Professor & Research Executive
Science-Native Artificial Intelligence
I develop AI that is born from engineering — not applied to it as an afterthought. My research creates new architectures, models, and optimization methods where physical laws, engineering constraints, and domain structure are foundational design principles. This work drives a new generation of intelligent propulsion systems, critical infrastructure security and resilience, and physics-inspired learning, bridging fundamental theory with real-world impact.
My work sits at the intersection of machine learning theory and engineering practice — creating AI methods that are structurally shaped by the physical systems they serve.
Designing novel neural architectures and optimization algorithms where engineering constraints — symmetries, conservation laws, stability requirements — are not regularizers bolted on after the fact, but the very scaffolding from which models are built.
Physics-inspired learning, digital twins, and simulation-augmented AI that respect the governing equations of real-world systems. Developing methods that achieve data efficiency and generalization by encoding physical structure into the learning process.
AI-driven design, optimization, and control of propulsion systems, hydrogen-electric powertrains, and energy networks. From grid security at national scale to the engines powering zero-emission aviation.
Recently accepted and published papers.
Systematic study of emotional reasoning in LLM strategic decision-making. Manipulative agents excel in information-asymmetric games, but emotional reasoning becomes counterproductive under high uncertainty. Humans elicited 2x more cooperation from emotionally attuned LLMs.
Unified open-source framework for neural downscaling of climate data. Transformer-based models achieve highest accuracy; generative models produce realistic fine-scale structure in complex terrain. Nine architectures benchmarked across three public datasets.
A robust neural network architecture for predicting extreme weather events that outperforms common baselines, leveraging physics encapsulated in CMIP data. Highlights the impact of rugged terrain on extreme weather risk distribution.
A two-level architecture for wildfire evacuation: RL-driven vehicle policies coordinated by an LLM commander that issues network-level routing rules, with a cognitive task analysis layer for human interpretability.
A new approach combining reward sampling with online stochastic optimization for contextual bandits. Attains near-optimal reward more quickly and with substantially lower computation than Thompson Sampling and UCB.
A multi-agent LLM-enhanced platform that jointly simulates natural hazards and human response, coupling high-fidelity flood AI forecasting with agent-based models of human behavior for emergency management.
First comprehensive study grounding intrinsic dimension in interpretable text properties. Scientific prose shows low ID (~8), encyclopedic content medium (~9), and creative writing high (~10.5) — revealing that LLMs find scientific text "representationally simple."
Research demonstrations and project visualizations.
Demonstration of a two-level architecture for wildfire evacuation combining RL-driven vehicle policies with LLM-based coordination and cognitive task analysis.
An autonomous system for predicting mechanical properties of steel alloys using physics-inspired models, with built-in generalization auditing across compositions.
Simulation of heterogeneous traffic environments where autonomous and human-driven vehicles interact with realistic behavioral profiles.
Principal investigator on more than a dozen projects supported by multimillion-dollar competitive funding.
Interdata Research, MV-Innovation
Applied research, scientific advisory services, technology development, and grant strategy and proposal assistance.
FlexPower.AI, smarter energy for distributed inference
Energy markets, distributed energy storage, and AI infrastructure.
ZeroAvia, #1 U.S. GreenTech company in 2024 (TIME)
Leading key R&D and business activities in hydrogen-electric aviation powertrains.
Los Alamos National Laboratory, pioneering science for national and energy security
Supporting internal AI/ML initiatives and projects at one of the world's most advanced scientific institutions.
Los Alamos National Laboratory, pioneering science for national and energy security
Led multi-lab grid security, hydrogen, and AI/ML research programs.
Skolkovo Institute of Science and Technology, established in collaboration with MIT
Research and teaching in AI/ML and Energy across Computational & Data Science and Energy departments. Approved by LANL part-time engagement.
Higher School of Economics — "Math of Machine Learning" M.S. Program
Co-founded what is now a top-3 national M.S. program in machine learning.
INRIA Rhône-Alpes & Université Grenoble-Alpes, advancing digital science and applied mathematics
ML theory research in applied mathematics.
M.S. & B.S. — Applied Physics & Mathematics, Summa Cum Laude
2007 – 2009
Published at top venues including NeurIPS, ICML, ICLR, and PSCC. Program committee member for NeurIPS, ICML, ICLR, PSCC. Invited plenary speaker at UN COP 29 and other top-tier scientific and business forums.
Taught at various European and US universities. Lecture notes are available upon request.
Multiple best-paper and national research awards among mentees.
I'm always open to discussing research collaborations, speaking opportunities, and advising engagements.