Perspectives of "low-resolution" data from computation, experiment, and automation communities.
Since yesterday's post , the Ceder group posted a reply with new characterization data, and crystallographers asked new questions . I'm grateful for many conversations that have given me a deeper appreciation of all perspectives, as well as sharpened my sense of where we can do better as a community.
Self-driving labs (SDLs) for materials discovery is a multidisciplinary field comprising computation, automation, and experimentation. We’re all climbing up the hill toward autonomous materials discovery. That’s the big prize.
?
Having been trained the “traditional way,” we all have different initial conditions. A computation group may have a high-fidelity materials-prediction algorithm, but less automation and experimental experience; for them, a huge step forward might be bringing synthesis in house. Even "lower resolution" synthesis and characterization data is far better than the status quo of not having an in-house way to test predictions, with future implications for how scientific collaborations are organized, a point made by Shyam Dwaraknath tweet here .
?
Experimentalists are right to call out areas for improvement in this approach. I really resonate with the rigor demanded by Robert Palgrave , Leslie Schoop , and others. I personally believe low-resolution data is really useful for screening, but I’m having issues with it being the standard of proof for discovery. But, the early steps of this path up the mountain may well offend the sensibilities of experimental groups, to achieve the advances mentioned in the previous paragraph.
?
领英推荐
Now flip this. Several experimental groups are taking a different path up the mountain. I come from this base camp, educated as an experimental materials researcher. We often use low-fidelity computation (generative/inverse design and ML methods) to guide our experimental work. Our standards are quite high for experiment, but admittedly less strictly defined for computation and automation. We accept the reality that having "lower-resolution" computation and automation is far better than having none. When I write “we predicted a material,” did we truly predict a synthesis path and structure-property relationship? maybe? not always? but if it helps us find a new material, it’s much better than what we as experimental groups had access to before, yes?
?
Automation groups are trying to make low-cost versions of SDLs to accomplish the same thing. My team in the MIT mechanical engineering department is one of those. I notice that these approaches can offend the sensibilities of computational AND experimental groups, because our hacking projects don’t produce the most reproducible and high-fidelity materials. But if it gets us close — accomplishes screening — is it not useful? And as our tools improve, with sufficient turns of the crank, might this also be a viable path up the mountain? (And a far less expensive, and more equitable and accessible, one?)
?
I suspect we have a few more Nature papers to go before we reach the top of that proverbial materials-discovery mountain. Guarantee some of these will rub some part of our community in all the wrong ways. But each, hopefully, will represent a major step up the mountain for another part, and a learning opportunity for us all.
The A-Lab will improve with time. I have to admit it's a gorgeous tool — a point Sterling G. Baird made in this post . Gerd’s LinkedIn post specifically opened the door to auto-XRD collaborations; I for one plan to take him up on this invitation, because solving this matter is useful for everyone’s research. Rob Palgrave made some specific suggestions too, and also shared his excitement about the future of autonomous research.
I'm grateful to have gained a better perspective for why experimental, computation, and automation communities have different burdens of proof associated with the words "discovery" and "prediction," and why they find different kinds of "low-resolution" data much better than the status quo. I hope we continue to set a supportive tone to attract the best young talent to the field, and support each other up the mountain.
Daniel M. Tellep Distinguished Professor in Materials Science and Engineering at University of California, Berkeley
11 个月totally agreed Tonio: very thoughtful comments. Indeed, we should all work together towards improved accelerated materials design and not trash each other. I promise never to call you out on Twitter for any imperfect (but possibly still very useful!) theory and instead partner with you ??
Gulsine | Faster to market with A2P2 | Extends lifespan of products
11 个月It needs to be celebrated and the comments must be in constructive nature.
Boeing Roundhill Professor of Chemical Engineering
11 个月Very insightful and balanced. I also believe the XRD is what the XRD is… it may or may not get any better for this SDL platform. The challenge seems to be the way it is superficially fit/interpreted and this should be a fix that helps advance the whole field. I hope the discussion triggers more work on automating data interpretation with rigor equating human experiments, which is still incredibly hard, but it would totaly change the outcome. Perhaps the algorithm, recognizing impurities, continues improving the synthesis recipe to get a better yield or move on to a different material. The ‘race to be first’ mentality is what is getting in the way of this hard work.
Assistant Professor, Computational and Statistical Materials Science
11 个月Excellent points! I believe thinking about how low resolution data can inform decisions on when to go after higher resolution data, with a proper uncertainty-aware cost-benefit analysis is key in accelerating discovery within multimodal setting.
Materials Scientist at Pacific Northwest National Laboratory - PNNL
11 个月Totally agree ?? collaboration is very important, there is No one general lab automation fits all materials discovery! We open the door to collaborate with any groups by providing lab automation tools in the field of wet chemistry.