This research pushes boundaries by fusing AI’s knack for spotting nonlinear links in massive datasets with the solid physics of climate models, proving we can boost reliable forecasts of extreme event stats—like how often or how bad they get—over subseasonal to decadal horizons.
Why Forecasting Climate Extremes Matters
In our warming world, ramped-up extremes like scorching heatwaves, brutal droughts, and deluge storms hammer societies, economies, and ecosystems hard. With human-driven climate shifts accelerating, spot-on predictions weeks to decades out are crucial now for engineers, policymakers, and industries to build resilience and slash risks from these chaotic beasts.
Before this review, nailing down climate extreme forecasts at subseasonal to decadal (S2D) scales was a tough nut to crack, thanks to their rarity, chaotic vibes, and gaps in dynamical models like CMIP6, which often lowballed the system’s built-in predictability beyond short-term weather. Key hurdles included faster anthropogenic forcing since the 2000s, timescale-specific drivers needing tailored extreme definitions, skimpy historical event counts and observations, and dodgy handling of feedbacks in numerical setups. Weather forecasts under 10 days had exploded with AI wins, matching top dynamical systems even for wild outliers, but S2D lagged due to way fewer training samples—think aggregating event stats over seasons or years instead of pinpointing one-offs. Efforts like the World Climate Research Programme’s Explaining and Predicting Earth System Change pushed for process-deep dives via experiments, while Earth observation booms (e.g., Copernicus, Sentinel sats) unleashed big data floods, sparking AI in geosciences to mine nonlinear ties without rigid physics assumptions. Still, AI for extreme predictions got short shrift, with prior overviews leaning more on attribution or traits than S2D links to drivers.
Research Question & Research Goal
This review tackles: How can AI amp up predictions of climate extremes at S2D scales and uncover ties to big-picture and local drivers? It stands out by surveying fresh AI apps for extreme forecasts, eyeing hybrid blends of data-driven empirics and dynamical physics, and hashing out data-method hurdles—arming climate pros with a solid playbook for sound future work.
The Innovative Step
This review energizes the field by dissecting AI’s role in forecasting top extremes—temps (heatwaves/cold snaps), droughts, cyclones, and heavy rains—at S2D scales, spotlighting hybrids as the smart bridge between AI’s pattern-hunting prowess and models’ physics backbone. It breaks down AI tools like random forests (RF), extreme gradient boosting (XGBoost), convolutional neural nets (CNNs), and long short-term memory (LSTMs), showing how they crush benchmarks in studies via spatiotemporal smarts. For temps, RFs with explainable AI (XAI) like SHAP unpack drivers from air, ocean, and land, outpacing climatology or trends, while NNs trained on sims tackle data droughts. Droughts get a boost from ANNs, LSTMs, and ELMs for indices like SPI, with ELMs shining for speed and accuracy. Cyclones and rains leverage CNNs and hybrids for better hits on counts and punch, often via downscaling. The real spark? Hybrid setups: coupled for subgrid tweaks (e.g., convection params), serial for bias fixes and downscaling, and statistical-dynamical with transfer learning from big ensembles to generalize. It flags persistent pains in data prep, uncertainties, out-of-distribution flops, reproducibility, and black-box vibes, pitching best practices like transparent code-sharing, multi-metric benchmarks, out-of-sample validation, uncertainty quant, and physics-guided designs to earn trust. Looking ahead, it envisions transfer learning for rare decadal events, probabilistic twists, and compound extreme forecasts.
Results & Key Insights
Across extremes, AI methods like RFs, XGBoost, CNNs, and LSTMs often top climatology, persistence, and dynamical baselines in metrics such as Brier skill score (BSS), Matthews correlation coefficient (MCC), and root mean square error (RMSE). Temp studies show RFs predicting heatwave odds up to 6 weeks out, with XAI highlighting soil moisture or stratospheric roles; NNs handle imbalances via undersampling for French heatwaves. Drought forecasts via ELMs and LSTMs nail SPI/SPEI months ahead, outperforming ANNs in speed. For cyclones/rains, CNNs with transfer learning boost seasonal TC counts, while hybrids refine precip via causal nets. Core takeaways: Data scarcity demands model sims for training; XAI spots links but needs causality checks; hybrids unlock skill jumps, yet rarity sparks verification woes like sampling errors.
Looking forward
This unlocks probabilistic models, compound event forecasts, and physics-tied AI for climate services in farming, infrastructure, and risk planning. Societies, engineers, and environments gain from proactive defenses. Building on it could mean transfer learning for decadal rarities, shared benchmarks for collab, and open workflows to speed progress.
As extremes intensify, blending AI’s ingenuity with climate know-how opens exciting doors to smarter, tougher systems against nature’s curveballs. Hats off to the authors for this insightful overview! Dive deeper: Materia, S., Palma García, L., van Straaten, C., O, S., Mamalakis, A., Cavicchia, L., Coumou, D., de Luca, P., Kretschmer, M., & Donat, M. (2024). Artificial intelligence for climate prediction of extremes: State of the art, challenges, and future perspectives. WIREs Climate Change, 15(6), e914.