Based on your own experience, what is a recommendation you’d make to someone getting into the space of self-driving labs for accelerated scientific discovery?
This question depends a lot on the audience and the goals of the audience, so feel free to clarify the assumptions about audience/goals in tandem with the advice.
My heavily biased recommendation would be to register for and complete " Introduction to AI for Discovery using Self-driving Labs" and keep an eye out for other AC microcourses as they launch. This assumes some technical background.
Perhaps also studying the following two articles: Self-Driving Laboratories for Chemistry and Materials Science and Review of low-cost self-driving laboratories in chemistry and materials science: the “frugal twin” concept.
I would recommend reviewing what has been done to automate the process of your interest, see which AI and automation tools can ACCELERATE the processes of your interest and how AI can help to make smarter decisions, not just simple automation of the process.
I definitely agree with @sgbaird on the course materials. They are very helpful!
There are lots of research groups that are working in this space that I think are great examples to show what is possible.
There’s also a lot of companies doing cool stuff as well for example Dow PaintVision.
There’s a lot more but that is definitely a sufficient starting point!
Different approaches will work for different people, but everyone has to start somewhere.
Building an entire SDL from scratch can be costly, time-consuming, and daunting. Yet, the fundamental concepts of automation and statistical-based sequential decision making in scientific discovery have been around for decades. They don’t need to be intimidating, we just have better, modern tools to implement them now.
I’d start with a particular experimental feedback loop that is determined to be a rate-limiting step in a discovery process (or just one that really, really annoys you). I would then work backwards, researching what is needed to achieve two goals: (1) how to do this process faster using available automation and high-throughput solutions, and (2) how to cycle through this loop more efficiently using AI-based sequential decision making as an alternative to conventional design of experiments.
I agree with all the previous comments. Researching existing tools and platforms is crucial for understanding the broader goal. I also suggest starting with a small or simple SDL project, such as the one detailed in this What is a minimal working example for a self-driving laboratory?, to familiarize yourself with the main components required to build an SDL.
There are a few points I find useful while start working in the area of SDLs:
• Explore existing training materials. As @sgbaird mentioned above by now we are fortunate to have some training courses on SDLs subject like Introduction to AI for Discovery Using Self-Driving Labs. It will help to get an initial overview of process design.
• Set up big goals but start simple. Establish your limits within desired chemical space to lower deployment barrier for your SDL
• Define your SDL plans. Decide would it be more favorable to follow fixed or flexible automation concept. For more details of benefits and disadvantages of each – check this paper Balancing act: when to flex and when to stay fixed
• Gain support/missing expertise on hardware and software sides. It will significantly slow down development if everything will be done by the same person. Create a team of experts with relevant knowledge. Collaboration is a key to fast development of SDL.
• Leverage previous expertise. SDLs are relatively young direction and often researches tend to create their own new solutions for known problems which leads to high divergence in the field. However, if we adopt common standards, we can boost rapid development more efficiently and get chemical autonomous systems to the next level
• Carefully plan your hardware and software set up based on your objectives. This perspective paper gives a lot of insights on this subject Automation isn’t automatic
• Be ready for challenges. There is a lot of troubleshooting (including big number of unforeseen challenges) before SDL will run smoothly