AI-Powered Carbon Capture Technology: What Machine Learning Can Do

Direct air capture (DAC) sounds simple: a machine pulls in air, pulls out the carbon dioxide, and stores it. The hard part is that CO₂ makes up only about 426 parts per million of the atmosphere, roughly one molecule in every 2,350. Finding materials and operating schedules that do this cheaply is a search problem, and that is where AI-powered carbon capture technology enters. This article separates what machine learning has demonstrated in published work from what companies and headlines imply.

Key takeaways

• The best-documented use of AI in carbon capture is materials screening: ML models trained on quantum-chemistry calculations rank thousands of candidate sorbents before anyone synthesizes them.
• The open ODAC23 dataset holds more than 38 million density functional theory (DFT) calculations on more than 8,400 metal-organic frameworks (MOFs). Its 2025 successor covers about 15,000 MOFs.
• The IEA counted 18 DAC facilities in 2022 capturing almost 0.01 million tonnes of CO₂ per year. Climeworks' Mammoth plant is designed for 36,000 tonnes per year and is reported at about $1,000 per tonne.
• Public, peer-reviewed evidence that AI has already cut costs at an operating plant is thin. Treat company statements about AI-driven savings as claims, not measurements.
• Energy is the binding constraint: at 1 to 2.5 MWh per tonne, removing a billion tonnes a year would need 1,000 to 2,500 TWh of electricity.

Why pulling CO₂ out of air is so hard

The NOAA Global Monitoring Laboratory reports a monthly mean of 426.07 parts per million (ppm) at Mauna Loa for September 2026. In other words, a machine has to find about one CO₂ molecule among 2,350 others, then hold onto it, then let go of it on demand. Picture trying to collect a few grains of one color from a stadium filled with sand: the sorting is cheap per grain, but you need to process a mountain of material to get a bucketful.

Thermodynamics sets a floor on the work involved. According to the Wikipedia summary of DAC research, the theoretical minimum energy to separate CO₂ from air is about 250 kilowatt-hours per tonne, compared with about 100 for a natural gas power plant's exhaust and about 65 for a coal plant's. The more dilute the stream, the more energy it takes, which is why capture at a smokestack is cheaper than capture from open air.

The scale of what exists today is small. In its 2022 report the International Energy Agency (IEA) counted 18 DAC facilities in Canada, Europe, and the United States, capturing almost 0.01 million tonnes (10,000 tonnes) of CO₂ per year, and all but two sold the gas for use rather than burying it. For comparison, the IEA's Net Zero Emissions by 2050 scenario calls for more than 85 million tonnes captured by 2030 and around 980 million tonnes by 2050. For wider context on how capture fits into climate policy, see direct air capture: how it works, what it costs, and where it fits.

How is DAC different from capturing CO₂ at a smokestack?

Point-source capture treats a gas stream that is already rich in CO₂, such as a cement kiln's exhaust. DAC treats ambient air. The Wikipedia summary of carbon capture and storage notes that, as of July 2024, 44 commercial-scale carbon capture and storage plants were operating worldwide, collectively capturing about one-thousandth of global greenhouse gas emissions. That figure is a reminder that even the easier version of the technology is still a rounding error against annual emissions. The chemistry of the easier case is covered in how carbon capture works in practice.

Where the energy and money go in a DAC plant

A DAC plant has three jobs: move a lot of air, capture the CO₂ in it, and release the CO₂ as a concentrated stream that can be compressed and stored. Two families of technology dominate, and they differ in what captures the gas.

Approach	What captures CO₂	How CO₂ is released	Example
Liquid solvent	An alkaline solution (hydroxide) or amine-based liquid	A second chemical cycle regenerates the liquid and releases pure CO₂	Carbon Engineering's air contactor
Solid sorbent	A solid material that binds CO₂ by chemisorption	Heat and vacuum	Climeworks plants (a life-cycle assessment in Nature Energy, 2021, studied this temperature-vacuum swing design)
MOF-based sorbent (emerging)	A porous crystal with tunable pores	Heat, often with moisture-tolerance as a design challenge	Svante's CALF-20 and Captivate's MUF-16 are commercializing MOFs for CO₂ capture

Sources: Wikipedia's DAC, Carbon Engineering, Climeworks, and metal-organic framework entries.

