Back to the Future
Predicting Long-Horizon Dynamics with Retrospective Correction for Disturbance-Aware Retrieval
Grabbing a bottle from a packed cabinet is easy for people and hard for robots. With no room to lift the target over its neighbors, a robot has to push through them, and the disturbance can travel past the nearest objects and topple one. As first author, I led DARE: its graph-neural-network world model predicts how contact spreads from object to object, and its learned corrector goes back to fix earlier predictions when a new observation arrives. Under noisy observations, that correction lifted 10-object retrieval success from 0% to 80–100%; with clean ones, my planner succeeded in 90–100% of simulated 20-object cabinets, against 10–20% for direct-retrieval and collision-avoidance baselines.
Popping Bubbles
Isogeometric Navier–Stokes–Korteweg Simulation of Bubble Nucleation and Cavitation
Collapsing vapor bubbles damage ship propellers, can damage the fine lines on semiconductor wafers during cleaning, and have been proposed as a mechanism of blast-induced brain injury. Predicting how a bubble forms and evolves means modeling a liquid–vapor phase change across a moving interface. With Dr. Chad Landis at the Oden Institute, I model it with the Navier–Stokes–Korteweg equations and am building a JAX isogeometric solver that reaches optimal O(h⁴) convergence. Resolving that moving interface stretched runs to as long as 166 hours, so I am designing a fully adaptive mesh with self-stopping refinement thresholds, projected to cut them to 25 hours, toward a first-author paper.
KLA 5D Analyzer
Chipmakers rely on KLA’s inspection and metrology systems to reach their target yields. Across two internships, I worked on KLA's 5D Analyzer. I adapted and refactored normalized spatial uncertainty (NSU) code into Python for a team metrology-targeting pipeline that cut runtime 33% over the grid-based approach. I also made NSU visible across the whole die, with gradients that auto-scale to chosen criteria instead of manual color and scale setup, and built an all-in-one data input node that replaced 15+ per-type input nodes and centralized data management for entire workflows.
Foam-Core Spacecraft Sandwich Panel
A satellite panel has to survive the shaking of launch and the temperature swings of orbit. On a CubeSat it must also carry loads, insulate and shield radiation inside a 10 cm × 10 cm square, where material, geometry and mass all compete. As engineering lead of a 9-person senior-design team working with Southwest Research Institute (SwRI), I directed the Ansys Mechanical vibration and thermal simulations of our foam-core sandwich panel and validated them against SwRI’s tests. The panel passed 14.1 g RMS random vibration on all three axes and came through thermal cycling from −40 °C to +61 °C without damage.