01
Research · 1st Author
Attributing Extreme Precipitation to Global Warming via AI
A FiLM-conditioned CNN that preserves a causal intervention pathway for GMT, enabling counterfactual attribution of extreme rainfall across the contiguous United States.
, infrastructure failure, and major economic losses. Yet attributing individual extreme rainfall events to anthropogenic warming has remained especially difficult: precipitation arises from complex multiscale interactions among moisture, circulation, and topography, and conventional statistical models rarely capture the heavy upper tail — precisely where the most societally consequential events live.
The Core Challenge: Predictor Redundancy Threatens Causal Validity
ML-based counterfactual attribution requires a warming variable (GMT) that can be meaningfully manipulated. But even after linearly removing the GMT signal from atmospheric predictors, an auxiliary CNN can still recover GMT with R² ≈ 0.77 from the residual fields. This means the network can infer warming from correlated dynamical patterns — functionally bypassing the explicit intervention variable. Good prediction alone is not enough: if GMT loses its independent causal pathway inside the model, counterfactual experiments become uninterpretable.
What I Built
I designed a FiLM-conditioned convolutional neural network trained on the 22-member CESM2 Large Ensemble (1850–2100) and fine-tuned on ERA5 reanalysis. The central architectural choice: Feature-wise Linear Modulation (FiLM) layers at every convolutional block, which give GMT a dedicated pathway to directly modulate intermediate spatial features — applying learned affine transformations to each channel — before spatial pooling collapses the representation. This preserves GMT as a causally valid intervention variable within a high-capacity nonlinear model. The framework covers 15 CONUS boxes spanning 30°–45°N and 75°–125°W, focused on November–March synoptically driven winter extremes.
- FiLM-conditioned CNN (PyTorch)
- 22-member CESM2-LE training pipeline
- ERA5 transfer + fine-tuning
- Hybrid KDE–GPD tail modeling
- 15-box CONUS spatial framework
- Counterfactual GMT sweeps (ΔT = 0 to +4°C)
Attribution Framework: Beyond Risk Ratios
Rather than relying solely on conventional metrics like the fraction of attributable risk (FAR), I incorporated probabilities of necessary and sufficient causation (PN and PS), computed as continuous functions of the exceedance threshold. This distinction is critical for precipitation extremes: as event severity increases, the causal role of warming shifts — from sufficient causation at common thresholds (warming alone can raise risk) toward necessary causation at rare-event thresholds (warming must be present for the event to reach that intensity). The upper tail is modeled via a hybrid KDE–GPD approach for stable, threshold-resolved estimation.
Key Findings
- R² = 0.86–0.98 on CESM2; 0.84–0.93 after ERA5 fine-tuning
- 15–20% precipitation intensity increase at +4°C warming
- PN ≈ 0.3 near historical event magnitudes, exceeding 0.6 at higher thresholds
- PS declines sharply with severity → warming is necessary but rarely sufficient
- Stronger amplification in the eastern US, consistent with greater moisture availability
Why This Matters
This work makes both a domain contribution and a methodological one. For climate attribution, it extends ML-based counterfactual approaches from heat extremes to precipitation — a far messier target — and provides the first PN/PS causal decomposition of extreme rainfall events over CONUS. Methodologically, it demonstrates a principle that extends well beyond precipitation: in ML-based attribution, architecture directly determines whether the intervention variable remains causally interpretable. The FiLM-conditioned design offers a blueprint for any attribution setting where the causal signal is dynamically mediated rather than thermodynamically direct.
Good predictive skill is necessary but not sufficient for causal attribution. The architecture must preserve the intervention pathway — otherwise counterfactuals lose their meaning.