LeJEPA World Models for Scientific Systems

• Exploring LeJEPA (SIGReg isotropic-Gaussian regularization) as a principled, heuristic-free route to self-supervised latent world models for scientific data, building on LeWorldModel’s stable end-to-end JEPA training
• Built an action-conditioned JEPA world model regularized with SIGReg on spatiotemporal physical fields (shallow water, 2D fluid flow, ERA5 reanalysis), evaluated downstream on data assimilation and sensor placement

ScatterPrism: Generative Model for Particle & Nuclear Physics

• Developed a Conditional Flow Matching generative surrogate that replaces slow Monte Carlo simulation of particle collisions and inverts it to infer physics from detector measurements; validated on Jefferson Lab scattering data
• Showed that low training loss does not guarantee physically correct outputs, and built a multi-metric diagnostic protocol to verify kinematic accuracy without memorization, motivated by the upcoming Electron-Ion Collider (EIC)

PyTorchFire: Differentiable Wildfire Simulation

• Architected a GPU-accelerated, differentiable wildfire Cellular Automata simulator that runs orders of magnitude faster than CPU-based simulators on real-world-scale, high-resolution fires
• Enabled real-time parameter calibration via gradient descent to match observed fire spread, with stronger generalization than supervised surrogates
• Extending to wildland-urban interface (WUI) scenarios; built FireDataForge, an open-source data framework that retrieves and harmonizes 11 wildfire-related geospatial sources (weather, vegetation, elevation, built environment, WUI, satellite imagery) into analysis-ready grids

Cosmo3DFlow: Generative Early Universe Reconstruction

• Co-designed a generative framework that reconstructs the early universe from present-day cosmic structure, using the 3D Discrete Wavelet Transform (DWT) with flow matching to compress the sparse, mostly-empty 3D volume into a spectral form
• Achieved up to 46\(\times\) faster sampling than diffusion models on 128³ N-body simulations, cutting early-universe initial-condition generation from minutes to seconds via stable large-step ODE solvers

DT-SegNet: Deep Learning for Materials Science

• Built an end-to-end two-stage pipeline (YOLOv5 detection + SegFormer segmentation) for automated precipitate identification and measurement in electron-microscopy images of chromium-based superalloys
• Outperformed established tools (Weka, ilastik) across accuracy, precision, recall, and F1, enabling high-throughput microstructural analysis for alloy development