I am an incoming EECS PhD student at MIT where I will be working with Professor Priya Donti. I am currently a Master's student at Cornell University (Cornell Tech) studying Electrical and Computer Engineering. My research interests lie in AI for Science.
Hello! I'm Anvita Bhagavathula, a Master’s student at Cornell University (Cornell Tech) studying Electrical and Computer Engineering. I recently graduated from Brown University with a degree in Physics (Sc.B) and Applied Mathematics (A.B).
I am interested in key questions such as how to best represent scientific data and how to modify model architectures by leveraging physical laws or novel learning strategies to improve their applicability to problems in the natural sciences. I have had several research experiences exploring these questions spanning materials science simulation development and scientific machine learning research in sustainable foods, drug-discovery, and mechanics. My research has a track record of being used to solve tangible problems in industry, academic, and startup settings.
I am passionate about incorporating physical inductive biases and constraints into deep learning workflows to improve their scientific applicability. This summer, I worked as a deep learning research intern at Aqemia, a Paris-based drug-discovery startup, where I designed a unique attention-based graph neural network that predicted likelihood of reaction synthesis of SMARTS reaction strings to accelerate the rate at which drug candidates are optimized. I was also a part of the The Crunch Group at Brown University where I investigated neural networks as function approximators to solve differential equations by integrating governing physical laws into their architectures. In addition to my experience with physically-motivated deep learning research, I have extensive experience using Density Functional Theory (DFT), an ab-initio quantum mechanical simulation technique, which I cultivated while studying the superconducting phase in 2D graphene systems as part of the Low-Dimensional Electronics Lab at Brown.
In the past, I have also worked with the Research for Industry group at Microsoft Research where I developed a multimodal machine-learning based approach to predict food protein digestibility to accelerate the development of alternative food proteins. More details about my research projects can be found here.
In my free time, I love to create recipes for my friends, cook, play the bass, and sing. I also love film photography and you can find some of my photos here!
Email: akb249@cornell.edu
[5th Mar 2024] Committed to MIT's EECS PhD program and will be joining in Fall 2024.
[19th Dec 2023] Preprint summarizing findings from Microsoft Research internship is under review at npj Science of Food.
[21st Aug 2023] Started my Master's in Electrical and Computer Engineering at Cornell Tech supported by a merit-based scholarship.
[5th Jun 2023] Began research internship with Aqemia, a French startup leveraging machine learning and quantum physics to accelerate the drug discovery process.
[27th May 2023] Graduated from Brown with degree in Physics and Applied Mathematics. My thesis recieved departmental honors and can be viewed here.
[1st Nov 2022] New preprint summarising findings from Microsoft Research internship available on arxiv.
[31st Oct 2022] Started new research project with The Crunch Group at Brown researching self-adaptive physics-informed neural networks.
[26th Aug 2022] Finished my internship with Microsoft Research and filed a provisional patent for our proposed methodology to predict protein digetibility.
[6th June 2022] Started my internship with Microsoft Research where I am working with Dr. Sara Malvar and Dr. Ranveer Chandra with the Research for Industry Group.
Reaction Filtering with Graph Neural Networks
Physics-Informed Neural Networks
Supervised by Jacques Boitreaud and Dr. Antoine Brochard, I undertook a project on automating the time-intensive process of screening synthetic feasibility of drug candidates to help the medicinal chemists in the company. The fundamental challenge with this project was to create an architecture that was robust enough to generalize to in-house testing data, which had a distribution shift compared to the USPTO-MIT training data. To address this problem, I wanted to create a context-informed and chemically-constrained model. I formulated and implemented a unique attention-based architecture that used reactant and product graphs with encoded reaction site masks, generated using substructure matching, to do so. By the end of the summer, my model was performing with a false positive rate less than 0.1, filtering out almost all of the same reactions as Dr. Nicolas George, the in-house medicinal chemist. The most rewarding part of this project was seeing the tangible impact my research could have, as my model ultimately accelerated the rate at which promising molecules were screened from hours to seconds.
Advised by Dr. Somdatta Goswami and collaborating with two other undergraduate students, I created a PINN to solve the 1D heat equation using a custom loss function that enforced the structure of the solution and its boundary condition behaviour. Optimizing this model’s performance involved investigating loss convergence, identifying poor learning of boundary conditions, and implementing adaptive regularization. This resulted in a reduced boundary loss by six orders of magnitude. Next, we shifted our focus to handling the increased computational cost of the 2D case. We explored low-rank decomposition using forward-mode auto differentiation to separate the PINN on a per-axis basis as a strategy to do so. The repository for this project is available here.
For my Honors Senior Thesis (available here), I worked at the interface of experiment and theory by applying ab-initio quantum mechanical simulations to investigate the mechanism behind the superconducting phase in 2D graphene systems.
Supervised by Dr. Jia Li at Brown University's low-dimensional electronics lab, I built a nano-electronic device from graphene to measure its unique superconducting and magnetic properties at cryogenic temperatures. Building this device provided me experience with several innovative fabrication techniques like graphene exfoliation using scotch tape, nano-circuit design in DesignCAD Express, and electron beam lithography. I received an Undergraduate Teaching and Research Award to pursue this research.
I then initiated a collaboration with the Department of Chemistry (PI: Professor Brenda Rubenstein) to better understand this graphene system theoretically. My contributions included implementing a program called TriCrystal (repository available here) that generated the system’s structure, running baseline band structure, density of states, and Fermi surface calculations, and proposing a cluster-based technique to overcome limitations in modeling large systems using DFT. My thesis laid the groundwork for my advisors to continue studying this system using a combined approach, which they were awarded a DEPSCoR grant to do. An example of the computational structures I generated are shown below.
Advised by Dr. Sara Malvar and Dr. Ranveer Chandra at Microsoft Research, I developed a multimodal machine-learning based method to predict food protein digestibility, a value normally determined using extensive animal experimentation. To tackle this problem, I first created a 2000-point ground-truth protein property dataset that mapped features such as food nutritional information and protein sequence embeddings, extracted from a pre-trained transformer model, to digestibility coefficients. Then, I used Shapley value analysis to only select features that had established relationships to protein digestibility in the literature to train tree-based models on. We found that this biologically-interpretable dimensionality reduction was key in achieving the high accuracies our models did through an ablation study. We filed a provisional patent for this methodology and after the summer, I co-authored a paper summarizing our findings that is available on arXiv. This paper has since then been revised further and submitted to npj Science of Food. The downstream impact of our results is the reduction of animal experimentation in the development of new food products.
I addressed diversity and inclusion issues within the Physics department in a committee of undergraduates, doctoral students, and faculty. One of the projects I significantly contributed to was designing a climate survey to collect data regarding the state of diversity in the department which received over 100 responses. I was a part of this committee between January 2021 to May 2023.
I organized numerous community-building initiatives such as group study sessions, lunches, and peer mentoring alongside two other lead coordinators. Our goals are to build a strong sense of community for women and other underrepresented gender identities in the Physics department. I was a coordinator for this group between January 2021 and May 2023.
Since August, I have been supporting 20 students who belong to underrepresented gender identities in STEM in their endeavor to develop machine learning solutions to industry-issued challenges. Responsibilities include holding meetings for students, providing debugging support, and evaluating coursework.