Skip to content

Artificial intelligence

Five research results where artificial intelligence meets science: a language model that designs proteins, a model that proposes new materials, a hybrid climate model, a simulator for fusion machines, and a controller that keeps a fusion plasma stable. Models and machines aimed at Earth systems, materials, grids, proteins, and recycling, along with methane and fire observation, flood forecasts, and land maps from 2024 on. Weather models depend on satellites.

GOES-16 full-disk image of Earth: the observations AI weather models train and run on

Artificial intelligence Deployed

AI weather models: GraphCast

A graph neural network now rivals classical medium-range forecasts, and it depends on satellite observations.

Source: Lam et al., Science, 2023: GraphCast

A quartz crystal: the kind of ordered solid materials models try to explore faster than trial-and-error

Artificial intelligence Research

GNoME: AI for materials

A model proposed hundreds of thousands of crystals that may be stable. Making them is work for the laboratory.

Source: Merchant et al., Nature, 2023: GNoME

High-voltage transmission pylons: the physical grid that optimization software sits on

Artificial intelligence Pilot

AI for the power grid

Better wind forecasts and dispatch tools shave reserves. The transformers and the rest of the physical grid stay in place.

Source: Google DeepMind: wind-energy value (26 Feb 2019)

A protein ribbon diagram on a dark field: the class of structure these models predict

Artificial intelligence Deployed

AlphaFold 3 in use

A 2024 model of proteins and their partners is now everyday laboratory software, used by millions of researchers. A working climate enzyme is a later laboratory step.

Source: Google DeepMind: AlphaFold

An industrial robot arm sorting a PET bottle on a recycling conveyor

Artificial intelligence Deployed

Recycling robots: AMP

Cameras and arms already pick cartons and bottles on real sorting lines. A buyer for the sorted material is a separate question.

Source: AMP Sortation

An IBM Q System One-class quantum computer, the chandelier of cables over a dilution refrigerator

Artificial intelligence Research

Quantum computing: chemistry hopes

Public machines from IBM, Google, and IonQ keep adding qubits. Designing a climate catalyst with them is later work.

Source: Google: Willow quantum chip (9 Dec 2024)

An oil refinery at dusk: the class of industrial site emissions inventories try to measure from space

Artificial intelligence Deployed

Climate TRACE v5.10

The August 2026 inventory put first-half global greenhouse gases 0.2 percent above 2025: a satellite-and-model estimate, not a court verdict.

Source: Climate TRACE: v5.10 / first-half 2026 emissions (27 Aug 2026)

GOES-17 full-disk image of Earth: the class of satellite observations WeatherNext 3 trains and runs on

Artificial intelligence Deployed

Google WeatherNext 3

A September 2026 global weather model that ingests live satellites and refreshes every hour: still not your national warning service.

Source: Google DeepMind: Introducing WeatherNext 3 (3 Sep 2026)

The European Centre for Medium-Range Weather Forecasts building in Reading: home of AIFS, not a screenshot of a forecast chart

Artificial intelligence Deployed

ECMWF AIFS 1.1

Europe’s operational AI forecast, updated in August 2025 with physics bounds so rain cannot go negative: running beside, not instead of, the physics model.

Source: VAST Data: The weather model that taught itself physics (AIFS 1.1)

Apollo 17 Blue Marble: the planet that Earth-observation foundation models are built to read

Artificial intelligence Research

Prithvi: Earth foundation models

NASA and IBM trained an open model on years of Harmonized Landsat-Sentinel imagery. It is a starting weight, not an oracle.

Source: NASA Earthdata: Prithvi geospatial model (2023)

Carbonized Herculaneum papyrus fragments (PHerc. 1103 and PHerc. 110): the class of sealed rolls Vesuvius Challenge reads without opening

Artificial intelligence Research

Vesuvius Challenge: sealed scrolls

X-ray CT and machine learning recovered first passages in Feb 2024 and a full Herculaneum scroll in June 2026: a prize contest, not magic OCR of all antiquity.

Source: Vesuvius Challenge

Suomi NPP night lights over the Permian Basin in May 2024: gas flares in an oil-and-gas region MethaneSAT measured, not a MethaneSAT scene

Artificial intelligence Pilot

MethaneSAT: oil & gas methane from space

EDF’s MethaneSAT (launched 4 Mar 2024; lost contact 20 Jun 2025) mapped area + point methane over ~45 basins: data and ACP 2026 science remain public. Not Climate TRACE’s inventory product and not a weather model.

