I am a fourth-year Physics Ph.D. candidate at the Department of Physics at the University of Maryland, College Park. I am advised by Prof. Joshua Weitz and mentored by Dr. Stephen Beckett. My Ph.D. thesis is on, "Bayesian learning of phage-microbe interaction traits from population dynamics across scales".
My research focuses on solving inverse problems and optimization problems using time series data. I develop interpretable models to uncover parameters and dynamics of complex systems; and optimal methods for downstream decision-making under realistic constraints. My mission is building quantitative models for the human good. So far, my models have:
I am fortunate to have worked alongside excellent scientists and engineers at
University of Maryland
Georgia Tech
Simons Foundation
University of British Columbia
IISER Kolkata
Mission: Building quantitative models for the human good
Phage Therapy: Some viruses (phages) can infect and kill bacterial cells, and are hence used for treating resistant infections. We design pharmacological models of phage therapy using phage cocktails and antibiotics. Aided by HFIM data, we use inverse models to inform such infection treatment models and ask the question: is combination therapy better than single therapy? what is the optimal timing and dosage of such treatments?
Network of community infections: In natural ecosystems, multiple species of phages coexist with their bacterial hosts. From observed population dynamics data, we address the question: which phage infects which microbes? Which microbial groups cooperate vs compete? How to infer underlying traits from the observed population dynamics? How different are pairwise interactions from community interactions? How are the populations stabilized in a natural ecosystem? We use ODEs and interaction models to parametrize these systems and constrained optimization techniques to solve inverse problems in these complex systems.
Optimal sensing via microrheology: In experiments guided with Optical Tweezers, a submicroscopic probe is embedded in complex non-Newtonian fluid. From the observed motion of the microscopic probe, the rheological properties of the fluid can be determined. We design(ed) optimal control algorithms to optimize the signal-to-noise ratio in wideband response measurements, such that they remain in a linear-response regime. Our constrained optimization algorithm outperforms standard techniques in terms of speed and accuracy, which is essential for time-sensitive experiments with biological media. We have used these to study how viscoelastic properties of mutated Lamin differ from the wild type, a mutation linked to cardiomyopathy.
Sequential Bayesian localization of acoustic emitter via mobile drone: Drones are widely used in search and rescue. Through numerical simulations, we devised a sequential Bayesian framework, for optimally moving a receiver drone to localize an acoustic emitter. The drone uses phase difference of arrival from the source at different time intervals, and each step identifies a posterior for possible location; then it computes where to optimally move. (Hint: moving directly towards the source is not optimal.)
LLM-based time series forecasting: LLMs being trained on large datasets have reliable autoregressive next-word prediction capabilities. Tokenizers are being developed where numbers are mapped to tokens and solving the next token estimation problems with pre-trained LLMs solves the next number prediction problem for time series. For different LLMs and tokenizations, we investigate how zero-shot or few-shot learning is effective for chaotic and stochastic systems (Hint: it is not!).
Gen-AI for time series extrapolation: Gen-AI models such as VAE, GAN, and diffusion models are excellent in approximating the underlying probability distributions from samples. For physics-derived dynamical systems, we explore Gen-AI models for predictive tasks for undersampled data and sparse time series.
Real random numbers from Brownian trajectories: Algorithms for random number generation produce pseudo-random numbers. Through experimental trajectories of Brownian probes captured via our optical tweezers setup, we developed random number extraction algorithms, based on NIST tests for randomness. We found that the entropy of randomness improves asymptotically as the sampling rate increases.
Non-equilibrium statistical mechanics: We study energy dissipation and thermodynamic currents driving stochastic probes in fluids. With experimentally recorded Brownian trajectories, we discovered how these currents converge to an arcsine distribution, and how they have directional preference depending on the timescale of observation.
Optical Tweezers: A collimated beam of laser when focused by a high numerical aperture lens can capture microscopic probes, such that they can be moved along with the laser beam. At the Light-Matter Lab (IISER Kolkata) , I developed custom-built optical tweezers and imaging systems for the simultaneous capture of multiple sub-microscopic particles; along with modulation and tracking with nanometer resolution.
Quantum Spectroscopy: Entangled photons interacting with matter have their properties altered due to light-matter interactions. At the Silva Lab (Georgia Tech), I worked on building time-synchronized dual single-photon detection systems, for quantum metrology with these entangled photon pairs.
Microscopy and imaging: At the Li Lab in the University of British Columbia, I build a custom microscope with two high numerical aperture lenses and 3D motor controls enabled by gaming joysticks.
