Case study 01 — Undergraduate thesis
Hazard zones in seconds, not simulation queues
ALOHA dispersion modeling answers one scenario at a time. For n-Hexane tank overfill, I built ML surrogates that answer 1,261 of them — then shipped the result as a live web app with an interactive plume map.
- Problem
- Consequence modeling for overfill scenarios is run case-by-case in ALOHA, which is too slow for wide-range screening or rapid emergency-response decisions.
- Approach
- Generated 1,215 full-factorial scenarios across six parameters (Pasquill stability class, wind, temperature, spill rate, duration, liquid thickness), extended to 1,261 with targeted out-of-distribution additions, then trained MLR, Random Forest, and ANN surrogates to predict red/orange/yellow LEL zone radii.
- Outcome
- The ANN surrogate reaches R² 0.9975 with a red-zone MAE of 1.82 m — 12× lower than the MLR baseline — deployed as a Flask web app with an ALOHA-style plume map, batch processing, and KML export, plus a BI dashboard for scenario analytics.
- 1,261 scenarios
- ANN R² 0.9975
- 3 LEL zones
- Live web app
ALOHA · Python · ANN · Random Forest · Flask · Leaflet
Try the live app on Hugging Face Spaces
