Scientific research is a messy business. The road to learning new things and making discoveries is paved with hard labor, tough thinking, and plenty of dead ends. It’s a time-consuming, expensive endeavor, and for every success, there are thousands upon thousands of failures.
It’s a process so inefficient, you would think someone would have automated it already. The concept of the self-driving laboratory aims to do exactly that, and could revolutionize materials research in particular.
Leave It To The Auto-Lab
Materials research is a complex field and a challenging one to work in. Much of the work involves finding improvements on existing materials to make them harder, better, faster, or stronger. There’s also the scope to discover entirely new materials with unique properties and capabilities beyond those already known to us.
In the modern world, it’s not enough to just go outside and dig up a new kind of rock or find a new kind of tree. All the low-hanging fruit are already gone. Materials research now requires a sound understanding of physics and chemistry. It’s all about figuring out how to exploit those principles to make something better than what we’ve seen before.
This is where artificial intelligence and computers come in. Rules we’ve discovered in chemistry and physics can be programmed into an intelligent system. It’s then a straightforward leap to have that system apply those rules in varying ways to optimize desired outcomes. For example, an AI system can be asked to synthesize a given chemical in the most efficient way possible given a certain set of precursor chemicals. It’s then possible for the AI to run through all the possibilities and determine the best course of action.
Where the concept gets most compelling, though, is where an AI system is given the capacity to run its own experiments in the real world. Laboratory automation is advanced to the point where robots can readily run experiments far quicker and more efficiently than human scientists can do by hand. Give the AI the hardware to do experiments and to measure the results of its work, and it can then use the results to guide further experiments towards its given research goals. Congratulations – you’ve built a self-driving laboratory!
In Practice
Far from a mere theory, researchers around the world are already building self-driving laboratory systems. One of the most well-known is the so-called Artificial Chemist, developed by researchers at the University of Buffalo and North Carolina State University. The project’s goal was to develop an automated system to perform chemistry research for the research and development of commercially-desirable materials.
It’s designed to perform chemical research into materials that can be made using liquid solutions. The system is tasked with finding a way to synthesize a material that meets a set of desired parameters, and performs experiments on its own to determine how to achieve that. In testing the system was tasked with synthesizing quantum dots with various desired parameters. Through experimentation, Artificial Chemist was able to figure out ideal techniques on how to make the dots, including the identification of the correct chemical precursors.
Far from a simple computer simulation, Artificial Chemist does real chemistry on its own, and measures the results. The system was outfitted with chemical reactors that are entirely autonomous. They’re also designed to remain clean without picking up chemical residues that would throw off the experiments. The system can mix chemicals and run an entire chemical synthesis all on its own.
The system was developed with an eye to both research and manufacturing. It can be tasked to produce quantum dots for a given wavelength of light, and will first spend time doing research experiments to determine the best way to make them. Once that process is complete, usually after 1-10 hours, the system can then begin producing the dots en masse.
Research+
Overall, though, the basic principle can be applied to all kinds of research processes. One need only give a suitable AI system the means to experiment and the means to examine the results of its work. It can then take the logical steps to further its work in the direction of its given research goals.
The benefits of such systems are manifold. Where parts of experiments may have been automated by robots before, self-driving laboratories go further. They enable scientists to set a goal and the automated lab works its way to a solution entirely indepdently. This enables research to be carried out with less labor and human effort, with progress made far faster and far cheaper than before. Plus, the ability for quick calculation and experimentation may allow an AI to quickly run tests on combining regular ingredients in unexpected ways, netting surprise unconventional results. Some researchers expect these systems to provide a tenfold benefit to costs and time, where goals that once took ten years and $10 million dollars completed in one year and for just $1 million.
Of course, such systems won’t make human researchers obsolete. Creativity is of huge importance in science and engineering disciplines, and has led to some of our biggest advances. For example, an AI could be tasked to make stronger and more lightweight metal alloys. However, given those human-spawned preconceptions, it would never come up with the brilliance of composite materials like carbon fiber.
A great corollary is the image synthesis AIs which have skyrocketed in popularity this year. Initially, hyperbole stated that artists and photographers would be out of a job and human endeavour in this field was over. Then, weeks later, it turned out that these were just a new kind of tool that could be guided and put to work by humans best experienced in exploiting them.
These “self-driving laboratories” will likely become major tools in industrial R&D labs, doing everything from developing new materials to uncovering new molecules of potential medical interest. Talented research scientists will work to best employ the robotic resources they have, ensuring they’re put to work in the most effective manner for their broader research goals in general. With much of the research drudgery handed off to the robots, that will leave human scientists more time to think about the bigger picture.
