Ekin Dogus Cubuk is a researcher at Google DeepMind where he leads the materials science and chemistry team.
He received his Ph.D. from Harvard University where he studied the physics of disordered solids and battery materials using density functional theory and machine learning.
After a brief postdoc at the Materials Science Department of Stanford University, he joined Google Brain in 2017. Since then, he has been studying the scaling and out-of-domain generalization properties of large neural networks, and their applications to materials discovery for applications including clean energy and information processing.
Can AI Revolutionize Materials Discovery?
Artificial intelligence, and especially deep learning, is generating excitement across many scientific disciplines—including materials discovery. But despite remarkable progress on academic benchmarks, major challenges remain in applying these models to real-world discovery tasks. In this talk, I will explore the question: Can AI truly revolutionize materials discovery?
Drawing on examples from computational materials science, I will highlight where AI tools have made genuine progress—and where they have fallen short. For instance, I will discuss when the ability to predict formation energies of inorganic crystals may not translate into more efficient discovery of new 0K stable materials. These kinds of gaps reveal the difference between academic benchmarks and more realistic performance, and underscore the need for skepticism, nuance, and better evaluation strategies. Along the way, I will discuss how concepts like active learning, training size scaling, and hybrid approaches that combine physics-based and data-driven methods are shaping the future of the field.
