Research
What I work on
Machine learning is now good enough to be asked engineering questions, and bad enough at arithmetic that we should not trust its answers on their own. My research builds systems where the model is held to something external and checkable — a numerical solver, a circuit equation, a test suite — so that the result is both data-efficient and verifiable. Most of this work is grounded in power systems, where being wrong is expensive.
Interests: agentic AI, physics-informed machine learning, foundation models, scientific machine learning, optimization, and smart grids.
Projects
-
Hybrid surrogates across discontinuities
A surrogate of a hybrid system should itself be a hybrid system. Contact and impact do not live in one smooth equation: a bouncing ball reverses its velocity at an instant, and a press-fitted pin sticks, slips, and lets go. DA-PINNs put one small network on each smooth interval and train the networks, the event times, and the unknown physical parameters together. The number of events is identified rather than fixed. On the ball the method finds four impacts and the restitution to 0.1%. On a MuJoCo gantry pulling a press-fitted pin, it recovers the stick–slip–release sequence and both friction coefficients within 2%, including the pin velocity, which is never measured.
-
Tool-augmented LLMs for power systems
Answering questions about distribution optimal power flow requires a solver, not recall. VeraGrid-Agent writes the feeder and DER simulator input, executes the open-source VeraGrid solver, and reads back only the relevant result sections, keeping the context compact and every claim traceable to solver output. On VeraGrid-MCQ-150, a deterministic benchmark of 150 expert-templated questions we introduced, no-tool reasoning scores 42.7–49.3% while the agent reaches 97.3–100%.
-
Agents that are made to prove themselves
Two lines of work on holding generated artifacts to executable specifications. A test-driven agentic framework synthesizes deployable low-level robot controllers, alternating between prompt-level refinement and direct code editing in response to structured PyTest feedback, evaluated on 2D navigation and 3D Webots scenes. MathAgent, with Kishor Kumar Bhaumik, builds a SciPy-augmented benchmark for agentic scientific computation.
-
Physics-informed synchro-waveform analytics
Inverter-based resources are hard to model from data because disturbance events are rare. Embedding the circuit relationships into the learning problem compensates for that scarcity: the framework estimates the IBR terminal current response from only a few network disturbance signatures, and in the harder case jointly identifies the transient dynamics and the unknown line parameters. Validated on real waveform measurement unit data across several sampling rates.
-
Event detection in continuous waveforms
Recasting event detection in continuously recorded IBR waveforms as temporal object detection on spectrograms. Short-time Fourier transforms per channel are stacked into a tensor, which exposes transient and harmonic signatures more explicitly than raw time series; a U-Net over that representation beats an equivalent-capacity raw-waveform detector on both single-phase disturbances and three-phase faults.
-
Online optimization with latent parameters
What if the cost function you are minimizing depends on parameters you cannot measure, evolving under dynamics you do not know? Using control-theoretic identification on gradient measurements alone, the parameter dynamics can be recovered in the original coordinates — and, under mild conditions, from only finitely many measurements. A follow-up handles noisy gradients via a Gauss-Markov reconstruction and an instrumental-variable estimator, with a bound on expected tracking error.
-
Distributed optimization for multi-agent systems
Earlier and continuing work on coordination without a coordinator: a partially distributed economic dispatch algorithm that converges in fixed time while accounting for transmission losses through Kron’s model, algorithm fusion for heterogeneous online learning, and exact distributed attitude alignment on SO(3) for robot teams whose local objectives conflict.
Talks & presentations
-
Mar 2026
Physics-informed learning
NSF presentation, Rutgers University, March 17, 2026.
-
Sep 2025
NC4 talk
Army Research Laboratory, September 18, 2025.
-
Sep 2024
NC4 work
Army Research Laboratory, September 26, 2024.
-
Ongoing
Quarterly project updates
Presented to the Army Research Laboratory as the funding agency.
Professional service
-
Power systems
Reviewer for IEEE Transactions on Smart Grid, IEEE Transactions on Dependable and Secure Computing, and IEEE PES Grid Edge.
-
Control
Reviewer for IEEE Transactions on Automatic Control, IEEE Control Systems Letters, the Conference on Decision and Control, the American Control Conference, the Conference on Control Technology and Applications, and the European Control Conference.
Tools I use
-
Languages
Python, C++, MATLAB
-
Learning
PyTorch, TensorFlow
-
Solvers
CPLEX, OSQP
-
Simulators
PSCAD, OpenDSS, VeraGrid, Webots, MuJoCo, ns-3