HOME JOURNEY LEADERSHIP RESEARCH
STATUS: RESEARCH_ACTIVE

ARCHITECTING THE FUTURE OF INTELLIGENT SYSTEMS.

A repository of deep research across scientific machine learning, structural & biomedical materials, and population health. Spanning Brown University, NVIDIA's Deep Learning Institute, NASA, and the NIH.

01 // APPLIED ML

Physics-Informed Neural Networks BROWN_UNIV // NVIDIA_DLI
Drug-release data and model fits for flat, 1D-wrinkled, and 2D-crumpled films, comparing the Higuchi, Peppas, and Fick's Law classical models against the PINN model.
PINN VS. CLASSICAL RELEASE MODELS

PHYSICS-INFORMED NEURAL NETWORKS

Developed at Brown in collaboration with NVIDIA's Deep Learning Institute, working with Prof. George Em Karniadakis and Dr. Khemraj Shukla. It is a PINN and Bayesian PINN (BPINN) method for predicting drug-release rates from planar, 1D-wrinkled, and 2D-crumpled controlled-release films, built in PyTorch and JAX. By integrating Fick's diffusion law directly into the network alongside limited experimental data, the model benchmarks against, and outperforms, the classical Fick, Higuchi, and Peppas release models, cutting prediction error by up to 40% and reaching target accuracy for the planar-film case using just 6% of the typical experimental timeframe.

NVIDIA DLI PyTorch JAX Bayesian PINNs

Read the preprint on arXiv →

As Featured in Brown University News JULY_2026
0%
Reduction in Required Experimental Data

For simple, planar controlled-release materials. The reduction is 67% for more complex, folded geometries. Read the Brown University News feature →

DRUG-RELEASE PREDICTION

Collaborative research with Prof. Vikas Srivastava (Brown Engineering), Prof. George Em Karniadakis, and Dr. Khemraj Shukla. This work could cut development time for therapeutic patches, bandages, and implants. The press coverage above and the peer-reviewed preprint at left describe the same underlying model.

Predictive Modeling // Healthcare Provider Behavior IMPIRICUS

PREDICTING HEALTHCARE PROVIDER ENGAGEMENT IN SMS CAMPAIGNS

Applied predictive modeling to forecast how individual healthcare providers engage with SMS-based outreach campaigns. This is the same predictive-modeling-of-provider-behavior work that underpins product development at Impiricus, formalized and written up as a standalone study.

Predictive Modeling Healthcare Provider Behavior

Read the paper on arXiv →

02 // BROWN UNIVERSITY RESEARCH

Mussel Byssal-Plaque Mechanics MIT COLLAB // QIN LAB
Figure showing mussels attached to a surface by byssal threads, a close-up of a byssal plaque, a fluorescence microscopy image of the plaque, and a schematic of the modeling, 3D-printing, and mechanical-testing workflow used to study plaque attachment strength.

WHY MUSSEL BYSSAL PLAQUES ARE TINY YET STRONG

First-author study, conducted during his Brown years through a collaboration with Dr. Zhao Qin's lab at MIT, combining mechanical testing, 3D printing, and numerical modeling to explain how mussels anchor themselves to wave-battered surfaces. Found that byssal plaques 3–5x the diameter of their thread give optimized attachment strength, and that overall bonding strength depends more on network architecture than on plaque size alone.

First-Author Mechanical Testing 3D Printing Numerical Modeling

Read the peer-reviewed paper on Cell Press →

First-Author Publication // Ultrasonic ML Crack Detection SRIVASTAVA LAB // BME
Scatter plots of CNN-predicted crack length and crack position versus actual measured values on experimental high-density polyethylene ultrasound data, both tightly following the prediction-equals-actual line.
CNN PREDICTION VS. ACTUAL (EXPERIMENTAL)

MEASURING EMBEDDED CRACKS IN HDPE VIA ULTRASOUND

High-density polyethylene is used in nuclear-plant cooling pipelines and biomedical implants. It can develop embedded crack-like flaws during fabrication or operation that risk catastrophic failure if undetected. Working with Prof. Vikas Srivastava's Biomedical Engineering lab, built a finite-element-simulation-trained CNN that predicts crack length and position simultaneously from only tens of microseconds of ultrasound A-scan signal, achieving 3.2% mean absolute percent error on crack length and 3.8% on position.

