Machine Learning Research Intern
2025 — PRESENTUT Dallas — Prof. Xinfang Jin · GAN-based 3D microstructure generation
- Built a Python preprocessing pipeline converting DREAM.3D exports into 64³ voxel stacks with measured volume fractions and surface areas — replacing an entire MATLAB workflow, with batch --multi and spatial --tile-xy modes for 500³+ volumes.
- Diagnosed two bugs causing pore-phase collapse in a WGAN-GP, then designed a differentiable connectivity loss (3D-convolution isolation penalty + percolation term) — fixing collapse on synthetic data (0.48 → 0.90, 0.59 → 0.86, FAIL→OK); a real-data ablation then isolated when the term helps versus overshoots.
- Validated on 101 real DREAM.3D microstructures — generated structures reproduced real Ni and pore connectivity in the OK band (S = 0.90 / 0.87), with Ni connectivity emergent (never a training target).
- Built 4_CNNCT, an analysis module for phase percolation, triple-phase-boundary density, and tortuosity — shipped with a 23-test suite.
- Automated a manual ParaView workflow with pvpython; its ground-truth test harness caught 4 silent label-corrupting bugs before handoff.
- Ran the full pipeline end-to-end: CNN estimator → connectivity-aware WGAN-GP → persistent-homology validation → transport validation.