AI Applications in Climate Science and Energy (2025-2026)

Comprehensive Research Report

---

1. Breakthrough Weather Prediction Models

1.1 Google DeepMind: GraphCast and GenCast

GraphCast (deployed 2023, iterated through 2025) marked a paradigm shift in numerical weather prediction (NWP). Built on graph neural networks operating over a mesh representation of Earth, it produces 10-day global forecasts in under a minute on a single TPU – compared to hours on supercomputers for traditional models like ECMWF’s HRES (High Resolution forecast).

Key accuracy results:
- Outperformed HRES on 90% of 1,380 verification targets (variables x pressure levels x lead times)
- Particularly strong at 3-7 day lead times for tropospheric variables
- Notable improvements in tropical cyclone track prediction (lower mean track error at 3-5 day leads)

GenCast (published late 2024, operational testing 2025) represented the next evolution – a probabilistic/ensemble diffusion model. Unlike GraphCast’s single deterministic forecast, GenCast generates calibrated ensemble forecasts:
- Outperformed ECMWF’s ENS (ensemble) system on 97.2% of targets at lead times up to 15 days
- Superior probabilistic skill (CRPS metric) for extreme weather events
- Better calibrated uncertainty quantification – critical for decision-making
- Enabled probabilistic wind power forecasting 2-3 days earlier than ENS with equivalent reliability

Architecture: Conditional diffusion model on an icosahedral mesh, trained on ERA5 reanalysis data (1979-2018), generating 0.25-degree resolution forecasts.

1.2 Huawei: Pangu-Weather and Successors

Pangu-Weather (2023) was among the first AI weather models to match ECMWF HRES accuracy. It used a 3D vision transformer architecture with pressure-level-specific processing.

Successors and related developments through 2025:
- Pangu-Weather v2 iterations improved handling of precipitation and extreme events – initial versions were known to under-predict intensity of extremes due to regression-to-the-mean effects inherent in deterministic ML training
- Huawei continued work on higher resolution (0.1-degree) regional models for East Asia

1.3 NVIDIA: FourCastNet and CorrDiff

FourCastNet pioneered the use of Adaptive Fourier Neural Operators (AFNOs) for weather prediction, processing data in spectral space for computational efficiency.

CorrDiff (2024-2025) applied diffusion models to weather downscaling:
- Took coarse-resolution (25 km) forecasts and generated high-resolution (2-3 km) outputs
- Achieved 1000x speedup over traditional dynamical downscaling
- Preserved physically consistent fine-scale features (orographic precipitation, sea-breeze convergence)

1.4 ECMWF’s AI Integration: AIFS

The European Centre for Medium-Range Weather Forecasts developed its own Artificial Intelligence Forecasting System (AIFS), entering operational testing in 2024-2025:
- Graph transformer architecture trained on ECMWF’s own operational analyses (higher quality than ERA5 reanalysis used by most competitors)
- Designed for integration into ECMWF’s operational pipeline rather than as a standalone replacement
- Produced ensemble forecasts with physical consistency constraints
- ECMWF began issuing experimental AI-based forecasts alongside traditional ones

1.5 Accuracy Improvements and Limitations

What AI weather models do well:
- Medium-range (3-10 day) large-scale pattern prediction
- Tropical cyclone track forecasting
- Computational cost reduction: minutes vs. hours, single GPU/TPU vs. supercomputer
- Ensemble generation (GenCast, AIFS) at fraction of traditional cost

Persistent challenges through 2025:
- Intensity of extremes: Deterministic ML models tend to blur/smooth extreme values. Diffusion-based approaches (GenCast) partially address this but do not fully resolve it
- Precipitation: Remains harder to predict than temperature/pressure fields; convective precipitation especially challenging
- Physical consistency: AI models can produce fields that violate conservation laws (mass, energy) – an active area of research for hybrid approaches
- Sub-seasonal to seasonal (S2S): 2-6 week forecasts remain difficult; AI models show modest improvements but the predictability barrier is fundamental
- Training data dependence: Models trained on reanalysis inherit its biases; climate non-stationarity means historical training data may not represent future extremes

---

2. AI for Grid Optimization

2.1 The Grid Challenge

Modern electricity grids face a fundamental transformation: integrating variable renewable energy (wind, solar) at 30-60%+ penetration while maintaining reliability. This creates optimization problems of unprecedented complexity – balancing supply, demand, storage, and transmission across millions of nodes with uncertainty in both generation and consumption.

