Skip to main content
v2026.11,986 entries · CC-BY 4.0

CarbonTracker: NOAA CO2 System, Python Package and Think Tank

CarbonTracker names three unrelated projects: NOAA’s CO2 flux system, a Python tool for ML training footprints, and a finance think tank. Find yours.

Written and maintained by CASRAI Editorial Board

Published

On this page

CarbonTracker is not one thing. In research it usually means one of three unrelated projects: NOAA’s CarbonTracker, an atmospheric CO2 measurement and modeling system that estimates where carbon dioxide enters and leaves the atmosphere; the open-source carbontracker Python package, which tracks and predicts the energy use and carbon footprint of training deep learning models; or Carbon Tracker, an independent financial think tank focused on the energy transition. Which one you need depends on whether you study the carbon cycle, compute emissions, or capital markets.

Which CarbonTracker do you mean?

The name is shared by projects with different owners, different outputs and different audiences, so a search for the bare word returns a mixture. The quickest way to land on the right one is to decide what you want to measure. If the quantity is carbon dioxide moving between land, ocean and atmosphere, you want the NOAA system. If it is the electricity and emissions of your own model-training job, you want the Python package. If it is the financial exposure of fossil-fuel assets, you want the think tank. The data table below sets the three side by side, using each project’s own description of itself.

None of the three is a successor, fork or version of another, and the shared name is a coincidence of naming rather than a relationship between the organizations. Citing the wrong one is an easy error in a methods section, because each is cited differently: the NOAA system through its methodology paper, the Python package through its software repository and workshop paper, and the think tank as an organization.

NOAA CarbonTracker: CO2 sources and sinks

What it is

NOAA’s Global Monitoring Laboratory describes CarbonTracker as a CO2 measurement and modeling system developed by NOAA. Its purpose is to estimate surface-atmosphere CO2 fluxes, meaning the net amount of carbon dioxide taken up or released by the surface in a given place and time. The current release, CT2026, provides global flux estimates from January 2000 through December 2025. The project’s methodology reference is Peters et al. (2007), published in Proceedings of the National Academy of Sciences, volume 104, pages 18925 to 18930.

How it works

NOAA’s tutorial explains the approach in plain terms. The system compares atmospheric CO2 mole fractions predicted by a model with the mole fractions actually observed, and it solves for the surface sources and sinks that best make the two agree. It starts from a first guess of the fluxes, uses atmospheric transport driven by weather forecasts to carry those fluxes through the atmosphere, and adjusts them until the predicted atmosphere matches observations. Separate modules represent air-sea exchange, photosynthesis and respiration, fire, and fossil fuel emissions.

The observational input in CT2026 is the GLOBALVIEWplus 11.0 ObsPack, a compilation of atmospheric CO2 measurements contributed by about 80 laboratories. Other inputs are an atmospheric transport model, emissions inventories and output from a terrestrial biosphere model that supplies first guesses for land exchange.

What it estimates, and what it does not

The estimated fluxes are the land biosphere (excluding fire), fire emissions and air-sea exchange. Fossil fuel emissions are deliberately not estimated: they are taken from emissions inventories and subtracted from the observed signal. NOAA notes that errors in those inventories carry through into the estimated natural fluxes, which matters more for the smaller natural fluxes. Anyone who cites a CarbonTracker number as a fossil emissions estimate has misread the product.

Resolution and uncertainty caveats stated by NOAA

  • The ecoregion is the scale at which the system actually operates. The 1°×1° flux maps are model-derived, and NOAA states they should not be interpreted as quantitatively meaningful for each individual block.
  • Adjustments are applied uniformly across each ecoregion, which assumes ecosystems behave coherently over large areas.
  • Where observations are sparse, estimated sinks may drift back toward the near-neutral first guess, which could bias results.
  • Flux uncertainties are probably somewhat underestimated because of the initialization scheme used in this release, and NOAA says this is expected to be corrected in future versions.
  • The biosphere process model is basic and NOAA says it will be improved. Additional measurement sites are expected to produce the largest improvement.

The practical reading for a researcher is that country-scale and continental totals are better supported than block-by-block maps, and that the uncertainty ranges should be treated as a lower bound.

The CT2026 page lists related products: CarbonTracker Near-Real Time, CarbonTracker Methane, CarbonTracker Europe, CarbonTracker-Asia, CarbonTracker Lagrange and the U.S. Emission Tracker for Potent Greenhouse Gases. When a paper says it used CarbonTracker, check which of these, and which release, it means, because they differ in gas, region and time period.

The carbontracker Python package: compute footprint

What it is

The carbontracker package is an open-source Python tool for tracking and predicting the energy consumption and carbon footprint of training deep learning models. It is described in the paper “Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models” by Wolff Anthony, Kanding and Selvan, submitted to arXiv on 6 July 2020 and presented at an ICML 2020 workshop. The authors argue for reporting a model’s carbon footprint alongside its performance metrics. The code is hosted on GitHub under the MIT license and installs with pip install carbontracker.

How it is used

There are two entry points. From the command line, carbontracker python script.py wraps an existing script. In code, you create a CarbonTracker object with the number of epochs, call epoch_start() and epoch_end() around each epoch, and call stop() to report actual consumption if training ends early.

