Joint NJIT Physics & Rutgers Physics Seminar

February 13th, Friday (*SPECIAL DAY*)

Machine Learning Across Scales:

Bridging Quantum Many-Body Theory and Large-Scale Materials Modeling

Dr. Bikash Kanungo

Univ. of Michigan, Ann Arbor

(Materials Phys., Host: Rutgers-Newark Physics Dept.)

SPECIAL ROOM at RUTGERS NEWARK CAMPUS:

Dana Room (4th floor of Dana Library) map

(**BRING YOUR PHOTO ID TO ENTER BUILDING**)

SPECIAL TIME: 11:30 am - 12:30 pm (Coffee: 11am)

Density functional theory (DFT) and atomistics have long remained two complementary pillars in computational physics and chemistry. DFT, as an electronic structure method, informs us about electron-electron and electron-ion interactions. Atomistic methods strip away the electrons to describe the dynamics of atoms at length- and time-scales beyond the capabilities of DFT. Together, they account for over 40% of global scientific computing usage. Despite their impressive record, they are severely limited in their accuracy, transferability, and bridging the relevant length- and time-scales. In DFT, approximations to the exchange-correlation (XC) functional—the cornerstone in DFT that describes the quantum-mechanical interactions between electrons—remain far from the desired chemical accuracy of 1 kcal/mol in energies. Plus, the high computational demands of DFT limit their routine usage to length-scales of ∼1000 atoms. In atomistics, the interatomic potentials (IPs), including the machine- learned ones (MLIPs), being fitted to specific datasets suffer from poor transferability across different chemistry. Importantly, their lack of electronic information limits their ability to describe electronically driven phenomena, such as in surface chemistry, emergent behavior in quantum materials, biochemical reactions, to name a few.

In this talk, I will discuss approaches to address these fundamental challenges of accuracy and scale in DFT and atomistics. I will present a machine-learning (ML) framework that bridges quantum many-body theory, DFT, and atomistics through universal, physics-informed maps. I will describe how accurate quantum many-body data can be used to systematically learn exchange–correlation functionals via the inverse DFT problem. I will also introduce field-theoretic atomistics (FTA), as a large- scale machine-learned surrogate to DFT. Unlike IPs/MLIPs, FTA is an electronic structure method, as it yields not only energies and forces but also accurate electron densities and other electronic properties at similar cost as MLIPs. Combined with scalable numerical methods and high-performance computing, these approaches demonstrate how machine learning, when tightly integrated with physical principles, can overcome long-standing accuracy and scale barriers in materials modeling.