Events Calendar

MFD Distinguished Lecture Series: Dr. Albert Musaelian

Speaker: Dr. Albert Musaelian, Harvard University

Talk Title: Designing Neural Network Architectures for Effective Scientific Computing

Abstract: Machine learning models hold significant potential to accelerate and enhance scientific computing. A prominent example is the development of machine learning interatomic potentials (MLIPs), which address the trade-offs between accuracy and computational cost in atomic-scale simulations of chemical and material systems. These models have been successfully applied to simulations of systems ranging from batteries to pharmaceuticals—simulations that would have been infeasible without machine-learning techniques.
This talk will cover the background of MLIPs and their machine learning aspects, with a focus on the “E(3)-equivariant” neural network MLIP architectures, NequIP and Allegro, developed to exploit the symmetries inherent in the physical problems. The presentation will explore their architecture, the design process, and the relationship between network architecture, domain science, and practical engineering, which together enable new capabilities for downstream scientific applications.

Biography: Albert Musaelian researches novel neural network architectures that can improve atomic-scale simulations in computational chemistry and materials science, in particular leading significant work on the NequIP and Allegro architectures and the software frameworks underlying them. He completed his PhD at Harvard University in the Materials Intelligence Research (MIR) group under the guidance of Prof. Boris Kozinsky and with the support of the DOE Computational Science Graduate Fellowship (CSGF).

Host: Dr. Paulo Branicio & Dr. Ken-ichi Nomura

Location: James H. Zumberge Hall Of Science (ZHS) - 352

Audiences: Everyone Is Invited

Contact: Candy Escobedo