Carbon Capture Has an Optimization Problem

Climeworks opened its Mammoth plant in Iceland in May 2024—at the time, the largest direct air capture facility on Earth. Mammoth can remove approximately 36,000 tons of CO2 from the atmosphere per year. It does this by pushing air through filter beds containing amine-functionalized sorbents, heating the beds to release concentrated CO2, and injecting the resulting stream into basalt rock formations where it mineralizes permanently.

The cost per ton of CO2 removed: roughly $1,000 at current scale. The IPCC’s median estimate for the carbon price needed to limit warming to 1.5°C is approximately $200 per ton by 2030. The gap between where direct air capture costs and where it needs to be is the central problem in carbon removal, and it is a hard engineering problem at its core.

AI enters this picture as an optimization layer. Whether it’s a decisive one or a marginal one depends on where in the cost stack you’re looking.

The direct air capture cost structure breaks down roughly as follows: about 60% is energy, 20% is capital (the fans, heat exchangers, sorbent regeneration systems), 15% is operations and maintenance, and 5% is the CO2 compression, transportation, and storage. These percentages vary significantly by design and location—Iceland is cheap because geothermal provides near-zero-carbon heat, which is why Climeworks built there—but the energy dominance is a consistent feature.

Optimizing a process that is 60% energy costs means optimizing energy use. That is a domain where AI has demonstrated real impact. The same category of reinforcement learning that DeepMind used to reduce Google’s data center cooling energy by 40% in 2016 applies conceptually to direct air capture plant operations. The specific control problem—managing sorbent bed temperatures, fan speeds, and regeneration timing to maximize CO2 capture per unit of energy—has the right structure for RL: continuous state space, clear reward signal, and a system complex enough that heuristic rules leave significant room for improvement.

Heirloom Carbon, a startup focusing on a different approach (enhanced weathering of calcium carbonate rather than amine sorbents), has been using ML-optimized process control in its pilot operations in Tracy, California since 2023. Their reported result: a 12-18% reduction in energy consumption per ton compared to baseline operations running standard control loops. That’s not nothing. At scale, a 15% energy reduction translates directly to a 9% reduction in total cost of capture.

The more ambitious AI application in carbon capture is materials discovery. The bottleneck in sorbent-based direct air capture is the sorbent itself: the material needs to bind CO2 from air at concentrations of roughly 420 parts per million, release it cleanly when heated, survive hundreds of thousands of regeneration cycles without degrading, and be cheap enough to manufacture at scale. No material currently checks all of these boxes simultaneously.

Computational materials discovery—using machine learning to screen candidate materials before synthesis—has compressed the experimental cycle for sorbent development meaningfully. Microsoft’s Azure Quantum Elements platform, working in collaboration with the Pacific Northwest National Laboratory, screened approximately 32 million potential solid-state materials for battery electrolyte applications in 2023; the same approach is being applied to sorbent candidates for DAC.

The catch is that computational screening for sorbent performance involves properties—flexibility under cyclic thermal stress, resistance to water vapor interference, real-world binding kinetics—that are difficult to predict accurately from first-principles calculations alone. The models that screen candidates fast are less accurate. The accurate models are slower. The workflow involves multiple rounds of computational screening followed by experimental synthesis of the most promising candidates, and the experimental step remains rate-limiting.

There’s a second category of carbon capture where AI’s role is more straightforwardly impactful: natural carbon sinks. Forests, wetlands, and agricultural soils capture and store carbon through biological processes. Managing these systems—deciding where to plant trees, how to prevent fires, how to maximize organic carbon accumulation in soil—is an optimization problem with enormous spatial scale and heterogeneous conditions.

The Pachama platform, and similar tools from Land Life Company and Terrasos, use satellite imagery and machine learning to monitor the carbon content of forest restoration projects. The AI tasks here are site selection (which degraded lands have the most carbon accumulation potential per dollar of restoration cost), growth monitoring (is the planted forest actually surviving?), and leakage detection (is deforestation happening in adjacent areas, canceling out the restoration gains?).

This is a significantly easier AI application than designing novel sorbents, because the data is richer (satellite time series going back decades), the physics is better understood, and the optimization objectives are cleaner. Whether it actually results in more atmospheric carbon removal depends on the integrity of the projects and the governance of the carbon markets they feed—questions that AI cannot answer.

The geological storage side of the equation is also seeing AI applications. Sequestering CO2 in deep saline aquifers or depleted oil and gas reservoirs requires understanding subsurface geology well enough to predict where injected CO2 will migrate and whether it will stay put. The subsurface interpretation problem—reading seismic reflection data to characterize rock formations—has been the core data science challenge in the oil industry for decades. The tools developed for reservoir characterization in oil exploration are now being adapted for CO2 storage site assessment.

One number worth keeping in mind: humanity emits approximately 37 billion tons of CO2 per year. The entire global direct air capture industry currently removes around 0.01 million tons annually. The gap is roughly seven orders of magnitude.

No plausible trajectory of AI-assisted cost reduction closes that gap by 2050. The IEA’s net-zero scenario requires approximately 980 million tons of carbon removal per year by 2050, mostly from bioenergy with carbon capture and storage (BECCS) and direct air capture. Getting from 0.01 to 980 million tons in 24 years is primarily a question of industrial deployment, capital investment, and policy—not a question of whether the AI can shave a few percent off the sorbent regeneration energy.

This is worth saying plainly because the technology coverage of carbon capture often elides the difference between “AI is meaningfully improving the technology” and “the technology is close to solving the problem.” Both propositions might be true; the first does not imply the second.

What AI is genuinely doing in carbon capture: it is compressing the materials discovery cycle, improving operational efficiency in existing plants, enabling better site selection and monitoring for nature-based approaches, and advancing subsurface characterization for geological storage. These are legitimate technical contributions.

What AI cannot do: make carbon capture cheap fast enough to be a primary climate strategy without a parallel collapse in the cost of zero-carbon energy. The $1,000 per ton cost at Mammoth is mostly an energy cost at current electricity prices. At $20 per megawatt-hour electricity—which is approaching feasibility with aggressive solar and wind deployment—the economics of direct air capture look fundamentally different. That electricity price reduction is not an AI story; it’s a solar manufacturing learning-curve story that has been playing out for thirty years.

The honest synthesis is that AI’s role in carbon capture is probably most valuable as a multiplier on investments that would be worth making anyway. Better sorbent materials, identified faster through ML screening, make planned plants cheaper. Better process control makes operating plants more efficient. Better site selection makes nature-based projects more credible. None of these contributions is decisive; collectively, they move the cost curve in the right direction faster than without them.

Given that the cost target for DAC is essentially “get from $1,000 to under $200 in less than 15 years,” faster movement matters. The analogy to solar is instructive: photovoltaics dropped by roughly 90% in cost from 2010 to 2020 through a combination of Chinese manufacturing scale, design optimization, supply chain efficiency, and—increasingly—computational design of cell architectures and materials. AI was part of that story, not the whole story.

Direct air capture needs its own version of that trajectory. AI is one instrument in the orchestra. The music isn’t going to play itself.