2.2 Deployed Systems and Key Players

Google DeepMind / Google:
- Applied AI to optimize Google’s data center energy use, achieving 30% reduction in cooling energy (deployed since 2016, continuously improved)
- Expanded to grid-scale: partnered with utilities to improve wind power value by ~20% through 36-hour ahead wind power forecasting, enabling better day-ahead market bidding
- By 2025, these approaches were being adopted more broadly

AutoGrid (acquired by Schneider Electric):
- AI-driven virtual power plant (VPP) platform managing distributed energy resources (DERs)
- Optimized dispatch of batteries, EVs, smart thermostats as aggregated grid resources
- Deployed with multiple US and European utilities

Utilidata:
- Partnered with NVIDIA to deploy AI chips at the grid edge (on transformers and distribution equipment)
- Real-time local optimization of voltage, power quality, and fault detection
- Edge AI approach addresses latency requirements that cloud-based solutions cannot meet

National Grid / UK System Operator:
- Deployed ML-based demand forecasting and optimal power flow calculations
- Reduced balancing costs and improved frequency response management
- AI-assisted constraint management for transmission bottlenecks

2.3 Technical Approaches

Reinforcement Learning for Grid Control:
- Deep RL agents trained to manage voltage control, frequency regulation, and economic dispatch
- L2RPN (Learning to Run a Power Network) competition series drove research progress
- Challenges: safety guarantees (RL agents can find adversarial states), sim-to-real transfer

Graph Neural Networks for Power Flow:
- Power grids are natural graph structures; GNNs approximate power flow solutions 100-1000x faster than Newton-Raphson solvers
- Enables real-time contingency analysis (N-1, N-2 security assessment) that was previously computationally prohibitive
- Active area: physics-informed GNNs that guarantee constraint satisfaction

Probabilistic Forecasting for Renewable Integration:
- Quantile regression, normalizing flows, and diffusion models for probabilistic load and generation forecasts
- Improved reserve margin calculations, reducing need for expensive spinning reserves
- Real impact: 5-15% reduction in operational reserves needed while maintaining reliability standards

2.4 Real-World Impact Metrics

• AI-optimized battery dispatch typically improves revenue 10-20% over rule-based systems
• Demand forecasting errors reduced from 3-5% MAPE to 1-3% MAPE for day-ahead horizons
• Predictive maintenance on grid assets reduces unplanned outages by 20-40% in deployment studies
• Dynamic line rating (using AI weather prediction) increases transmission capacity utilization by 10-30% without new infrastructure

---

3. AI for Carbon Capture Design

3.1 The Carbon Capture Landscape

Carbon capture, utilization, and storage (CCUS) encompasses point-source capture (from industrial emissions), direct air capture (DAC), and ocean-based approaches. AI is accelerating every stage: sorbent/solvent design, process optimization, and geological storage site characterization.

3.2 AI-Driven Sorbent and Solvent Discovery

Microsoft Research and Pacific Northwest National Laboratory (PNNL):
- In a landmark 2024 study, used AI to screen 32 million potential materials, narrowing to 18 top candidates for battery materials – the same methodology was being applied to CO2 sorbents
- Accelerated the traditional materials discovery pipeline from years to weeks

Machine Learning for Metal-Organic Frameworks (MOFs):
- MOFs are porous crystalline materials with enormous design space (millions of possible structures)
- Research groups at Georgia Tech, MIT, and EPFL trained graph neural networks and transformers to predict CO2 adsorption capacity, selectivity, and stability
- By 2025, ML-screened MOFs were entering lab synthesis and testing phases
- Key result: ML models could predict CO2 uptake with R-squared > 0.9, enabling virtual screening of 100,000+ candidates before any synthesis

Solvent Design for Amine Scrubbing:
- Traditional amine-based capture (MEA) is energy-intensive due to regeneration heat
- ML models (molecular fingerprint-based, graph neural networks on molecular structure) predicted novel amine blends with 15-25% lower regeneration energy
- Carnegie Mellon and academic groups published optimization frameworks combining molecular ML with process simulation

3.3 Process Optimization

Digital Twins for Capture Plants:
- Physics-informed neural networks (PINNs) used to create real-time digital twins of absorption/desorption columns
- Enabled model-predictive control that adapts to varying flue gas composition and flow rates
- Reported 5-10% energy savings over conventional PID control in pilot plant studies

Direct Air Capture Optimization:
- Climeworks (solid sorbent DAC) and Carbon Engineering/1PointFive (liquid solvent DAC) both incorporated ML into operational optimization
- Key challenge: DAC is extremely energy-intensive (~6-9 GJ/tonne CO2); any efficiency improvement has outsized impact on cost
- AI-optimized cycling (adsorption-desorption timing, temperature swings) based on real-time ambient conditions

3.4 Geological Storage and Monitoring

• ML-based seismic interpretation for identifying suitable CO2 storage reservoirs
• Physics-informed ML for predicting CO2 plume migration underground (replacing expensive numerical simulation)
• Anomaly detection on monitoring data (pressure, seismic, satellite) for leakage detection
• Stanford and UT Austin groups published surrogate models that were 10,000x faster than physics simulators while maintaining accuracy for reservoir-scale predictions

---

4. AI for Fusion Energy Modeling

4.1 The Fusion Control Challenge

Magnetic confinement fusion (tokamaks, stellarators) requires maintaining a plasma at 100+ million degrees while controlling instabilities that can terminate the reaction (disruptions) in milliseconds. This is fundamentally an AI-amenable control problem.

