-low and ultra-high lattice thermal conductivity compounds are predicted. The property map is generated from the high-throughput approach, and four simple features directly related to the physics of lattice thermal conductivity are proposed. The performance of the model is far superior than the physics-based Slack model, highlighting the simplicity and power of the proposed machine learning models. For the second aspect, we have connected the otherwise independent electronic and thermal transport properties. The role of bonding attributes in establishing this relationship is unraveled by machine learning. An accurate machine learning model for thermal transport properties is proposed, where electronic transport and bonding characteristics are employed as descriptors. I will discuss the application of ML in establishing the complex structure-property relations in alloys.