It measures elapsed time, energy in kilowatt-hours and emissions in grams of CO2 equivalent over the epochs it monitors. After a configurable number of epochs (the default is one), it predicts the consumption of the full run from what it has observed. An optional setting expresses the result as an equivalent car distance. The README also describes a setting that pauses training after the prediction until the user confirms, and a log parser that totals logs across a directory.

Hardware and environment support

According to the README, it supports NVIDIA GPUs through NVML, Intel CPUs through Intel RAPL, Slurm, and Google Colab and Jupyter notebooks. The README does not list AMD GPUs or other non-NVIDIA accelerators as supported, so check this before assuming a reported figure covers all of your hardware.

The README also records limits that affect shared systems. GPU selection looks at the CUDA_VISIBLE_DEVICES setting first and, failing that, measures all available GPUs, which matters on shared clusters. The option that restricts measurement to devices running the main process cannot find GPUs inside containers that were not started with --pid=host.

Where it fits in research reporting

The tool produces a compute-side figure that can be reported next to accuracy or loss. For the concepts behind such a figure, see training carbon footprint and compute carbon footprint. If your scheduling can shift work to cleaner grid hours, carbon-aware computing is the related practice. With verbose output the package also reports the average carbon intensity, in grams of CO2 equivalent per kilowatt-hour, for the detected location.

Carbon Tracker: the financial think tank

Carbon Tracker describes itself as an independent financial think tank working on the energy transition and capital markets. Its stated role is to help markets understand and quantify the risks implied by that transition. It presents itself as a not-for-profit research house and part of the Tracker Group, alongside Planet Tracker. Its founder and director is Mark Campanale, and the organization is credited by its own account with the terms carbon bubble, unburnable carbon and stranded assets. It launched a Global Registry of Fossil Fuels in 2021.

This organization does not measure atmospheric CO2 and does not provide software for estimating a model’s emissions. Its relevance to research administration is as a source for financial-risk analysis of fossil-fuel assets, which is a different question from either of the two technical tools above.

Where it matters for research practice

Three situations make the distinction matter in practice.

  • Methods sections and data citations. A paper that uses NOAA flux estimates should name the release and cite the methodology reference; a paper that reports training emissions should name the package, its version and the hardware it could measure.
  • Grant and sustainability reporting. Projects that must report emissions from computing need a measurement approach for their own footprint, which is a Python-package-type question. Projects that need atmospheric context need NOAA-type flux data. For the general accounting concepts, see carbon accounting in research, and for a different open emissions dataset see the guide to Climate TRACE.
  • Search and literature review. When screening papers, include a qualifying term such as NOAA, flux, deep learning or think tank, or the results will mix all three.

Common misunderstandings

“CarbonTracker tells me a country’s fossil fuel emissions.”

The NOAA system does not estimate fossil fuel emissions. It takes them from inventories and subtracts them to isolate land and ocean fluxes.

“The 1°×1° maps are measurements for each square.”

NOAA states that the maps should not be interpreted as quantitatively meaningful for each block. The system operates at ecoregion scale.

“The Python package measures the whole footprint of my model.”

It monitors the training epochs it wraps and predicts the rest of the run from them, on supported hardware. It is not described as covering data collection, inference after deployment or the manufacture of the hardware.

“These are versions of one tool.”

They are unrelated projects from different owners that share a name.

Sources

  • NOAA Global Monitoring Laboratory, CarbonTracker CT2026: https://gml.noaa.gov/ccgg/carbontracker/ (checked 2026-11-15)
  • NOAA Global Monitoring Laboratory, CarbonTracker tutorial: https://gml.noaa.gov/ccgg/carbontracker/tutorial.php (checked 2026-11-15)
  • carbontracker repository and README, GitHub (lfwa/carbontracker): https://github.com/lfwa/carbontracker (checked 2026-11-15)
  • Wolff Anthony, Kanding and Selvan, Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models, arXiv:2007.03051: https://arxiv.org/abs/2007.03051 (checked 2026-11-15)
  • Carbon Tracker, About: https://www.carbontracker.org/about/ (checked 2026-11-15)

Data: Three projects called CarbonTracker compared

Compares the three unrelated projects that share the name CarbonTracker by owner, what each measures or analyses, main output and typical user. Compiled from each project's own official page, README or paper, read 2026-11-15.

Period covered: 2000/2026.

Three projects called CarbonTracker compared
ProjectRun byWhat it addressesMain outputTypical user
NOAA CarbonTracker (CT2026)NOAA Global Monitoring LaboratorySurface-atmosphere CO2 fluxes: land biosphere, fire and air-sea exchange; fossil fuel emissions taken from inventories, not estimatedGlobal flux estimates, January 2000 to December 2025Carbon-cycle and atmospheric scientists
carbontracker Python packageAuthors Wolff Anthony, Kanding and Selvan; open source on GitHub (MIT license)Energy use and carbon footprint of training deep learning modelsTime, energy in kWh and CO2 equivalent in grams, with a prediction for the full runMachine learning researchers
Carbon Tracker (carbontracker.org)Carbon Tracker, an independent financial think tank in the Tracker GroupFinancial risk from the energy transition and capital marketsResearch and analysis on implied risks to marketsInvestors, analysts and policy audiences

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · full access

Ask about CarbonTracker: NOAA CO2 System, Python Package and Think Tank

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →