Romit Chakraborty, Ph.D. 

Quantum Chemical Predictions for Clean Energy Materials

Welcome to my humble internet abode!

I leverage the basic tenets of quantum chemistry to conduct simulations of clean energy materials with a broad view toward developing efficient energy storage and conversion systems. Central to these pursuits are the core principles of quantum mechanics, whose application enables the prediction of the physical and chemical characteristics of materials needed for the realization of a sustainable energy infrastructure. My scholarly pursuits are informed by the foundations of Density Functional Theory (DFT) and Reduced Density Matrix (RDM) theory, both of which furnish succinct descriptions of the Hilbert Space for wave functions that convey information about the electronic and vibrational degrees of freedom. DFT, a workhouse for practical materials simulations, is an effective one-electron theory that elegantly side-steps the well-known N-representability problem that applies to the treatment of many electron quantum systems. For a formal and concise outline of effective one-body constraints that ensure Pure N-representability, which Generalise the well-known Pauli Conditions as they apply to some chemical and biological systems, please peruse my dissertation on this topic.

Following  this  journey of abstractions  that many would consider tangential  to formal electronic structure theory, I worked with Prof. Martin Head-Gordon at Berkeley where I assisted in  the rational design efforts for clean energy materials: metal-organic  frameworks for gas storage and separations. This gave me a broad perspective on practical electronic structure theory and simulations. Despite improvements,  the computational complexity of the CI (Configuration Interaction) and CC (Coupled Cluster) Hamiltonians entails that only a handful of electrons, about ten on your laptop, and about twenty if you have the luxury of high-performance compute, can be treated accurately in molecular systems. This barrier is not likely to be overcome by quantum hardware of any architecture: superconducting qubits, trapped ions, atoms, or silicon photonics in the next two years. Density Matrix Renormalization Group (DMRG) remains the best CI solver, and Density Functional Theory (DFT) remains the workhorse for practical computations for material systems, and is likely to remain so in the foreseeable future. 

As an AI optimist, I am enthused by the prospect of using tranformer-based architectures to learn more about the physical world. I enjoy building extensions to solutions in the natural sciences by programming multimodal large language models. Please find below some examples of applications and services derived from programming multimodal large language models. These are relatively simple augmentations of the underlying llms (GPT-4o for instance) that i) employ low rank adaptations of a large model by fine-tuning it with the QM9 dataset, and ii) train a large vision model (GPT-4o) to recognize molecular orbitals via prompt engineering.

Some examples of ongoing work include
i) Agentic Workflows that partition the Hilbert Space into interacting and non-interacting subsystems,
ii) LoRA fine tuning of Multimodal Large Language models for Orbital Recognition,
iii) Models for competitive adsorption in MOFs with open metal sites, and
iv) Accurate Depictions for Anharmonicity in Small Molecule Binding to Open Metal sites in MOFs.

Find more info about peer reviewed and archived articles on my Google Scholar.

The following are simple demonstrations of ongoing work, some of which is in process, and some of it proprietary due to my affiliations with PsiQuantum.

I am actively seeking research and development opportunities that lie at the intersection of quantum chemistry, materials science, quantum and classical information theory, and that can leverage my skills in programming multimodal large language models.

AI usage in quantum chemistry applications. A minimal full stack application that uses a GPT-4o model fine-tuned on a subset of QM9. Details in the github repo.
On the right GPT-4o is learning to classify atomic and molecular orbitals based on their symmetry, occupations and bonding attributes via few shot prompting. *model may make mistakes

You may find more about my work on my google scholar page which showcases my research.  Outside of work you can find me playing squash or tuning my guitar.