Source: MethaneSAT: successful launch (4 Mar 2024)

Wildfires in Portugal and Spain seen from orbit: the class of fire FireSat is built to flag earlier, not a FireSat frame

Artificial intelligence Pilot

FireSat: wildfire early detection from space

Earth Fire Alliance / Muon Space / Google Research FireSat: Protoflight Mar 2025; first three operational sats launched 7 Jul 2026 (Transporter-17). AI compares 5×5 m IR spots for early fire flags: not a weather FM and not Climate TRACE.

Source: Earth Fire Alliance: first three operational FireSats reach orbit (7 Jul 2026)

Landsat and MODIS view of Rufiji River flooding in Tanzania, April 2024: a riverine flood of the kind Flood Hub forecasts, not a Flood Hub screenshot

Artificial intelligence Deployed

Google Flood Hub: AI river flood forecasts

Google’s Flood Hub (Nature, 20 Mar 2024; coverage expanded Nov 2024 to 100+ countries / ~700M people) issues free AI riverine forecasts up to seven days ahead: hydrology, not GraphCast-style global weather.

Source: Nearing et al., Nature: global flood prediction (20 Mar 2024)

Landsat view of farmland in Taiwan: the kind of land pattern geospatial embeddings summarize, not an AlphaEarth rendering

Artificial intelligence Deployed

AlphaEarth Foundations: Earth embeddings

Google DeepMind’s AlphaEarth Foundations (30 Jul 2025) turns multi-sensor Earth data into 64-D embeddings at 10 m; annual Satellite Embedding V1 is in Earth Engine. A mapping foundation: not Prithvi HLS weights and not a weather forecast.

Source: Google DeepMind: AlphaEarth Foundations (30 Jul 2025)

Illustrative stock photograph of a crystal jelly (Aequorea victoria), the jellyfish species in which green fluorescent protein was first found.

Artificial intelligence Research

ESM3 protein language model

In January 2025 the journal Science published ESM3, an artificial intelligence model from the company EvolutionaryScale that reads and writes proteins. Asked to design a glowing protein, it produced one that matches its closest natural relative at only 58 percent of positions, a gap the authors compare to about 500 million years of evolution.

Source: Europe PMC

Illustrative stock photograph of a natural quartz crystal cluster from Tibet, shown as an everyday example of a crystal.

Artificial intelligence Research

MatterGen materials design

On 16 January 2025 the journal Nature published MatterGen, an artificial intelligence model from Microsoft that proposes new crystals for a chosen property, designing them from scratch. One proposed material was made in a laboratory, and its resistance to being squeezed, estimated from laboratory tests, came within 20 percent of the 200 gigapascal target.

Illustrative stock photograph of Earth’s atmosphere and a crescent Moon taken from the International Space Station in 2006.

Artificial intelligence Research

NeuralGCM hybrid climate model

On 22 July 2024 the journal Nature published NeuralGCM, a Google Research model that keeps the physics of large-scale air movement and uses a neural network for clouds and other small-scale processes. Run over 40 years of the past, its temperature error was 0.25 degrees Celsius, against 0.75 for standard atmosphere-only models, and it ran more than 3,500 times faster than a detailed physics model.

Illustrative stock photograph of the inside of the Joint European Torus, a tokamak in England and a different machine from SPARC.

Artificial intelligence Research

TORAX fusion plasma simulator

In May 2024 Google DeepMind released TORAX, a fast open-source simulator of the hot gas inside a fusion machine. On 16 October 2025 Google DeepMind and Commonwealth Fusion Systems announced a partnership to use it, together with artificial intelligence, to plan how their SPARC machine will be run.

Source: DeepMind

Illustrative stock photograph of a worker inside the DIII-D vacuum vessel during a maintenance period in 2017. It was taken years before the experiment described here.

Artificial intelligence Research

Artificial intelligence against tearing instability on DIII-D

On 21 February 2024 the journal Nature published an experiment on the DIII-D tokamak in California in which an artificial intelligence controller learned by trial and reward to adjust heating and plasma shape in real time, keeping the predicted risk of tearing, a leading cause of plasma collapse, under a chosen limit.