Most of the energy goes into heating the capturing material to release the CO₂, plus fans to push air through and compressors to prepare the gas for storage. Wikipedia's DAC page puts total electricity use at roughly 1 to 2.5 megawatt-hours per tonne of CO₂, with additional heat in some designs. Those numbers vary by design and by what counts as electricity versus heat, so read them as order-of-magnitude figures rather than a spec sheet.

Why sorbent choice dominates the bill

A sorbent has to do three things that pull in different directions. It must bind CO₂ strongly enough to catch it from a dilute stream, release it without a huge heat input, and survive thousands of cycles in humid, dusty air. Think of an ideal sorbent as a bouncer who recognizes only one kind of guest, lets them in from a thin crowd, and then opens the door quickly when it is time to leave. Bind too tightly and the energy cost of letting go rises. Bind too loosely and nothing stays.

This is a multi-objective trade-off, and the stakes are concrete. Humidity makes it worse, because water competes with CO₂ for binding sites. The ODAC23 paper's abstract says that screening is hard in part because performance depends on humidity and temperature, and Wikipedia's DAC entry notes that moisture stability is a key challenge for MOFs.

How machine learning screens sorbent materials

Metal-organic frameworks are crystals built from metal nodes joined by organic linkers, which make them porous like a molecular sponge. Chemists can swap metals and linkers in nearly endless combinations, so the number of conceivable MOFs is far too large to synthesize one by one. The 2025 Nobel Prize in Chemistry went to Susumu Kitagawa, Richard Robson, and Omar Yaghi for their work on MOFs, per the Wikipedia entry.

The standard way to evaluate a candidate on a computer is density functional theory (DFT), a quantum-mechanical method that estimates how strongly a molecule such as CO₂ sticks to a surface. DFT is accurate but slow, because each material may need many calculations for different binding positions. That is the bottleneck machine learning targets.

The screening funnel

The workflow is like hiring. Machine learning is the résumé screen: fast, cheap, and imperfect. DFT is the interview: slower and more reliable. Lab synthesis and testing is the job trial: slowest, most expensive, and the only one that counts.

1. Generate a large library of candidate frameworks, either from known structures or by computer-built variations.
2. Run DFT on a subset to create training data.
3. Train a model to predict the DFT answer (binding energy, adsorption strength) directly from atomic structure.
4. Use the model to rank millions of candidates quickly, keep the top few hundred, and verify them with real DFT.
5. Synthesize and test the best few in a lab.

The ML step only makes sense if the model's errors are small compared with the differences between good and bad candidates. That is why validation against held-out DFT results matters, a theme explored in overfitting and underfitting.

What the open datasets contain

The Open DAC 2023 dataset (ODAC23), described in an arXiv paper by Anuroop Sriram and colleagues, contains more than 38 million DFT calculations on more than 8,400 MOFs, including CO₂ and H₂O adsorbed in the pores. The authors describe it as the largest MOF adsorption dataset at DFT accuracy available, and they trained ML models to approximate DFT as a baseline. By my arithmetic, 38 million calculations over 8,400 materials works out to roughly 4,500 calculations per material, which shows why brute-force DFT is expensive and why a surrogate model is attractive.

The 2025 follow-up, ODAC25, expands to nearly 60 million DFT single-point calculations on about 15,000 MOFs and four adsorbates: CO₂, H₂O, N₂, and O₂. The authors add functionalized MOFs and synthetically generated frameworks, improve DFT accuracy, and release machine-learned interatomic potentials trained on the data. The same paper notes that ODAC23 was published in ACS Central Science in 2024. In plain terms, the field has moved from "can ML predict one number?" to "can ML approximate the physics of a flexible, humid, multi-gas pore?"

