AI in Climate Modelling and Carbon Accounting: From Forecasts to Verifiable Emissions

Climate change is, among other things, a data problem. We need to know what the atmosphere will do next week, what a region's climate will look like in thirty years, how much carbon a company, a city or a forest is actually emitting or absorbing, and where the next flood is likely to hit. The IPCC Sixth Assessment Report makes clear that both mitigation and adaptation depend on reliable information at scales that matter for decisions.

As founder and CTO of StrikeClimate, I spend a lot of time thinking about the gap between what climate science knows and what a local government, a farmer or a business can act on. Traditional climate and weather models are extraordinarily sophisticated, but they are computationally expensive and often too coarse for local decisions. Carbon accounting, meanwhile, still relies heavily on self-reported estimates and generic emission factors. AI is beginning to change both.

Learned weather models

Numerical weather prediction solves physical equations on supercomputers. A new generation of machine learning models instead learns the dynamics of the atmosphere directly from decades of reanalysis data such as ECMWF's ERA5. DeepMind's GraphCast, described in Science, uses a graph neural network over a mesh of the globe and was reported to outperform a leading operational deterministic system on a large majority of verification targets for medium-range forecasts. Huawei's Pangu-Weather, published in Nature, uses a 3D transformer architecture. Once trained, these models produce a global forecast in minutes on modest hardware rather than hours on a supercomputer.

Downscaling and emulation

Climate projections from global models have grid cells tens of kilometres wide. Deep learning downscaling, often using convolutional networks or diffusion models, translates these into finer-resolution local projections, learning the relationship between coarse fields and high-resolution observations. Emulators approximate expensive model components so that scientists can run many more scenarios and quantify uncertainty better.

Hybrid physics and ML

The most promising direction is hybrid: keep the physical core of a model but replace poorly understood or expensive sub-processes, such as cloud parameterisations, with learned components constrained by physics. This addresses a common worry that pure ML models might violate conservation laws or behave unpredictably outside their training climate.

Measurement, reporting and verification (MRV) is the backbone of any credible climate commitment. AI helps in several ways:

• Remote emissions monitoring. Satellite observations combined with machine learning estimate emissions from power plants, oil and gas facilities, shipping and land-use change. The Climate TRACE coalition publishes facility-level emissions estimates derived largely this way, offering an independent check on reported inventories.
• Forest carbon estimation. Models combining optical imagery, radar and lidar estimate above-ground biomass, which is essential for verifying forest carbon projects whose credibility has been widely questioned.
• Corporate Scope 3 estimation. Supply chain emissions are notoriously hard to measure. Machine learning maps transaction and procurement data to emission factors and fills gaps, aligned with frameworks such as the Greenhouse Gas Protocol.
• Energy optimisation. Forecasting demand and renewable output allows grids and buildings to cut emissions by shifting loads and reducing curtailment.

The influential paper Tackling Climate Change with Machine Learning by Rolnick and colleagues remains one of the best maps of where ML can contribute, from electricity systems to carbon removal, and I recommend it to every engineer moving into climate work.

Climate is a domain where overclaiming is particularly harmful.

• Extrapolation. ML weather models are trained on the past climate. Their behaviour under unprecedented extremes is an open research question, which is exactly where the stakes are highest.
• Data gaps in the Global South. Weather station coverage across much of Africa is sparse, which weakens both training data and local verification. Models may look accurate on global averages while underperforming where people are most vulnerable.
• Greenwashing risk. An AI-generated emissions estimate with no stated uncertainty can be used to make a company look better than it is. Estimates should always come with error bars and methodology.
• The footprint of AI itself. Training and serving large models consumes significant electricity and water. Climate teams should account for their own compute emissions and choose efficient architectures and cleaner regions where possible.

A model that cannot tell you how uncertain it is should not be used to make a climate claim.

1. Build on open foundations. ERA5 reanalysis, Copernicus satellite data, NASA Earth data and open-source ML weather models mean you do not need a supercomputer to start. Focus your effort on the local last mile.
2. Invest in local observations. Low-cost weather stations, air quality sensors and community reporting improve calibration where official data is thin. My interest in embedded systems is partly driven by this: cheap, reliable sensors are climate infrastructure.
3. Translate forecasts into decisions. A farmer needs to know whether to plant this week; a city needs to know which drainage channels to clear before a storm. Design outputs around those decisions, not around raw model fields.
4. Quantify and communicate uncertainty. Use ensembles or probabilistic outputs and present them in plain language.
5. Make MRV auditable. Store inputs, model versions and outputs so that any emissions number can be traced and reproduced.
6. Partner with meteorological agencies. In Nigeria, NiMet and similar agencies elsewhere hold historical data and public mandates; collaboration beats competition.

Operational forecasting centres are now running ML models alongside traditional ones, and foundation models for Earth system data are emerging. For carbon accounting, the combination of more satellites dedicated to greenhouse gas monitoring and better AI inversion methods means that, increasingly, emissions will be observed rather than merely reported. That shift, from trust to verification, could be one of the most important contributions AI makes to climate action.

For builders in Africa, the opportunity is to make these global advances useful locally: forecasts in local languages, flood warnings that reach informal settlements, and MRV systems that let African carbon projects prove their integrity.

• ML weather models such as GraphCast and Pangu-Weather now produce competitive medium-range forecasts at a fraction of the compute cost.
• Hybrid physics-ML approaches and deep learning downscaling make climate information more local and actionable.
• Satellite-based AI is moving carbon accounting from self-reporting toward independent verification.
• Always quantify uncertainty; extrapolation to unprecedented extremes and sparse African data are real limits.
• Climate tech teams should build on open data, invest in local sensing and design outputs around concrete decisions.

1. AR6 Synthesis Report: Climate Change 2023 — IPCC
2. Learning skillful medium-range global weather forecasting — Science
3. Accurate medium-range global weather forecasting with 3D neural networks — Nature
4. Tackling Climate Change with Machine Learning — arXiv
5. Climate TRACE — Climate TRACE
6. Greenhouse Gas Protocol — GHG Protocol

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.