idalab seminar #13: Exploring Chemical Space with Deep Learning
What if we could build batteries for electric cars that would take us further than a full tank of gasoline? If we could grow affordable, tasty and nutritious meat in the laboratory instead of occupying one third of the land on our planet with animal farming? What if we could easily identify promising targets in the human body for new cancer drugs?
Underpinning all of these challenges is: chemistry. The challenge is not a scarcity of potential high-value compounds. Quite the opposite: The problem is that it just takes too long to test every possible compound for the desired chemical properties. Therefore, we must find ways to reliably predict these properties.
While quantum-chemical simulations allow us to accurately calculate chemical properties, their large computational cost as well as the huge number of molecules and materials make an exhaustive exploration infeasible. This talk introduces deep learning models for a variety of use cases in quantum chemistry. By analyzing the learned representations, we get a glimpse into the inner working of the neural network to find out whether the model has learned known chemical concepts – or has even uncovered hitherto unknown mechanisms!
Friday, September 28th, 5 pm | doors open at 4.30 pm | Potsdamer Straße 68, 10785 Berlin
About idalab seminars: idalab seminars are open to all interested parties. Once a month, we invite scholars, data scientists, business experts and big data thought leaders to discuss their work, gain new perspectives and generate fresh insights.
After the talk, we invite you to stay for drinks. We’re looking forward to seeing you there!
Dr. Kristof T. Schütt is currently a postdoc at the machine learning group at Technische Universität Berlin and the Berlin Big Data Center. His primary research interests cover learning from structured data, deep learning as well as its application to quantum chemistry. Furthermore, he has been working on explaining neural network decisions and building interpretable models.