What these datasets do not show

Neither paper claims that an AI-discovered MOF is now running in a commercial plant. What they supply is infrastructure: shared data, benchmarks, and models that other labs can use to propose candidates. Predicted performance still has to survive synthesis, scale-up, thousands of heating and cooling cycles, and real weather. The broader field of using algorithms to find new compounds is surveyed in how machine learning is accelerating materials discovery, and the model families used on molecules are explained in graph neural networks in chemistry.

Where AI could help run a plant

The second use case is operational. A DAC plant is a machine with many adjustable settings, including how long to expose the sorbent to air, how hot to heat it, how fast to run the fans, and when to run the energy-hungry steps. In principle, an ML model trained on sensor data could suggest better settings, predict when a component will fail, or schedule the plant around weather and electricity prices.

Here the honest answer is that the public evidence is sparse. The literature includes studies such as "The impact of climate on solvent-based direct air capture systems" (An, Farooqui, and McCoy, 2022) and a 2024 paper on warehouse automation for scalable DAC (McQueen and Drennan), which show that location and automation shape cost. No peer-reviewed measurement among the sources reviewed shows how much an AI controller has reduced the cost of a specific operating plant. Several percentages circulate in press materials, but without a named dataset or paper they cannot be checked.

A scheduling example, with illustrative numbers

The logic of scheduling is easy to show with made-up numbers. Suppose a plant needs 2 megawatt-hours per tonne, and it can delay half of its regeneration energy by a few hours without losing capture. If power costs $20 per megawatt-hour at midday and $80 in the evening, shifting that half from evening to midday saves $60 on one megawatt-hour per tonne, or $60 per tonne. These numbers are illustrative only. The point is that a forecasting model, which predicts prices or solar output, can turn flexibility into money, but only if the hardware can actually pause and restart without damage.

The same logic is behind AI in power systems generally, covered in smart cities and AI. The caveat is that every model also needs electricity to train and run. See why AI needs so much computing power and how AI data centers use electricity for the scale of that overhead.

What operating plants show today

A short tour of the best-known plants shows how far the technology is from gigatonne scale. Capacities below are nameplate (design) figures from the cited summaries, not measured output.

Plant	Operator	Location	Design capacity	Notes
Hinwil	Climeworks	Switzerland	900 t/yr	Opened May 2017, uses waste heat from a local incinerator, stopped October 2022
Orca	Climeworks	Hellisheiði, Iceland	Up to 4,000 t/yr	Opened September 2021, geothermal power, CO₂ mineralized by Carbfix
Mammoth	Climeworks	Iceland	Up to 36,000 t/yr	Activated May 8, 2024; reported at about $1,000 per tonne
Squamish pilot	Carbon Engineering	British Columbia	About 1 t/day	Operating since 2015; liquid hydroxide process
Stratos	Occidental (1PointFive)	Permian Basin, Texas	Not stated on Occidental's page	Described by Occidental as under construction; expected to be the largest facility of its kind

Sources: Wikipedia entries on Direct air capture, Climeworks, and Carbon Engineering; Occidental's DAC page; IEA (2022).

Two caveats matter. First, nameplate is not output. A May 2025 analysis by the Icelandic outlet Heimildin, summarized on Wikipedia's Climeworks page, concluded that as of 2023 the machines captured only a fraction of their claimed volumes. Wikipedia's DAC page likewise says Orca has averaged around 600 tonnes per year against a 4,000-tonne nameplate. Second, Occidental bought Carbon Engineering in 2023 for $1.1 billion, according to Wikipedia, and its own page says Stratos will be the largest facility of its kind once complete. Occidental's page described Stratos as under construction when this article was checked in October 2026.

What is in the pipeline for AI

Where AI shows up in these plants is mostly not documented in public. The sources reviewed for this article contain no audited breakdown of AI contributions from Climeworks or Carbon Engineering. If a company states that an algorithm saved a given percentage, ask three questions: relative to what baseline, measured over what period, and verified by whom.

The energy problem at gigatonne scale

A gigatonne (one billion tonnes) of CO₂ per year is the scale many climate scenarios imagine for removal. Multiply by the 1 to 2.5 MWh per tonne figure and the electricity requirement becomes 1,000 to 2,500 terawatt-hours (TWh) per year, by my arithmetic. Ember's Global Electricity Review 2025 implies total world generation of roughly 30,800 TWh in 2024, so one gigatonne of DAC would take about 3% to 8% of today's global electricity supply.

That is a large load, but it is not impossible, and it shows why location and power source matter. DAC powered by coal or gas would emit CO₂ while removing it, which is why Orca's geothermal site and the IEA's emphasis on regions with high renewable potential recur in the literature. The IEA's 2022 report lists energy prices, siting near abundant renewable energy, and sustained public and private support among the main drivers of future costs.

Scenario	Annual removal	Electricity at 1 to 2.5 MWh/t	Share of 2024 world generation (about 30,800 TWh)
IEA Net Zero, 2030	More than 85 Mt	More than 85 to 213 TWh	Under 1%
One gigatonne	1,000 Mt	1,000 to 2,500 TWh	About 3% to 8%
IEA Net Zero, 2050	About 980 Mt	About 980 to 2,450 TWh	About 3% to 8%

The rows combine IEA scenario volumes with the electricity range above, so they are projections built on one source's estimate, not forecasts. The grid constraint is covered in why grid storage is hard and the supply side in renewable energy explained.

What DAC costs, and what pays for it

Cost estimates for DAC span a huge range depending on who is estimating and whether the plant exists yet. The IEA's 2022 report estimated that capture at a large plant built at that time would cost between $125 and $335 per tonne, and called the estimates "wide-ranging and uncertain." It said costs could fall below $100 per tonne by 2030 in places with high renewable potential that use the best available heat and power. A widely cited 2018 engineering study by David Keith and colleagues estimated about $94 to $233 per tonne for Carbon Engineering's design (per the Wikipedia entry).

Real plants have not matched those numbers. Wikipedia's DAC entry reports more than $1,000 per tonne in 2025 per BloombergNEF, and about $1,000 for Mammoth. The same entry cites a 2025 study estimating that plants below 0.05 million tonnes a year cost more than $1,000 per tonne while a 1-million-tonne plant could fall to roughly $97 to $168. The gap between a design study and an early plant is typical of first-of-a-kind technology, and it is why scale matters as much as chemistry.

Source	Cost estimate (per tonne CO₂)	Basis
Keith et al., 2018	About $94 to $233	Engineering design study, large plant
IEA, 2022	$125 to $335	Estimate for a large plant built then; possible fall below $100 by 2030 in favorable sites
Mammoth, reported	About $1,000	Operating plant, per Wikipedia summary
BloombergNEF, 2025	More than $1,000	Market estimate, per Wikipedia summary

The 45Q tax credit

In the United States, the federal 45Q credit pays for captured and stored carbon oxide. The IRS page lists $36 per metric ton for DAC as the base amount, multiplied by five for facilities that meet prevailing wage and apprenticeship requirements, which gives $180. A DAC facility must capture at least 1,000 metric tons a year to qualify. The credit amounts and eligibility rules have been revised by legislation before, so check the current IRS guidance rather than relying on any article, including this one.

A little arithmetic puts the numbers in perspective. At $1,000 per tonne, removing one million tonnes costs $1 billion. At $180 per tonne of credit, that same million tonnes earns $180 million. The credit narrows the gap, but it does not close it at today's reported prices, which is one reason cost reduction through better sorbents and operations is so valuable.

Common misconceptions about AI and carbon removal

"AI will make DAC cheap on its own." Cost is shaped by energy, capital equipment, sorbent lifetime, and scale. AI can help with the first and third, but it cannot change the thermodynamic floor.

"DAC replaces emission cuts." The IEA frames DAC as a tool for emissions that are hard to avoid and for legacy emissions. It is a supplement to emission reductions, not a substitute. For the underlying science of why concentrations keep rising, see climate change science: how it actually works.

"A neural network predicting a material is the same as discovering one." Prediction is the first filter. A material becomes useful only after it has been made, tested in humid air, and cycled for months.

"AI makes carbon markets trustworthy." Machine learning can help with measurement and monitoring, as discussed in real-time emissions tracking and AI and carbon markets, but accounting integrity depends on rules and verification, not on algorithms alone.

"Point-source capture is just a smaller version of DAC." It is a different engineering problem, because the gas is richer in CO₂ and the energy floor is lower. The chemistry behind both is explained in how carbon capture works.

The takeaway

The most defensible claim about AI-powered carbon capture technology is a modest one: machine learning is becoming a standard way to narrow a vast space of candidate sorbents before expensive experiments, and open datasets such as ODAC23 and ODAC25 have made that work reproducible. The claim that AI has already made DAC affordable is not supported by public measurements, which show early plants running well above the $100-per-tonne targets and, in some reports, well below their nameplate output.

What AI cannot change is the arithmetic. Air is dilute, separation takes energy, and a gigatonne of removal would draw a few percent of the world's electricity. If the next decade delivers a sorbent that binds CO₂ firmly in humid air and releases it with modest heat, a prediction model will probably have helped find it, and a lab will have had to prove it. That division of labor is the real story.

FAQ

Q: How does captured CO₂ get stored permanently? A: Some plants inject it underground. In Iceland, Climeworks works with Carbfix, which injects CO₂ dissolved in water into basalt about 700 meters down, where it mineralizes into rock, according to the Wikipedia summary. Other captured CO₂ is sold for use, and the IEA noted in 2022 that all but two operating DAC plants sold their output.

Q: Is AI essential to carbon capture, or just helpful? A: Helpful in specific steps. Machine learning lets researchers screen thousands of MOFs in far less time than DFT alone, as the ODAC23 and ODAC25 datasets show. But no source I reviewed shows that a plant's performance or cost was transformed by AI, and every candidate still has to be validated in a lab.

Q: How much DAC capacity exists today? A: The IEA counted 18 facilities with almost 0.01 million tonnes per year in 2022. Since then, Climeworks' Mammoth plant has opened with a design capacity of up to 36,000 tonnes per year, and Occidental describes its large Texas plant as under construction. Actual output can lag nameplate, as the 2025 Heimildin analysis of Climeworks suggested.

Q: Can point-source capture, such as at cement plants, be optimized with AI too? A: In principle, yes, since it is also a chemical process with many tunable settings. The Wikipedia summary reports 44 commercial-scale carbon capture and storage plants as of July 2024, capturing about one-thousandth of global emissions, so the potential for process improvement is large in both cases.

Q: What does the U.S. 45Q credit pay for DAC? A: The IRS lists a base of $36 per metric ton for DAC, multiplied by five when prevailing wage and apprenticeship requirements are met, giving $180. A facility needs to capture at least 1,000 metric tons a year. Check IRS guidance for the current version, because the rules have changed over time.

Q: Why is humidity such a problem for sorbents? A: Water molecules compete with CO₂ for binding sites and can degrade some materials. The ODAC23 authors say performance depends on humidity and temperature, and Wikipedia's DAC entry calls moisture stability a key challenge for MOFs. This is why ODAC datasets include H₂O adsorption alongside CO₂.

Q: Could trees do this job instead? A: Forests and soils also store carbon, and they do not require a new industrial plant, but they need land and can be reversed by fire or logging. Most analyses treat nature-based and engineered removal as complements. For a sense of where the forest area is, see countries with the most forest cover, and for a soil-based method see how biochar stores carbon in soil.

Sources

• IEA, Direct Air Capture 2022 and its executive summary
• Sriram et al., The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture (arXiv:2311.00341)
• Sriram et al., The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture (arXiv:2508.03162)
• IRS, Credit for carbon oxide sequestration (Section 45Q)
• NOAA Global Monitoring Laboratory, Trends in atmospheric carbon dioxide
• Ember, Global Electricity Review 2025
• Occidental, Direct air capture
• Wikipedia: Direct air capture, Climeworks, Carbon Engineering, Metal-organic framework, Carbon capture and storage

---

Read next:

• Machine Learning for Climate Prediction — How AI Forecasts Climate Change
• Neural Networks for Ocean Monitoring — How AI Protects Marine Ecosystems