Case studies

Research

Combustion kinetics and mechanism reduction, fair multi-objective optimization and uncertainty quantification, reduced-order and scientific machine learning.

Overview

MUQ-SAC in motion: each frame is one sampled Arrhenius curve tightening the joint uncertainty band for an n-heptane reaction (10× real speed) — see MUQ-SAC.

I develop fair multi-objective optimization, uncertainty-quantification and reduced-order/scientific-ML methods for expensive, stiff physical simulations — using combustion chemical kinetics and mechanism reduction as a primary proving ground, and reaching into CFD, propulsion and quantum chemistry as the same methods generalize. My PhD work produced MOSIAC, an open-source workbench for uncertainty-aware combustion mechanism optimization, alongside hands-on experimental and numerical combustion work: pool-flame characterization, a downdraft gasifier, and the computer-vision measurement instrument (flamelab) that quantifies pool-flame video in physical units.

Research themes

MOSIAC ignitability 3D map

01

Combustion Kinetics & Mechanism Reduction

Chemical kinetic mechanism development, validation and reduction to tractable, simulation-ready form — from H2/O2 to n-heptane, methyl butanoate and 1,3-butadiene. Workbench: MOSIAC.

MOSIAC ignitability 3D map

Open project
MUQ-SAC correlated-parameter uncertainty ellipsoid

02

Fair Multi-Objective Optimization & UQ

Multi-stage, multi-dataset optimization of Arrhenius rate parameters built around a fair mean objective across datasets; correlated-parameter Bayesian uncertainty quantification and its propagation forward through the model. Method: MUQ-SAC, applied via MSO objectives.

MUQ-SAC correlated-parameter uncertainty ellipsoid

Open project
SciML noise-floor figure from the Neural ODE surrogate study

03

Reduced-Order Modelling & Scientific ML

Neural and response-surface surrogates that stand in for expensive stiff simulations — where they work, where they don’t, and why. Deep dive: SciML; live demo: the Neural ODE playground.

SciML noise-floor figure from the Neural ODE surrogate study

Open project
Vaporizer rig pressure-drop-vs-diameter design chart

04

Experimental & Numerical Combustion, CFD

Pool-flame characterization and downdraft gasification alongside an in-house CFD/multiphysics solver, reacting-flow studies, vaporizer/burner rig design, and early-stage quantum-chemical rate-constant estimation. Rig design: vaporizer rigs; imaging: flamelab.

Vaporizer rig pressure-drop-vs-diameter design chart

Open project
RL training curve for rotor-tuning conformer search

05

Optimization & Reinforcement Learning

A from-scratch, dependency-light optimization/RL workbench with a plug-and-play graph IR — solvers cross-benchmarked head to head, sampling & UQ, a jax-backed NN/RL training layer, a browser pipeline editor, and an RL agent that tunes a real molecule’s rotatable bonds. Workbench: dokimazo.

RL training curve for rotor-tuning conformer search

Open project

Browse the full, filterable project list →

The strongest single thread: MOSIAC, the optimization workbench built during the PhD — see its step-by-step guide — feeding into the MSO objective-function work and the published MUQ-SAC uncertainty quantification method.

Research experiments

Smaller, self-contained studies that show how I work, not just what I’ve published:

  • Scientific ML for Stiff Kinetics — a documented, 50-generation journey from an untrainable neural ODE to a working surrogate, closing with a 115-run statistical audit of my own results.
  • Symbolic Regression for Rate Rules — using SR to rediscover the Evans–Polanyi relation from data, with a leave-one-class-out generalization check.
  • flamelab — a computer-vision instrument that measures flame length, width and flicker from pool-fire video in physical units with propagated uncertainty: eight-run ensemble campaign, regime classification, thirteen analysis methods, honest limitations included.
  • dokimazo — an RL agent tunes a real molecule’s rotatable bonds one at a time, independently reaching the same global-minimum conformer as a from-scratch CMA-ES implementation, checked against a directly-computable structural ground truth rather than solver agreement alone.

Methods & tools

Every tool below links to a project that actually used it — no skill bars.

  • Fair Multi-Objective Optimization — genetic algorithms, CMA-ES, gradient/NLLS, Bayesian & GP optimization, surrogate-assisted search — see MSO objectives
  • Uncertainty quantification — correlated-parameter UQ, Sobol sensitivity analysis — see MUQ-SAC
  • Reduced-order modelling & scientific ML — physics-informed neural networks, neural ODEs, response-surface surrogates — see SciML
  • Simulation & CFD — Cantera, CHEMKIN-PRO, FlameMaster, OpenFOAM, an in-house CFD/multiphysics solver — see MOSIAC and PANSOPHIA
  • Experimental combustion — rig design and commissioning, gas chromatography, thermocouple instrumentation, gasification, computer-vision flame measurement — see gasification and flamelab
  • Optimization & reinforcement learning software — a from-scratch graph-IR workbench, cross-benchmarked solvers, REINFORCE — see dokimazo

Publications

Full publication list →

5 refereed journal articles, 1 published conference paper, 1 manuscript under review, and 3 accepted conference abstracts (64th Symposium (Japanese) on Combustion, Fukuoka, Nov 2026).

CV

View the full CV →