AI Is Now Designing Molecules That Absorb Carbon Directly From the Air — What the Latest Chemistry Breakthrough Means

For decades, designing a new molecule has been one of the slowest, most painstaking jobs in science. A chemist starts with a target — a drug, a material, a compound with some desired property — and has to work backward, step by step, figuring out which simpler building blocks could be combined, in which order, under which conditions, to eventually arrive at the thing they actually want. It is a process that can take years of trial, error, and hard-won intuition. Now artificial intelligence is starting to compress that process dramatically, and one of the most promising places this could matter is in the fight to pull carbon dioxide directly out of the air.

In April 2026, researchers at EPFL in Switzerland published a new tool called Synthegy that lets chemists describe, in plain everyday language, the kind of molecule or reaction strategy they want — and have an AI system interpret that description, evaluate possible synthesis routes, and guide the chemist toward a workable plan. It is part of a much larger wave of AI-driven materials discovery tools, including systems purpose-built for carbon capture chemistry, that together are reshaping how scientists search for the molecules that could help solve one of the hardest problems in climate science: getting carbon dioxide out of the atmosphere cheaply enough to matter at a global scale.

What Direct Air Capture Is, and Why It's So Expensive

Direct air capture, or DAC, is exactly what it sounds like: technology that pulls carbon dioxide straight out of the ambient air, rather than capturing it at the smokestack of a power plant or factory. This matters because some of the carbon dioxide already in the atmosphere needs to be removed if the world is to meet its climate targets, regardless of how quickly new emissions are cut. At the heart of every DAC system is a sorbent — a material that selectively binds CO₂ from air at concentrations of around 420 parts per million, then releases it in concentrated form so it can be stored or reused.

The trouble is that capturing carbon at such a low ambient concentration is fundamentally difficult. Air contains only about 0.04% CO₂, which means DAC systems have to process enormous volumes of air to extract a comparatively tiny amount of carbon, and that takes a lot of energy and expensive materials. Most DAC systems today operate at costs closer to $500 to $1,000 per ton of CO₂ removed, and Climeworks, one of the largest companies in the space, has reported costs as high as $600 to $1,200 per ton at its plants in Switzerland and Iceland. For comparison, almost no other industrial commodity sells for anywhere near that price per ton — even cement, a relatively expensive bulk material, trades for roughly $100 per ton. The US Department of Energy has set a target of bringing DAC costs down to $100 per ton through its Carbon Negative Shot programme, but even optimistic industry roadmaps suggest $200 to $350 per ton is a more realistic target for the early 2030s.

The bottleneck driving these high costs is largely a materials problem. Two main types of sorbent dominate the field: amine-based materials, which are relatively cheap but degrade over repeated use, and metal-organic frameworks, or MOFs — porous crystalline structures with enormous internal surface area, sometimes exceeding 3,000 square metres per gram, that can be tuned to grab CO₂ molecules with high selectivity. MOFs perform impressively but are expensive to synthesise, typically running $50 to $200 per kilogram compared to $5 to $20 per kilogram for simpler polymer-based sorbents. Finding a MOF or other sorbent that combines high performance with low cost, and that holds up to repeated capture-and-release cycles without degrading, has historically meant sorting through an almost unimaginable number of possible chemical structures — billions of plausible candidates — a search space far too large for human chemists to explore by trial and error alone.

How AI Is Changing the Search for Better Sorbents

This is precisely the kind of problem artificial intelligence is well suited to help with. Rather than synthesising and testing candidate materials one at a time in a lab, researchers can now use generative AI models to predict which molecular structures are likely to have the properties they want — high CO₂ capacity, low cost, durability under humid conditions — before ever making them physically. Tools like GHP-MOFassemble, a generative framework developed by US national laboratory researchers, use diffusion models, the same broad family of AI technique behind image-generating systems, to design novel MOF linker molecules optimised for CO₂ adsorption and built from components that can realistically be synthesised. Startups such as CuspAI are pursuing a similar strategy, running multiple AI models together — some generating candidate molecules with specific target properties, others assessing how well those candidates would actually perform — to narrow billions of theoretical possibilities down to a shortlist worth testing in the lab.

Synthegy approaches a related but distinct problem from a different angle. Rather than generating new molecular structures directly, it acts as an intelligent reasoning layer that sits on top of existing computational chemistry tools, helping chemists plan the sequence of reactions needed to actually build a target molecule once they know what they want. A chemist can type an instruction in ordinary language — asking the system to favor an early ring-closing step, for instance, or to avoid a particular class of reactive group — and Synthegy interprets that instruction, evaluates the available synthesis routes, and returns options that reflect the chemist's stated strategy. In testing against a panel of expert human chemists, the system's rankings of synthesis routes matched expert judgment most of the time, at a level of agreement comparable to how often two human experts agree with each other. Tools like this do not replace the harder scientific work of discovering which new sorbent chemistries are worth pursuing in the first place, but they meaningfully shrink the time and expertise required to turn a promising idea into an actual, synthesisable molecule — which matters a great deal in a field where every new candidate material has historically taken months of specialist planning before it ever reaches a test rig.

