Ekin Dogus Cubuk

Ekin Doğuş Çubuk is a materials scientist and Periodic Labs co-founder working to connect machine learning with physical experiments. His work addresses a practical difficulty: a model may predict that a crystal is stable, yet scientists still have to find a way to make it and measure what they made. At Periodic he is building tools to close that loop.[1][2]

At Google DeepMind, Çubuk co-authored the account of GNoME, a system for exploring candidate crystal structures. One pipeline proposed structures related to known crystals; another explored chemical compositions more broadly. A graph model estimated stability, and density functional theory calculations evaluated selected candidates and fed new data back into the model. The resulting candidates were computational predictions, not a guarantee that each material could be synthesized or would possess a desired property.[3]

Periodic’s physical lab takes the next problem seriously: mixing ingredients is not enough if nobody can tell what came out. Çubuk identifies automated characterization as a bottleneck. Measurements such as X-ray diffraction reveal patterns associated with the arrangement of atoms; interpreting those patterns lets the team identify phases, incorporate simulation, and make a more informed choice for the next experiment.[1]

Çubuk argues that experiments differ from a fully specified mathematical problem because scientists never record every detail of the physical system. A furnace’s heating element can change as material accumulates on it; placement inside the furnace can change the temperature experienced by a sample. Even an optical measurement can be disturbed by vibration. The point is not to model every atom perfectly, but to choose useful representations and account for the information that measurements miss.[1]

That incomplete picture also limits simulation. Çubuk describes density functional theory as powerful for some systems but difficult to apply reliably to strongly correlated electrons, where interactions matter in ways a simple approximation can miss. A prediction therefore helps choose an experiment; it does not eliminate the need for one.[1]

Çubuk’s longer-term goal is to make synthesis more systematic. He points to nickelates: a promising chemical family did not yield superconductivity in an early search, but later work demonstrated superconductivity in thin films. The example illustrates why a scientific idea and the physical route to a material are separate challenges. His proposed automation would expand the number of candidates a lab can test; it remains an ambition rather than a demonstrated recipe for a room-temperature superconductor.[1]