Physics-aware Causal DL Attribution

Overview

The goal is not only to predict precipitation extremes, but to quantify how key drivers (e.g., temperature forcing) contribute to extreme-event risk. The work emphasizes interpretable attribution outputs and robust, reproducible evaluation across multiple regions/blocks.

What I Built

• An end-to-end pipeline: data → preprocessing → training → attribution → robustness checks → reproducible artifacts.
• Attribution metrics based on fixed factual extreme sets and counterfactual responses (risk curves / CCDF-style views).
• Batch visualization and case studies for individual extreme events (sequence alignment and model responses).

Method (high level)

• Define a factual “extreme set” using quantile thresholds (e.g., top tail of observed precipitation).
• Train a deep model to learn the mapping from climate variables to precipitation outcomes.
• Evaluate distributional changes under counterfactual settings to produce attribution conclusions.

Results (replace with your numbers)

• Key plots: CCDF/risk curves, attribution curves, sensitivity analyses (add screenshots).
• Comparisons: different blocks/thresholds/training settings and stability over time.
• One-sentence takeaway: how risk changes as a key driver is perturbed.

Artifacts

• 📄 Paper/report: TODO
• 💻 Code repository: TODO
• 🖼️ Figures: TODO (2–4 per page works well)

Next Steps

• Standardize the attribution interface: (region, threshold) → curves + uncertainty.
• Expand robustness: dataset splits, alternate reanalysis/model sources, and sensitivity testing.