How do you begin to describe a "self-driving laboratory?"
With our blossoming knowledge of "AI" and how it has become more prevalent within technology, let alone how it has been influencing art - it feels like we've barely scratched the surface of how to apply it in a scientific setting. Material researchers are spearheading self-driving laboratories as a solution that combines chemistry, physics, robotics and, of course, AI.
One such material scientist, Sterling Baird, a Ph.D student with the University of Utah. Sterling and his team have developed a bare-bones, "hello world" kit of a DIY self-driving laboratory for academic settings. For the full interview, watch the video below. An abridged transcript is below!
We normally do a deep-dive of internet history and presence of our creators, and we found that your work has been cited in over 38 publications, and you've authored 32 research papers. Your Github has also been recently praised for how in-depth of commits you have. What other shoutouts can we include?!
Maybe just a couple extras: you know, it's been a long journey, my wife and my daughter are very supportive, and I've had some amazing mentors and advisors over the years. Dr. Robert Davis, my undergraduate partner, Oliver Johnson, at my master's program, and I'm currently working with Dr. David Sparks as he advises for my Ph.D.
Who makes up your team, and what was the mission of this collaboration?
There's a couple of different scopes. Within the inner-circle, there's me and Dr. Taylor Sparks. We've been using these machine learning models to predict new materials and set out in the lab, and validate those experiments. And what we're looking to get into more is using this demo and connections with the broader community are "self-driving laboratories": this idea of bringing the AI and robotics and other automation tools together in a way to accelerate our discovery of new knowledge, as well as new materials that could be used for energy applications like batteries, solar cells, thermoelectrics, as well as different chemical engineering processing, structural materials, simply all the materials that make up our everyday products and use cases.
What, in your opinion, is the most challenging aspect of establishing an SDL (Self-Driving Laboratory)?
That's a great question. I'm sure there are a lot of people that are much more qualified to answer that myself, but if I were to take a stab at it, it's that you can spend a year working on a self-driving laboratory, collect your data in three days. And there's this question of "what is the startup cost of setting it up or deceleration time that you actually get in running the experiments?"
I think there's some big challenges with making the system modular and reconfigurable so that if you did spend a year sort of cost to the first one, that you make that you spend a much shorter time, and the second one you may do the different iterations and also working to generate datasets, generate documentation, examples and other value added components along with it.
And then there's also the issue of maybe you're going to get some silly answers. You know, your algorithms have to be able to handle and recognize that, or you need to be well integrated with the optimization routine so that you can identify when things go wrong.
So those are kind of two tooling areas where there's a big challenge and kind of related there, where the the person really needs to be involved in the process, using it as these assistive tools, but make sure that it's the state of the art tools are being brought in in order to get that acceleration factor for the research discovery.
I have to ask, because it is on your GitHub, which, again, has been become very well known and kind of even like shout it out the other day on Twitter for how many commits and how much work you've done on the GitHub community: it says you like to breakdance - please divulge.
I'm getting back into it! I did it about a decade ago. In high school, I got started breakdancing. My friends were great. I kept going with that into college. I joined the big dance club at Brigham Young University and from there I became president for a little bit. We were the BYU dance group in a competition, with I think about 1500 dancers. They came and performed, and we we took first place amongst a lot of really, really talented people, amazing friends.
I started spending a lot more time in the scientific research side of things, and this resulted in really feeling the need to go back into that. And so I've been trying to, you know, several times a week just get outside a place, do some breakdancing. That's a lot of fun.
Without further ado, what was the purpose of creating your latest project, the Closed-Loop Spectroscopy Lab?
There is a huge startup cost involved with making all of these self-driving laboratories or some national labs where they have these incredibly expensive lights for the synchrotron, for example. There is a lot of automation occurring there, but the hassle is immense and the skills and expertise required are also very high. So there's a large barrier to entry for students, even for researchers and professors.
In order to break into this field, generally, they need a grant specific for this. Lots of grant money is spent on existing equipment that isn't very amenable to automation or the autonomous laboratory setup. And so a big reason for this is to help lessen that barrier. And there's also some really great teaching examples out there. This one takes it to an even lower cost, a simpler task, and really reduces it to a the encoding terms, a minimal working sample or a "hello world" for these advanced laboratories. So designing it and looking at what that might look like in a very small, contained scale that could be used in a classroom, that could be used for prototyping things and stuff like going on in a very low cost and low risk setting certainly.
