U.S. Progress on Self-Driving Labs
In a world where inference and ideas are cheap...
I left the Department of Energy last month. You can find some closing thoughts from my time in federal service here:
More to come on whats next, but in the meantime, expect to see more content coming from
in the coming months!For those who have not been keeping up, the past year has seen tremendous progress in the development of self-driving labs and in the policy interest around supporting this new paradigm of doing science.
I have written extensively about the need for more public R&D investment in self-driving labs, starting on this blog in October 2023, and with multiple think tanks in January 2024 [here][here].
For a primer on self-driving labs and why they matter, see here:
So what has been going on in the past year in self-driving labs?
The U.S. Government Steps Up
A small group of us across the federal government have been working to coordinate and highlight the importance of autonomous experimentation i.e. self-driving labs.1
In June 2024, our interagency group hosted a workshop on “autonomous materials innovation infrastructure” (read: self-driving labs) and had over 50 scientists and researchers from universities, national labs, and companies attend to identify gaps in the self-driving lab ecosystem for different material classes. You can find the workshop readout report here.
Up to this point, the US government has not directly funded any self-driving labs investments. Most of the funding has been through national lab directed R&D funding or through NSF grants, but there has been no federal science funding program created specifically for self-driving labs in any domain. This is in stark contrast to Canada, which has pooled over $500M of public-private funding into self-driving labs at University of Toronto.
But thanks to the efforts of our interagency group, that has now changed:
In October 2024, the CHIPS R&D office announced $100M funding for autonomous experimentation and AI systems for semiconductor material discovery.
In November 2024, the Department of Energy’s ARPA-E announced $40M funding for self-driving labs in catalyst discovery.
Note that Trump funding freeze impacts both of these programs - neither of them will be able to move forward with awardees and disburse funding until the freeze is lifted.
AI Industry Investments in SDLs
And it is not just governments that are recognizing the importance of self-driving labs for scientific discovery.
Meta, with Carnegie Mellon University, launched the Open Catalyst challenge in 2021, releasing a dataset with millions of density functional theory (DFT) simulations. The hope was that this kind of dataset would spur new AI models that could discover new catalyst materials for hydrogen production.2
My critique of this project had always been that training AI models on DFT data is inadequate for materials discovery, as simulation data is not always accurate with real-world results, nor do simulations give you instructions on how to synthesize materials.
I was pleasantly surprised in November 2024, when Meta announced the Open Catalyst 2024 dataset, which included experimental data for hundreds of synthesized nanomaterials. This was possible through a partnership with University of Toronto and VSparticle, to develop automated synthesis and characterization self-driving labs for these catalyst materials.34
What comes next
AI models have continued to advance, surpassing every human benchmark and making rapid progress on even the hardest scientific and mathematical benchmarks.
The Google AI Coscientist paper from last week is an indicator that AI systems are increasingly able to meaningfully contribute to scientific ideation and hypothesis generation.
In a world where inference and ideas are cheap, testing and verification become rate limiting.5 Self-driving labs will only become more important this next year. If the Trump funding freeze is lifted, the $140M in self-driving lab funding would be a transformative investment from the U.S. government. This public funding would jumpstart the complementary infrastructure we need to realize the full potential AI models, and the companies that build them, to accelerate scientific progress.
Formally, the Material Genome Initiative (MGI) Autonomous Materials Innovation Infrastructure (AMII) subcommittee working group (or at least I think thats the name).
You can find more from Larry Zitnick’s CVPR 2023 talk on Open Catalyst.
Google and Lawrence Berkeley National Lab also worked together on a self-driving lab for inorganic powder synthesis in late 2023, although later their results were criticized as overstating the extent to which new materials were synthesized and discovered.
The excellent Materialism podcast interviewed Meta and VSParticle about Opencatalyst as well if you want to learn more
The 2024 Toner-Rodgers paper highlights this as well: scientist who use AI spend less time on ideation and are uniformly less satisfied with their jobs as a result.
