Field notes The Energy Transition for the Rest of Us

Steel For Fuel N° 019 of 56 19 Aug 2024

Direct Air Capture and (machine) learning curves

by Andy Lubershane, Partner and Head of Research, Energy Impact Partners

In this note
  1. 01The question
  2. 02The answer
  3. 03The argument
  4. 04What you need to know first
  5. 05Details worth keeping
  6. 06Claims worth citing
  7. 07Where it’s contested

The question

What does Climeworks’ first real-world direct air capture plant missing its design targets tell us about the technology?

The answer

Less about direct air capture than about building anything industrial for the first time. The fix is mostly iteration rather than insight, and machine learning appears to be speeding part of it: finding better materials.

03The argument

Climeworks published results from its first plant, Orca, and they fell short of design targets, by amounts given only in a chart. Lubershane reads most of the shortfalls as classic first-of-a-kind problems for a new industrial facility rather than anything peculiar to pulling CO2 from the air. The story, he says, is not that direct air capture is hard, which everybody knew, but that the real world is hard, and Iceland’s winter is hard. Hence learning curves: progress comes from running something, seeing what failed and trying again, and he says Climeworks appears to have learned a lot in just two years of operating Orca, with the further performance data again in a chart. Then “learning” turns on him. He is a declared skeptic that generative AI is the next industrial revolution, but excited about one use, screening candidate materials for specific properties, which in this field already appears to be yielding better sorbents for capturing CO2.

04What you need to know first

Sorbent
The material that does the grabbing: CO2 sticks to it and is later driven off so it can be reused, so a better one moves the whole plant’s cost.

05Details worth keeping

  • The charts came from Eve Hanson, an Energy Impact Partners colleague leading its carbon capture and clean fuels research.
  • Running Tide, another carbon removal company, local to him in Portland, Maine, shut down around the same time.

06Claims worth citing

All figures as stated on 2024-08-19.

  • Orca’s real-world performance fell short of its design targets. Climeworks, May 2024, cited by Lubershane
  • Two metal organic framework structures were identified with AI and confirmed to capture CO2 more efficiently than established sorbents. Garcia et al, Nature, July 2024, cited by Lubershane
  • Svante is “apparently” moving to commercialize one of them. Lubershane

07Where it’s contested

Nobody argues back. He is a declared skeptic of the largest claims for generative AI and carves out materials discovery as his exception, a preference stated rather than defended. The load-bearing claim is the generalizing one, that Orca’s problems were mostly generic rather than specific to this technology, asserted from a chart rather than worked through.

Cite as: “Direct Air Capture and (machine) learning curves,” The Energy Transition for the Rest of Us, note on Steel For Fuel, August 19, 2024. CC BY 4.0. View the Markdown