Banner image: © xiaolangge / Adobe Stock.
Programmed to deal with the occasional BOOM!
Now, I am also curious about how they deal with accidental (or sometimes intentional) KABOOMs. A workaround that might already be in place would be that individual experiments are done in very small scale, probably in magnitude of microlitres or so.
Having been in this field, it is amazing how sensitive experiments can be, and how much luck is involved in finding the right combinations for things to work. Drug candidate screenings are already hyper automated, and it’s exciting to see science discovery do the same! It’s like a “for” loop for reality!
A friend of mine told me that the majority of the time spent in a bio lab would be to wait for each instruments to finish their job.
It’s more complex, but feels a bit like the computer programs that play chess. Maybe a researcher will one day write an exclamation point next to a result that surprises everyone.
Interesting and has obvious applications. There is of course that laboratory experiment that has been operating for over 3 billion years, namely evolution on planet earth. Numerous iterations have resulted in a mass of compounds perfected to perform in a vast array of conditions. Most of the compounds have yet to be discovered and investigated but many have already been adapted for use by humans. I’m sure the autopilot laboratory will come up with similar compounds but will be hard put to replicate the range of materials already produced by various lifeforms. There’s obviously a place for the laboratory but it will not displace the search for useful compounds to be found in our natural environment.
Neat. With computers to run simulations that a ‘creative’ scientist puts in motion is one thing, but then if the scientist has encouraging results, I can envision then the person can hand off the results to one of these ‘labs’ to validate, and if good, it can figure out a way to produce the material (whether for a building material, drug, whatever) efficiently…. Save a lot of time. Sounds wonderful…. Unfortunately computers/robots only do what they are programmed to do, so a tough nut to crack for the ‘general’ case as at that point the scientist might as well do it him/her self as he/she has to do the programming to do the testing and setup a lab for that case… Sort of the chicken or the egg dilemma.
Yesterday there was a big story about AI and how it doesn’t necessarily do what it’s programmed to do, it can do things above and beyond that. If an AI can make inferences about x ray results, or compute object shapes from 2 d images, then it can most certainly make inferences about chemical reactions. It can memorize that organic chemistry textbook a lot better than any human. It can stay up all night running thousands of experiments that no human would be able to accomplish, gathering enormous amounts of data to use for further analysis.
Someday we will be able to tell the robot to “find a better food preservative” or “find a cheaper way to produce this drug” and it will do it.
No human can tell the robot how to do these things, if we could, then we wouldn’t need the robot.
> It can memorize that organic chemistry textbook a lot better than any human.
Sure, but whether it actually understands any of its implications is a different question. The AI is good at finding correlations, it’s weak at identifying causation, and bad at inferring anything beyond the data and knowledge it is already given. That’s because there are an arbitrary number of theories/models that can be made to fit a data set, and a subset of those will even turn out correct predictions just outside of the data set, yet only one of them will be the correct model or theory, so the AI is statistically likely to find a plausible but wrong model to the phenomenon it is studying.
Dumping a load of data in and letting the AI loose on it is cargo-cult researching. You can easily end up amplifying random artifacts, so you have to constrain the AI to test particular models, or use it to come up with candidates for modeling. It’s not press-button-receive-science kind of deal.
Of course there’s the version where you just try every possible permutation and see what happens, and then iterate the same on the results to explore every possible branch, but that is both prohibitively costly and slow even with automated robots because the number of possible experiments you can make quickly explode towards infinity – and it’s technically not “AI” to begin with…
Reminds me of this oldish (Harlem Shake era, so some 10 years ago?) project[0], involving the reuse of different automated machines and instruments on a lab, with some glue code and drivers to get all working. An incredible feat.
[0] http://www.zachbjornson.com/projects/robotics/
the big problem will be the patent trolls that will stop the development
Imagine just brute forcing the discovery of all possible wonder drugs in this way. All the AI needs is a constant input of raw materials and test subjects. Actually that gives me an idea for a screenplay.
Well, if you start with three substances and combine them to create two new ones, you then have to combine each with the one you didn’t include in the first place, and between each other, and then cross check the results with all the other stuff… the task at hand starts to branch out to more and more combinations very quickly if you’re simply trying every possible thing.
Just don’t let GLADOS run the experiments!
Sounds like a great tool to formulate new licit novel recreational designer drugs on an unprecedented scale!