How AI Is Fighting Climate Change: From Power Grids to Carbon Removal

I'm Parikshet. Climate change is the biggest problem my generation inherits. And AI is one of the most powerful tools we have to address it — alongside policy, behaviour change, and investment in clean technology. I want to be honest about what AI can and cannot do here, because overblown claims about "AI will save the climate" are just as unhelpful as dismissing the technology entirely.

The Grid Problem: Matching Supply and Demand

Fossil fuel power plants have a useful property: you can burn more coal or gas when demand increases. Solar panels and wind turbines do not work that way — they produce when conditions are right, not when you need more power. Managing a grid with high renewable penetration requires predicting both supply (when will the wind blow?) and demand (when will people charge their electric cars?) hours or days ahead.

Google's DeepMind worked with the UK's National Grid to develop AI models that predict wind farm output 36 hours ahead, enabling grid operators to sell renewable power on the day-ahead market rather than the more expensive intraday market. In 2019, DeepMind also used reinforcement learning to reduce the energy used to cool its own data centres by 40%, applied to cooling systems managing thousands of variables simultaneously.

GraphCast: AI Weather That Beats the Old Models

Weather prediction has been dominated for decades by numerical weather prediction — supercomputers solving fluid dynamics equations across global atmospheric grids. It works well but requires massive computing resources and hours of processing per forecast.

Google DeepMind's GraphCast (2023) was trained on 40 years of weather data. It produces 10-day global forecasts in under one minute — 1,000 times faster than traditional models, on hardware that costs far less. It outperformed the European Centre for Medium-Range Weather Forecasting (ECMWF), the global benchmark, on 90% of forecast variables including temperature, precipitation, and wind speed.

Better weather forecasting improves renewable energy scheduling, disaster preparedness, agricultural planning, and supply chain resilience — all with significant climate implications.

Discovering New Materials at Machine Speed

Solar panels are made of silicon. Better, cheaper materials could make solar power more accessible globally. Batteries store energy in lithium-ion cells. Better chemistries could make battery storage cheaper and longer-lasting. Finding these materials traditionally takes decades of laboratory experiments.

AI-driven molecular simulation screens millions of candidate compounds by predicting their properties computationally — without physically synthesising each one. DeepMind's GNoME (Graph Networks for Materials Exploration), released in 2023, discovered 2.2 million new stable crystal structures — including 380,000 that could be manufactured. For comparison, the entire history of human materials discovery before GNoME had found around 48,000 stable inorganic compounds. This is a 45× expansion of the known material design space.

The Honest Limitations

AI cannot make political decisions to reduce emissions. It cannot override economies that profit from fossil fuels. It cannot replace the investment in infrastructure needed to deploy clean energy at scale.

And AI itself has a carbon cost. Training GPT-3 emitted an estimated 552 tonnes of CO₂ — equivalent to driving a car 700,000 km. Google, Microsoft, and OpenAI have all committed to running on renewable energy, but current AI training and inference consumes substantial power. The net climate impact of AI — benefits from optimisation vs. energy consumed — is a genuine research question without a clear answer yet.

My position: AI is a powerful accelerant for climate solutions that humans have already chosen to pursue. It makes clean energy grids more efficient, material discovery faster, and weather prediction more accurate. But it does not replace the hard political and economic decisions that climate solutions ultimately require.