4.2 DeepMind and EPFL: Tokamak Plasma Control

Landmark work (published 2022, continued through 2025):
- Deep RL agents learned to control the TCV tokamak at EPFL, Switzerland
- Controlled 19 magnetic coils simultaneously to shape and position plasma
- Achieved plasma configurations that had never been produced before, including a droplet-shaped configuration theoretically proposed but never experimentally realized
- Demonstrated the ability to maintain multiple simultaneous plasma configurations in sequence

Subsequent developments (2024-2025):
- Extended to disruption prediction and avoidance, not just shape control
- Integration with real-time diagnostic interpretation (ML processing Thomson scattering, interferometry data)
- Work toward transfer learning: training on one tokamak and deploying on another

4.3 Disruption Prediction

FRNN (Fusion Recurrent Neural Network) and successors:
- Princeton Plasma Physics Laboratory (PPPL) led development of LSTM and transformer-based disruption predictors
- Trained on large databases from DIII-D (General Atomics), JET (UK), and other tokamaks
- Achieved >95% detection rate with <5% false alarm rate at 30ms+ warning time on tested machines
- Challenge: cross-machine generalization remains difficult – disruption physics varies between devices

Relevance to ITER:
- ITER (under construction in France, targeting first plasma mid-to-late 2020s) will have disruptions carrying enormous energy (hundreds of MJ)
- Disruption prediction and mitigation systems are safety-critical
- AI-based predictors being developed and validated as part of ITER’s disruption mitigation system

4.4 Stellarator Optimization

Wendelstein 7-X (Max Planck Institute):
- Stellarators have inherently more complex 3D magnetic geometries than tokamaks
- AI used for:
- Coil design optimization (reducing engineering complexity while maintaining plasma confinement)
- Real-time equilibrium reconstruction from diagnostic data
- Turbulence prediction using neural network surrogates for gyrokinetic codes

Private Fusion Companies:
- TAE Technologies: Used ML extensively for optimizing their field-reversed configuration approach; Google collaboration on Optometrist Algorithm (human-in-the-loop Bayesian optimization)
- Commonwealth Fusion Systems (CFS): Applied ML to high-temperature superconducting magnet design and plasma scenario modeling for SPARC tokamak
- Helion Energy: Used ML for pulsed fusion optimization

4.5 Simulation Acceleration

Traditional fusion plasma simulation (gyrokinetic codes like GENE, GS2) is extraordinarily expensive – a single simulation can take millions of CPU-hours. AI surrogates:
- Neural network emulators trained on simulation databases predict turbulent transport 1,000-10,000x faster
- Enable optimization loops (plasma scenario design) that would be impossible with physics codes alone
- Physics-informed approaches ensure predictions remain within physically valid regimes
- Active learning strategies efficiently explore parameter space

---

5. AI for Renewable Energy Forecasting

5.1 Solar Power Forecasting

Intra-hour (Nowcasting):
- Sky-imager-based models using CNNs to track cloud motion and predict irradiance 5-30 minutes ahead
- Satellite-based models (geostationary satellites) for 1-6 hour horizons
- Achieved rRMSE improvements of 20-40% over persistence models

Day-ahead:
- Transformer architectures processing NWP outputs, satellite data, and historical generation
- Probabilistic forecasts using quantile regression or conformal prediction for uncertainty bands
- State of the art by 2025: 5-10% NMAE for day-ahead solar forecasts in well-instrumented regions

Degradation and Soiling:
- ML models predicting panel degradation rates and soiling losses from environmental data
- Optimizing cleaning schedules – reducing O&M costs by 3-8% while maintaining output

5.2 Wind Power Forecasting

Turbine-Level:
- Wake effect modeling using physics-informed ML – predicting how upstream turbines affect downstream performance
- Digital twins of individual turbines detecting performance degradation
- Yaw optimization using RL: 1-3% annual energy production increase by better aligning with wind direction

Farm-Level and Portfolio:
- Spatial-temporal models capturing correlation between geographically distributed wind farms
- Essential for system operators managing wind-heavy grids
- GenCast (see Section 1) demonstrated 2-3 day improvement in reliable wind power probability forecasts

