My current research focuses on building agentic AI harnesses for scientific computing and complex-systems analysis. This includes LLM-based pipelines that extract and ground causal scientific claims against the literature, agentic IDE workflows that accelerate research-software development, multimodal extraction systems that parse structured information from documents combining narrative, tables, and figures, and developer-efficiency tooling that wraps frontier models around domain-specific codebases. I am also interested in autonomous discovery - agentic systems that close the loop between scientific hypothesis generation, robotic experimentation, and high-performance computing for analyzing results. This work aligns with broader DOE efforts such as the Genesis Mission and American Science Cloud.

As the SSS Division AI Representative at Argonne National Laboratory, I advise researchers across the division on the practical engineering of these systems and authored the Division's first AI training module.

My applied work centers on the intersection of AI and large-scale transportation and logistics simulation. This includes architectural work on high-fidelity discrete-event simulators, multiscale modeling for traffic flow, and the integration of LLM-based reasoning components into simulation toolchains. The long-term direction is building AI harnesses that allow analysts to drive complex simulations end-to-end — closing the loop between hypothesis, scenario design, simulation, and decision support. 

I have broader interests in applying machine learning to scientific simulation problems. Deep learning is particularly well-suited for analyzing the differential equations arising from traffic flow — physics-informed neural networks for traffic state estimation (TSE) and equation discovery are natural applications. I am also interested in graph embedding techniques such as Node2Vec for training heuristic functions that accelerate A* search on large multimodal transportation networks.

During summer 2023, I served as a graduate student intern at UMD ARLIS. With guidance from Dr. James Baeder, our team used machine learning to improve CFD solvers for the Spalart-Allmaras (SA) turbulence closure model for Reynolds-Averaged-Navier-Stokes (RANS) equations. While turbulence models for RANS are attractive compared to computationally intractable direct numerical simulation (DNS), current turbulence models perform poorly when adverse pressure gradients (e.g. separated flow) are present. We used Field-Inversion-Machine-Learning (FIML) to produce a neural network correction to the SA model by modifying the turbulence production term. In-house mesh generation and flow solvers were computed using the UMD Zaratan HPC cluster (80 Nvidia A100 GPUs). For the FIML step, adjoint methods were used to compute gradients.

Figure: One sees a recirculating flow in the separated flow region. The nlf-0416 airfoil is experiencing stall due to high angle of attack (20 degrees). The SA turbulence model is inaccurate in quantifying loss of lift owing to this phenomenon.