FEATURED — CONSTANT-POTENTIAL MLIP PIPELINE (PRIVATE LANL)
Three composable packages.
Together: generate, ground-truth, and learn electrified-interface physics for CP-MLMD.
Stage 1 — Structures
electrode-electrolyte_interface_builder
Building physically valid electrode-electrolyte structures for MLIP training is months of manual work most groups skip. I wrote a stage-validated pipeline that generates IrO₂/RuO₂ slabs across four facets, populates the full electric double layer with water, Na⁺, and ClO₄⁻, places OER/HER/CER adsorbates at chemically valid sites, and validates each. Input layer of the HIPPIE-NN pipeline.
Stage 2 — Ground Truth
qe-tools
Daily friction in computational electrochemistry is in the plumbing: months of manual HPC setup, monitoring, and recovery. I built a Quantum ESPRESSO automation suite that compresses full OER campaigns to 2–3 weeks of unattended compute. It also adds operando XAS under ESM-RISM — the first such implementation in QE. Every LANL electrochemistry project ran on this toolkit.
Stage 3 — Model
hippynn-softfqeq
Every charge-aware ML potential predicts a flat electrochemical potential across the electrode-electrolyte interface — an artifact of a single shared chemical potential. HIPPIE-NN doesn't. My HIPPYNN fork adds Soft-FQEq: per-fragment augmented-Lagrangian charge equilibration so the electric double layer emerges from training. Forces by autograd via Uzawa unroll. ~50,000× faster than DFT-MD.
PUBLIC RESEARCH SOFTWARE
Open code from PhD and pre-doc work.
FEM Multiphysics for Liquid-Crystal Networks
Light-responsive liquid-crystal polymers couple photochemistry, thermal transport, and large-deformation mechanics — no off-the-shelf FEM solver handles all three at once. I wrote a Fortran/Abaqus UTEMP/UFIELD platform (~2,200 lines, 15 subroutines) with a solver-within-solver architecture. Backbone for 13 publications, adopted by collaborators across three countries from documentation alone.
MD for Liquid-Crystal Networks
Photo-induced density changes in azobenzene-doped polymers were attributed to static geometric mismatch between trans and cis isomers. I built an all-atom LAMMPS framework — 7,278 atoms, PCFF force field, probabilistic dihedral switching — and showed dynamic trans-cis-trans cycling, not static geometry, drives the 15.7% density reduction. Static isomerization produces none. J. Chem. Phys. (2024).
Thermal DEM for Granular Beds
Fusion breeder pebble-bed thermal conductivity was modeled with assumed-uniform packings. The thermal-resistor-network DEM platform I built at KIT maps real particle configurations to conductance matrices via Batchelor-O'Brien contact and Smoluchowski rarefied-gas conduction. Validated across 8+ ceramics from 25–800°C — widest material coverage in the granular-bed ETC literature.
OPEN SOURCE — SHIPS FOR THE COMMUNITY
Tooling side projects.
Template_Scientific_Figures
Every researcher spends weeks tuning matplotlib defaults to look journal-publishable. I open-sourced the templates I built across 13 publications — single-column, double-column, and journal-specific layouts with sensible defaults for typography, color, and figure density. Drop in, adjust data, export.
claude-resume-kit
Applying to roles in computational science means rewriting your resume for every JD. I built a Claude Code-driven workflow that extracts a researcher's work from papers once, then generates tailored LaTeX resumes, cover letters, and critique reports per posting. Five skills handle the full pipeline: job scouting, resume generation, cover letter drafting, editing, and critique. MIT license.