Offshore Wind:
- Particularly high-value application due to larger turbines, higher capacity factors, and more complex meteorology
- AI-based wave and wind forecasting for installation vessel operations (reducing weather downtime)
- Structural health monitoring of foundations using ML on sensor data

5.3 Grid-Scale Integration

Ramp Event Detection:
- Sudden drops or increases in renewable generation (cloud fronts, wind shifts) are grid stability threats
- ML classifiers detecting approaching ramp events 1-4 hours ahead with 70-85% accuracy
- Enables pre-positioning of reserves

Hybrid Forecasting:
- Combining physics-based NWP with ML post-processing consistently outperforms either alone
- Multi-model ensemble approaches using gradient boosting or neural networks to optimally blend forecasts
- Operational at multiple system operators worldwide by 2025

5.4 Deployed Systems

• Xcel Energy: Partnered with NCAR to deploy ML-enhanced wind forecasting, reducing integration costs
• ERCOT (Texas): ML-based solar and wind forecasting integrated into grid operations
• Danish Energy Agency / Energinet: Advanced probabilistic wind forecasting for a 50%+ wind penetration grid
• AEMO (Australia): AI-assisted forecasting for managing one of the world’s most renewable-heavy grids
• Google/DeepMind: Deployed wind forecasting increasing wind energy value by ~20% for Google’s power purchase agreements

---

6. Cross-Cutting Themes and Outlook

6.1 Foundation Models for Earth Science

A major trend in 2024-2025 was the development of foundation models for Earth/climate science – large pretrained models that can be fine-tuned for multiple downstream tasks:

• Aurora (Microsoft Research, 2024): A foundation model for atmospheric science, trained on over 1 million hours of diverse weather and climate data at different resolutions; demonstrated state-of-the-art performance on weather forecasting, air quality prediction, and ocean wave modeling from a single pretrained backbone
• ClimaX (Microsoft/UCLA): A vision transformer pretrained on CMIP6 climate simulation data, fine-tunable for forecasting, downscaling, and climate projection
• Prithvi (NASA/IBM): Geospatial foundation model for Earth observation, applicable to flood mapping, wildfire detection, crop monitoring

6.2 AI for Climate Projections

Distinct from weather prediction (days) is climate projection (decades):
- ML-based climate model emulators can approximate the output of expensive Earth System Models 1,000-1,000,000x faster
- Enables ensemble exploration of emission scenarios
- Hybrid approach: ML parameterizing sub-grid processes (clouds, convection) within physics-based climate models
- ClimateSet and similar benchmark datasets emerging to standardize ML-for-climate research

6.3 Persistent Challenges

1. Data quality and availability: Many energy applications lack labeled training data, especially for rare/extreme events
2. Physical consistency: Pure ML models can violate conservation laws; physics-informed approaches help but add complexity
3. Interpretability: Grid operators and fusion scientists need to understand why an AI makes a recommendation, not just what it recommends
4. Distribution shift: Climate change means historical data may not represent future conditions – a fundamental challenge for data-driven models
5. Equity and access: Advanced AI weather models require significant compute and data infrastructure, potentially widening the gap between well-resourced and under-resourced meteorological services

6.4 Impact Summary

Domain	AI Impact (2025)	Key Metric
Weather Prediction	Operational at major centers	Matches/exceeds NWP at 1/1000th compute cost
Grid Optimization	Widely deployed	10-20% efficiency gains in dispatch/reserves
Carbon Capture	R&D/pilot stage	100-1000x acceleration in materials screening
Fusion Energy	Experimental validation	RL plasma control demonstrated; disruption prediction >95%
Renewable Forecasting	Broadly operational	20-40% error reduction over baseline methods

---

7. Key References and Sources

The findings above synthesize information from the following landmark publications and programs:

• Lam et al., “Learning skillful medium-range global weather forecasting” (Science, 2023) – GraphCast
• Price et al., “GenCast: Diffusion-based ensemble forecasting for medium-range weather” (Nature, 2024) – GenCast
• Bi et al., “Accurate medium-range global weather forecasting with 3D neural networks” (Nature, 2023) – Pangu-Weather
• Degrave et al., “Magnetic control of tokamak plasmas through deep reinforcement learning” (Nature, 2022) – DeepMind/EPFL tokamak control
• Bodnar et al., “Aurora: A Foundation Model of the Atmosphere” (2024) – Microsoft Aurora
• ECMWF AIFS documentation and operational bulletins
• US DOE Fusion Energy Sciences reports on ML for fusion
• IEA reports on digitalization of energy systems

---

This report covers the state of play through early-to-mid 2025, which is the extent of my verified knowledge. The field is advancing rapidly; developments in late 2025 and into 2026 may include further operationalization of AI weather models, first results from ITER-relevant disruption mitigation AI, and potentially the first commercially deployed AI-discovered carbon capture sorbents at pilot scale.