For those interested, here is the paper and the GitHub repo. Happy building!.

The true metric for success here is not number of samples, but accelerated cycles of learning — a truly lagging indicator. Keep an eye on the Fenning lab at UCSD to see the fruits of PASCAL’s labor in the pipeline.

I am personally a huge advocate for autonomous experiments, but in my experience these workflows shine best in large-scale and repetitive experimental campaigns that are uncommon in university labs.

While capable of “self-driving” using Bayesian Optimization, >95% of samples were designed by human operators. The art of coaxing the machine too often gets in the way of the science — especially when experimental goals change too quickly to justify “getting it right”.

PASCAL was “down” for 1/3 of the time, with average ~2 months of operation between failures. PASCAL had ~10 trained operators from grad students to postdocs. (Note I’m assuming that any >2 week idle period is a system failure. This is an overestimate — people take vacations!).

In a typical (median) week, PASCAL made 29 samples. Sample throughput ebbs and flows with the demands of ongoing projects in the lab; we made over 100 samples in 1/3 of the operational weeks!

PASCAL is designed to make and measure perovskite solar cells. One “sample” could be a 30 minute heating test, or a complete solar cell. In its best 24 hour period, PASCAL performed 400 samples -- guess which kind that was 😅

I built a robotic lab in grad school. Last week it made its 10,000th sample. Here's some stats from three years of operation.

RT @SunShijing: We're hiring! We are looking for a postdoctoral researcher in energy materials and lab automation. Join us at the beautiful….

RT @berkeleygfx: McMaster-Carr, the gold standard of web design. They didn't have anyone to look up to, there is nothing like it. They didn….

Data-driven materials folks — how are you storing all your data? Especially in the context of automated experiments. Files and folders? Bespoke relational databases? Anyone using GEMD from Citrine? . Would love to discuss pain points!.

RT @AdamMGrant: When groups meet to brainstorm, good ideas are lost. People bite their tongues due to conformity pressure, noise, and ego t….

Human-AI partnership is the way forward.

PhDone! Had an amazing five years with great labmates+collaborators turned friends. Looking forward to my next chapter in Berkeley. Was also great to finally introduce my parents to their grand-robot 👶🏽🤖

RT @DanedeQuilettes: The long awaited Future of Energy Storage from @mitenergy is out! Over the last 3 years, experts in the field have tho….

Excited for #S22MRS, will be sharing our successes (and failures) in closed-loop robotic optimization of process/composition for halide perovskites (Tuesday DS01.05.03). Stop by and say hi!

This was a great collaboration with @TiihonenArmi @SunShijing @dpfenning @LiuZhe_MIT @toniobuonassisi and many fruitful discussions across academic and commercial #perovskite #solar. Typical tech adoption takes 10+ years — can we speedrun with ML?.

Finally, we envision a future where "gamechanger" ML like inverse design, natural language processing, and transfer learning enable rapid incorporation of new results into and across production lines.

We highlight “ready-now” ML methods (mostly demonstrated in halide perovskites) to identify optimal processing windows, to digest metrology, and to extract physical parameters from characterization data.

Four primary challenges from industry are 1) maintaining a baseline process, 2) performance parity between small- and large-area devices, 3) identify root causes of underperformance, and 4) developing in-house data science talent.