GMRT (Giant Meterwave Radio Telescope) in India, employs up to 33 radio telescope dishes to image pulsars (rapidly rotating neutron stars that emit beams of electromagnetic radiation). Due to RFID, these frames (lasting only a few seconds have low SNR). Using multi-detector observations and stacking, we developed a pipeline to filter RDIF corrupt data for radio frequency observation.
Designed an optimal control method to perform microrheology on proteins in solution such that the signal-to-noise ratio is maximized over a broadband frequency range (2 KHz).
Using a feed-forward neural network, we fit the best basis function to compute thermodynamic currents of a Brownian particle in water (a constrained non-convex optimization problem.) Showed how the skewness properties of these stochastic currents change non-mononotically over time.
Used our optimal control algorithm to maximise the broadband SNR for microrheology experiments on Lamin proteins. Laminopathies, such as cardiomyopathy are driven by mutations of these Lamin proteins. With some mutations the nuclear walls disintegrates. That changes the viscoelastic properties of the cell. However, we found out that even before the cell disintegrates, the protein in solution exhibits different viscoelastic properties, thereby aiding into fast detection.
Created an online feedback-control algorithm to maximise the signal to noise ratio (SNR) while performing broadband active microrheology measurement with modulated stochastic probes. We improved the benchmark SNR 10x on a 2 kHz bandwidth.
Developed a class of Machine Learning algorithms to extract random numbers from dampled driven Physics based stochastic time series. Demonstrated through an Ornstein-Uhlenbeck process with 3 different potential functions.
Invented a technique for algorithmically optimizing the signal to noise ratio (SNR) for wideband active microrheology.
Developed an optimization algorithm integrated with National Institute of Standards and Technology (NIST) tests for randomness suite to characterize how random are experimentally sampled tracjetories of Brownian particles. Our inverse model extracts real-random numbers from stochastic trajectories instead of algorithmitcally generated pseudo-randoms. Demonstrated on the experimentally measured trajectories of an optically trapped Brownian particle in water.
Formulated a novel two-point technique to probe viscoelasticity of a medium using modulated Brownian osciallators. Using two probes instead of one, creates motional resonance between the probes and provides a finer insight into medium viscoelasticity.
A lightweight, powerful and intuitive Python library for parameter optimization in ordinary differential equation (ODE) systems. InvODE combines global optimization with local refinement techniques to efficiently find optimal parameter sets for complex dynamical systems.
Tools and examples for integrating ODE-based models with Bayesian inference for biological systems.
GUI software for Bayesian analysis of phage and microbes, built with PySide, Qt, and PyMC (co-developed by: Timothy Cai).
Summer research intern
ML based virus-microbe network inference
Ph.D. Student (& GRA)
Physics
Ph.D. Student (GRA & GTA)
Physics
Mitacs Graduate Research Intern
Instumentation Engineering
B.S.-M.S.
Physics (Applied Optics and stochastic Physics)
Visiting Summer Research Program
Radio Astrophysics (Pulsar division)
UMD, Microbiome Center
UMD, Microbiome Center
Department of Physics, UMD
Graduate School, UMD
University of Syracuse
MKS Instruments
Canada
Government of India
Delivered a talk on using multimodal machine learning, that integrates data, ML model, Physics etc., to infer network edges. Showed how this multimodal approach is better than the conventional unimodal appoach in both parameter inference and time-series forecasting.
Delivered a talk on how life-history traits of a single strain of virus infecting a single strain of microbe is changed when put together in a complex community with multiple interacting strains. I start by building an inverse model for pairwise interactions. Talked how the forward version of that models cannot fit the community data. Next I build a scaled-up model for the community interactions and use Bayesian inference to infer life-history traits of these strains. But that cannot also describe the community completely. Only by integrating data and model iteratively and taking into consideration the higher order emergent effects the community transient dynamics of these strains could be recapitulated.
Talked about our new ML based model that integrates NIST tests for randomness with a stochastic trajectory to (a) calibrate the stochastic process through parameter estimation and (b) extract "true" random numbers from a stochastic trajectory that optimizes entropy of randomness.
A summer bootcamp on scientific computing for beginners with Python and PyTorch organized by Pratyush Tiwary, University of Maryland.
This introductory Physics course introduces programming and goal-driven problem solving tailored to physicists. The courses cover topics from basic mechanics to electromagnetism with hands-on problems solving sessions.
A deep dive into the advanced concepts of quantum mechanics, emphasizing modern research methods and computational techniques.
