High-resolution atmospheric dataset

Problem

Raw model output carries systematic biases and is too coarse for site-level decisions.

What I did

I trained LSTM and CNN models to correct systematic bias in forecast data, and used CNN and Kriging to downscale climate projections to finer resolution for modelling, visualisation and scenario analysis.

Later, as Climate Modeling Analyst, I supervised a small team cross-validating these ML/DL models against statistical bias-correction methods. Both ran in production on Azure, where the team monitored and retrained them, and I led the move of the team’s codebase to Azure.

Result

Forecast precision improved by 18%, and the bias-correction models, chosen on cross-validated evidence, ran in production with monitoring and retraining.

Tools

Python · LSTM · CNN · Kriging · Azure