In experimental chemistry research, people come up with hypotheses and test them. Jan Estrada Pábon’s thesis began by developing a procedure that would allow computers to develop these hypotheses themselves, instead.
“Having this capability would allow us to have machine-guided explorations instead of biasing our chemical exploration through human biases,” he said. “An example would be directly in Alex Norquist and Sorelle Friedler’s lab.”
Professor of Chemistry Norquist and Associate Professor of Computer Science Friedler collaborate across labs to use computers to understand the conditions under which perovskite crystals form. Estrada’s chemistry thesis outputs a set of chemical conditions that are necessary to form a crystal.
“My thesis’ stuff essentially can be used to give scientists a conceptual understanding of what’s guiding these crystals’ formation,” said Estrada. “This, in turn, might allow us to make better crystals!”
Estrada is considering continuing this research in his doctoral project. He is headed to Stanford University in the fall to pursue a Ph.D. in theoretical chemistry.
“I had a lovely four years,” he said, “but I’m ready to be somewhere near a beach again.”
What did you learn from working on your thesis?
Personally, I learned to stay motivated while locked up most of the day, to know when to lean on my friends when I needed support, and to never skip out on meals. In terms of research, I learned that it is possible to have a conversation with machine learning (ML) models through experiments. I hope a takeaway from my project is that we will be able to improve ML model building in a more efficient manner by better understanding what it is learning and testing for that directly until we (hopefully) find something completely discovered autonomously.
What are the implications for your thesis research?
I think it means some other senior will have a lot more work than I did. I think my thesis is not exactly finished, which is fine because I only had about a year to work on this. However, this thesis shows that I created the early stages of a workflow which will allow us to create experimentally testable hypotheses completely autonomously. … If other researchers implemented these similar strategies it would allow them to generate data in an unbiased fashion, maybe. Who knows?