Together, these two strands of AI-assisted chemistry — generative models proposing new candidate structures, and reasoning models helping plan how to actually build them — represent a meaningful shift in how carbon capture materials research is done. The promise is not that AI will single-handedly invent a miracle sorbent, but that it can compress years of iterative human search into a fraction of the time, surfacing candidate materials that a human team might never have thought to try, and helping researchers reach the lab-testing stage faster and with a higher hit rate.

Where Carbon Capture Technology Actually Stands Today

It is worth being clear-eyed about how early-stage this technology still is, even with AI accelerating parts of the discovery process. Global direct air capture capacity today removes only a tiny fraction of the carbon dioxide humanity emits each year — current operating and under-construction DAC facilities worldwide capture in the tens of thousands of tons annually, while global CO₂ emissions run into the tens of billions of tons. The gap between where the technology is and where it would need to be to meaningfully affect the climate trajectory is enormous.

Climeworks, the field's most visible commercial player, is in the process of scaling up with what it calls Generation 3 technology, aiming for costs of $250 to $350 per ton captured by 2030 — roughly half today's cost — using new filter materials that consume half the energy and last three times longer than earlier sorbents. The company's first Generation 3 plant is planned for Louisiana as part of the Department of Energy-backed Project Cypress DAC Hub, with construction set to begin in 2026 and intended as a step toward eventual megaton-scale capacity. Corporate demand for carbon removal credits is also growing, with airlines including Lufthansa Group signing offtake agreements for future carbon removal in 2026, providing the kind of long-term revenue commitment that helps developers secure financing for new plants.

At the same time, the sector remains exposed to real financial and political risk. Venture investment in carbon capture and removal peaked at over $2.2 billion in 2022 before contracting to less than $1 billion by 2025, reflecting a market maturing from speculative early funding toward harder questions about bankability. Late in 2025, uncertainty around the US Department of Energy's DAC Hub funding briefly threatened billions of dollars committed to projects involving Climeworks and other major developers, before that funding was preserved — a reminder that policy support, not just chemistry, will determine how quickly this technology can scale.

What Researchers Say About the Realistic Timeline

Ask researchers in this field for a timeline, and the honest answer is measured in years and decades, not months. Even the most optimistic industry roadmaps treat the early-to-mid 2030s as the point at which DAC costs might fall into the $150 to $300 per ton range at meaningful scale, and reaching the long-discussed $100 per ton target — first floated by Carbon Engineering in 2018 — is now widely regarded within the field as unlikely without substantial, sustained policy support, and possibly unrealistic given inflation and the underlying energy costs involved.

AI-assisted materials discovery does not change that timeline by itself, and researchers working in the field are careful not to overstate what it can deliver. Designing a promising candidate molecule on a computer is one step in a much longer chain that includes synthesising it in a lab, testing its real-world durability across thousands of capture-and-release cycles, scaling production from grams to tons, and then building and financing full-sized industrial plants. Max Welling, a co-founder of the AI-for-materials startup CuspAI, has been candid that even once you have a promising candidate, the actual chemistry of producing it reliably is finicky — small changes in humidity or air quality can derail a lab's attempt to make the same material that worked perfectly in simulation. What AI plausibly does is compress the front end of that long pipeline: cutting down the search space and the planning time so that more candidate materials reach the lab-testing stage sooner, with a better-informed starting hypothesis about why they might work.

Can AI-Designed Carbon Capture Genuinely Help Meet Climate Targets?

This is the question that matters most, and the honest answer is: it can help, but it is not, by itself, a solution to the climate problem, and no serious researcher in the field claims otherwise. The Intergovernmental Panel on Climate Change's own modelling suggests that meeting global net-zero targets will require some amount of carbon dioxide removal alongside, not instead of, aggressive cuts to new emissions — current international assessments put the eventual need at hundreds of millions to billions of tons of CO₂ removal per year by mid-century. Today's global direct air capture capacity is a rounding error against that target. Closing a gap of that size will require dramatic cost reductions, dramatic scale-up of manufacturing and deployment, and sustained political and financial commitment over decades — not simply better algorithms.

Where AI-driven molecule design genuinely earns its place in this story is in attacking the cost problem directly, at the materials level, which has been one of the most stubborn barriers to DAC ever becoming affordable enough to deploy at the scale the climate math demands. If generative AI models can reliably identify cheaper, more durable, higher-capacity sorbent materials — and tools like Synthegy can meaningfully shorten the path from a promising idea to an actual synthesisable molecule — the compounding effect over the coming decade could be significant. Faster discovery cycles mean more candidate materials tested per year, which means a higher chance of finding the breakthroughs that bring costs down from hundreds of dollars per ton toward the double digits where DAC starts to become economically viable at a planetary scale.

That is a genuinely useful contribution. It is also, by the field's own admission, a piece of a much larger puzzle that includes renewable energy availability, geological storage capacity, manufacturing scale-up, public investment, and political will that has shown itself to be unstable even in the recent past. AI is not going to single-handedly capture humanity's way out of the climate crisis. But by making the search for better carbon-capturing molecules faster and smarter, it is helping to remove one of the genuine technical bottlenecks standing between direct air capture and the kind of large-scale deployment the world's climate targets actually require.

*This article is for informational and educational purposes only. Data and research findings are sourced from EPFL, ScienceDaily, the World Economic Forum, the Belfer Center for Science and International Affairs, PatSnap, and Climeworks.*