Cosmo3DFlow: Efficient 3D Generative Modeling (2026)

• Co-designed a flow matching framework for cosmological inverse problems, using 3D wavelet representations to address spatial sparsity in 128³ N-body simulations (KDD 2026)
• Achieved up to 46× faster sampling than diffusion models with stable large-step ODE solvers, reducing early-universe reconstruction from minutes to seconds

PyTorchFire: GPU-Accelerated Differentiable Simulation (2024–present)

• Developed a PyTorch cellular automata simulator for high-resolution wildfire modeling, running orders of magnitude faster than CPU-based simulators
• Enabled real-time gradient-based parameter calibration through differentiable simulation, matching observed fire spread with stronger generalization than supervised surrogates
• Built FireDataForge, an open-source Python data pipeline that retrieves and harmonizes 11 geospatial sources into analysis-ready grids; extending simulation to wildland-urban interface scenarios

DeepVELA: Agentic Benchmark for Scientific Applications

• Explore agentic benchmarking within Genesis Mission, collaborating with MLCommons to evaluate time, energy, space, cost, and precision
• Design agent harnesses with structured outputs and verifiable results to support trustworthy, reproducible evaluation

ScatterPrism: Generative Simulation & Inverse Problems (2024–present)

• Developed a conditional flow matching surrogate for particle-collision simulation and inverse inference from detector measurements; validated on Jefferson Lab scattering data
• Designed a multi-metric model evaluation protocol to detect physical errors hidden by low training loss and assess kinematic fidelity without memorization

JEPA World Models for Scientific Systems (Exploratory) (2026)

• Prototyped and trained Joint Embedding Predictive Architecture (JEPA) models for spatiotemporal scientific data; explored action conditioning and LeJEPA/SIGReg regularization
• Investigated self-supervised representations of physical fields (shallow water, 2D fluid flow, ERA5 reanalysis) for potential data assimilation and sensor placement applications

DT-SegNet: Computer Vision for Materials Science

• Built an end-to-end object detection and semantic segmentation pipeline with YOLOv5 and SegFormer (vision transformer) for electron-microscopy images
• Outperformed Weka and ilastik across accuracy, precision, recall, and F1, enabling high-throughput precipitate identification and measurement