First-Author CNN Finite Element Simulation Nondestructive Evaluation

Read the arXiv preprint →

Broader Structural & Biomedical Materials Research SRIVASTAVA LAB

Beyond the two published studies above, ongoing work in Prof. Vikas Srivastava's lab at Brown applies machine learning to a wider set of structural and biomedical materials problems.

  • Analyzing structural defects in polymers with machine learning techniques
  • Understanding the mechanics of hydrogels for biomedical applications
  • Using machine learning to predict dietary intake and health status

03 // NASA

NASA Space Grant // Research Scholar
Photo of a researcher assembling a folded, faceted geometric paper structure on a workshop table, part of NASA Space Grant research into material behavior under stress.

A NOVEL GEOMETRIC STRUCTURE

Investigated a novel geometric structure I invented to study material behavior under stress for NASA. The approach combines mechanics and statistics to predict how materials behave under stress, applied to NASA space exploration technologies.

Research Program NASA_CT_SPACE_GRANT
Photo of a folded, faceted geometric structure made of blue paper mounted in a mechanical testing machine between two metal compression plates.

STRUCTURAL BEHAVIOR UNDER EXTREME LOADING

As a NASA Space Grant Research Scholar, applied statistical and mechanical modeling to the question of how materials behave under the extreme structural and thermal loads seen in space exploration. This work connects directly to the same physics-informed and data-driven modeling approaches used in the Applied ML and Brown University research above. The research looks at structural materials and how they respond to stress and strain under load, including how and when they start to fail. Understanding these failure modes under extreme conditions helps predict which materials and structures can hold up during space missions.

Structural Mechanics Statistical Modeling

04 // NIH & POPULATION HEALTH

NIH All of Us Program POPULATION_HEALTH
Photo from the CPGI 2023 presentation of NIH All of Us research findings, two presenters speaking at a podium with an All of Us Research Program banner in the background.

LARGE-SCALE STATISTICAL & ML ANALYSIS

Research through the NIH All of Us Program applying statistical and machine-learning analysis to large-scale population health datasets, exploring correlations between maternal mortality, cardiovascular disease, and social determinants of health. Presented these findings with Ketan Pamurthy at CPGI 2023.

  • Large-scale statistical & ML analysis of NIH All of Us data
  • Correlating maternal mortality with social determinants of health

See the CPGI 2023 presentation on LinkedIn →

Brown University // Biomedical Informatics SARKAR LAB

ETHNICITY, SOCIAL DETERMINANTS & CARDIOVASCULAR DISEASE

Research with Prof. Neil Sarkar in Brown University's Department of Biomedical Informatics studying the role of ethnicity and social determinants in cardiovascular disease, including an exploration of "Blue Zones" longevity patterns in the U.S., regions where residents live measurably longer, healthier lives.

  • Ethnicity & social determinants in cardiovascular disease
  • Exploring "Blue Zones" longevity patterns in the U.S.
Independent Study // Brown University UNPUBLISHED
Figure showing hand image segmentation into fingers and palm regions, illustrating the steps from a raw hand photo through background removal, contour detection, finger and palm segmentation, feature extraction, and modeling.
HAND IMAGE SEGMENTATION AND FEATURE EXTRACTION PIPELINE

ANALYZING HAND IMAGES FOR DIET AND HEALTH

This independent study at Brown looks at whether a photo of someone's hand can say something about their diet and health. The method starts by adjusting the colors in each photo so they match up across different lighting and cameras. It then finds the hand in the photo, separates it from the background, and splits the image into the left and right hand.

From there, it finds key points on the hand and uses them to split each hand into its fingers and palm. For each of these parts, it pulls out features like color patterns and shape measurements, and puts them on a common scale so they can be compared across photos. The next step, proposed but not yet done, is to use these features with statistics and machine learning to try to predict health outcomes.

This work is unpublished. If you are interested in it, feel free to reach out.

Image Processing Feature Extraction Health Prediction