This is very applicable within this academic setting, introducing, you know, students within material sciences or within the program world, just to kind of go into this SDL environment.
What applications has the kit been used for?
So this project is very new. I think it was within the last six months that the idea even came about, and it's quickly followed a rapid development cycle during this upcoming semester [spring semester], a number of professors at different universities have expressed interest in using this in their class classrooms for their classes that are teaching several of them, I think for the first time on the part of the science of discovery.
One of the big applications that I think is worth noting is teaching the "hello world" of optimization algorithms. If you Google search "grid search vs. random search vs. Bayesian optimization", they'll come up with a lot of machine learning. To bring that a little bit more to the material sciences or physical experiments, what we do in this demo is we take in red, green and blue LSD powers and we take a light sensor to measure the intensities at different color wavelengths. And we optimize the red, green and blue to match our target color spectrum.
And so that becomes the optimization task, and we can run it research where we can uniformly sample points along each dimension. In the search space, we can do random search where we randomly points red items, and then we can use a more informed optimization algorithm, something called Bayesian optimization to much more efficiently and in many fewer iterations reach our target value.
You've already documented really well on your GitHub and especially on the GroupGets campaign page. What is the best way for a user of this kit to get started?
[On the campaign, you're able to choose if you want it come pre-assembled.] First, you get to get the kit to the point where it's fully assembled. It's up to the sensor attached and facing the idea. And you run the starting introductory notebook, a Jupiter notebook. On this notebook, it lets you run Python commands in it. And if you can get that working where you're sending commands to your heart rate and reading the sensor data back, that's sort of the first of what we envision is like the first when you feel like, okay, I've been, you know, run my first autonomous drive, so to speak, in a self-driving lab with with the experiment.
And then from there, there are many different, tutorial notebooks that cover more advanced optimization topics, things like databases and, you know, the list goes on. And so that was kind of the flow that I see some people taking. And then also things that we might not even envision, but ideas that you have for how to modify it, how to extend it.
You mentioned that this project's unique approach is that there are no robotic movements. What is the main reason behind that decision and what is kind of like again, was made behind that decision?
To capture the key principles of a self-driving laboratory, as soon as you involve something with robotic movements, generally, you're either looking at several hundred dollars for the entry point or you can keep it in this very low cost, low budget thing, but you have dramatically more time in troubleshooting things out, working in breaking sort of time and so forth.
There's just this tradeoff between the cost of the demo and the time it takes to set the demo up. And one of the ideas I had in mind was to make something that was less than $100 and to be set up in less than an hour. And so with that, I started this idea of using LEDs in stead of something robotic, removing solid materials or liquid materials.
Some of the key principles in this multimillion dollar setup that people have been learning and experience over years of work and years of sweat and labor and and trying to condense that down into something that we saw in the classroom where someone could use for prototyping before they write a grant proposal, for example.
Congratulations to you and Dr. Sparks on achieving being 1 of the Top 10 Finalists for Hackaday.io's "Save the World: Wildcard Contest." Out of the 370 other projects that were submitted for this category, how did the Closed-Loop Spectroscopy Lab demonstrate a focus in "sustainability, resiliency and circularity?"
I think there's a couple of things I would point out here. One would be that the simplicity. By design, there's a lot of work involved in making a simple design that is meant to be user friendly. And there are a lot of competing objectives and I tried to make sure to document all that. On one hand, I think there's that level of documentation and reproducibility that was a big part that made it stand out.\
And then the other part is that it really is targeting some incredibly real world applications. You've got all sorts of groups that are focused on zero carbon materials, reducing the carbon footprint from a material standpoint. You know, looking at the energy being a big part of that as well. And so it's kind of this it's a tiny demo with really big goals and a big vision of what it can bring.
There are some fantastic projects on there, and I was I was excited to place as one of the finalists.
Links Referenced:
- Hackaday Project
- GitHub Source Code
- Google Colab Notebook Tutorials
- "From Scratch" Build Instructions
- Batch Optimization: 3 Cloud-Based Experiments controlled by a Central "Brain"
- Presentation Based on Accelerate Conference Talk