Batteries and semiconductor processing. Designed experiments, capability statistics, and every number regenerated from raw data before it is believed. Seven projects, each with a live 3D page and an audit trail from claim to script. The particles behind this text are those seven projects' own data.
Every repository carries a docs/CLAIMS.md that traces each number on its README to the script and data file that produce it, a one-command reproduce script, and a page built with three.js from the project's own data. The thumbnails below are live: the same data as the hero, drag to orbit.
A 300 mm showerhead ALD chamber in OpenFOAM in which every wafer face runs its own feature-scale model of a 100 nm hole at aspect ratio 100, two-way coupled to the gas through a C++ boundary condition. The TMA pulse that saturates a blanket wafer, 0.35 s, leaves the 3D-NAND-like wafer with 56× the wall area 40.7× under-dosed; it saturates at 14.05 s, and a one-line supply rule, tpulse = 1.2 (1 + Awall/Awafer) N₀/(U c₀), holds within 25 % over aspect ratios 20 to 100 and a 20× range of TMA feed. The 1D hole model sits within 3 to 11 % of a converged ballistic reference (Argonne's Machball and the repository's own 3D voxel Monte Carlo solver, which agree within 2 %) and fits published PillarHall profiles; the rule puts a measured fluidized-bed ALD run on 940 m² of silica gel at 1.12× the supply limit. A 25-run central composite design (JMP response surface, R² 0.993) with Gaussian-process surrogates in MATLAB and scikit-learn (LOO R² 0.999) spans 51 to 97 % precursor utilization and finds a 12 % precursor saving per wafer at the same pulse time.
2,774 Materials Project electrodes through an eight-layer funnel. A three-descriptor Degradation Screening Index outranks volume strain alone (0.55) and the metal-identity heuristic it replaces, and puts 10 of 11 experimentally characterised cathodes on the correct side of the durable split. Sole author; ChemRxiv preprint, manuscript under revision at Phys. Chem. Chem. Phys.
Commercial LMO cells cycled on a Bio-Logic VMP-300, CV 0.1 to 0.3 mV/s and C/10 to 2C. A DFN model fitted on C/5 and C/2 only predicts the unseen 2C voltage within 17 mV and six CV peaks within 80 mV, then splits the 2C loss 74 / 17 / 6 % between ohmic, charge-transfer and diffusion terms. 58 automated tests.
300 mm showerhead reactor, 2D axisymmetric. The starved deposit was root-caused as mass-transfer limited (Da above 10) and fixed at the inlet: 10 to 150 mm gives 6.1 %. Gaussian-process surrogate with leave-one-out R² 0.944, 10,000-draw Monte Carlo on MFC and thermocouple tolerances, 3D model with asymmetric exhaust at 5.9 %.
All-atom NAMD simulations of four imidazolium [BF₄] ionic liquids in a (15,15) nanotube: chain length sets orientation, binding and mobility. Co-author, Structural Chemistry 35, 1743–1753 (2024). The repository ships the 19,260-atom minimised structure the page renders.
Ten electron-transport layers and nine back-contact metals in a six-parameter DOE; VOC 1.21 V, JSC 25.96 mA cm⁻², FF 89.3 %. Spiral panel architecture with 17.9 % more active area: two granted Indian design registrations, one utility application under review. Best Poster Award, 61st Annual Convention of Chemists.
A 20-cell 5s4p pack of 4680 cells built on COMSOL's validated propagation model: 25 Ah and 110 kJ per cell, 80 °C trigger. Twelve transient one-hour studies over base cooling from 0 to 2,000 W m⁻² K⁻¹ locate a sharp threshold in through-layer (radial) conductivity between 1.0 and 1.25 W m⁻¹ K⁻¹, show the threshold is a property of the cell rather than the cooling, and identify internal conduction as the controlling resistance. Reduced-order model and tested scaling analysis in the repository.
Graduating May 2027. Open to process, cell and materials engineering roles. stg3719@seas.upenn.edu
Every project above has a repository with a reproduce script and a claims audit; the fastest way to check the work is to run it.