ABAB Pedia

Ekin Dogus Cubuk

Overview

Periodic Labs combines models and physical experiments to discover materials.⁠[1]

Contents13 sections
Key facts

Degrees

Doctoral Research: Atomistic Systems

His doctoral work examined structural complexity in disordered solids and catalytic surfaces, combining first-principles calculations, machine learning and stochastic simulation for applications including supercooled liquids and lithium-ion batteries.⁠[3]

Joining Google Brain

After a brief postdoc in Stanford University’s Materials Science Department, he joined Google Brain in 2017, studying large neural networks’ scaling, out-of-domain generalization and materials discovery.⁠[4]

AutoAugment

On June 4, he and Barret Zoph introduced AutoAugment, which uses reinforcement learning to select image transformations, probabilities and magnitudes for automated data augmentation. The team also reported transfer of learned policies between vision datasets.⁠[5]

RandAugment

In 2019, he, Barret Zoph and colleagues released the RandAugment preprint, reducing augmentation to a transformation count and shared magnitude without a separate proxy-task search. The NeurIPS 2020 paper evaluated image classification and object detection.⁠[6]

GNoME

On November 29, he and Amil Merchant presented GNoME, combining graph neural networks with density-functional-theory calculations to predict crystal stability. The team reported 2.2 million candidate structures, including 380,000 on the updated stability hull; these figures are not counts of synthesized materials.⁠[7]

Assessment of Materials Research

In their April commentary, Anthony K. Cheetham and Ram Seshadri questioned the evidence for combined novelty, credibility and utility, calling for synthesis and crystallographic expertise.⁠[8]

Periodic Labs

On September 30, Andreessen HorowitzAndreessen HorowitzAndreessen Horowitz (a16z) is a venture capital firm founded in 2009 by Marc Andreessen and Ben Horowitz. It invests from the seed stage through the growth stage and provides operational support to its portfolio companies.Open the full entry announced it was leading Periodic Labs’ $300 million founding round. The announcement described Cubuk and Liam Fedus’s collaboration on AI for science, emphasizing real experimental data.⁠[9]

Periodic Neon

On September 15, the company introduced Periodic Neon, describing a trillion-parameter model post-trained on its experimental data and deployed in its labs to interpret X-ray diffraction results. Its reported performance comparisons concern that analysis task, not general capability.⁠[10]

Research Approach

Scientific discovery requires experiments, he argues; simulation alone is insufficient.⁠[11]

Related Organizations, Websites, and Public Accounts

Periodic Labs: Official website: https://periodic.com/ (opens in a new window)⁠[1]

Sources

  1. Periodic Labs (opens in a new window)
  2. Ekin Dogus Cubuk runs a startup to accelerate physics R&D using AI (opens in a new window)
  3. Ekin Dogus Cubuk (opens in a new window)
  4. Applied Physics/Physics Colloquium: Ekin Dogus Cubuk- Combining Experiments, Large Language Models, and Theory to Discover Quantum Materials (opens in a new window)
  5. Improving Deep Learning Performance with AutoAugment (opens in a new window)
  6. RandAugment: Practical Automated Data Augmentation with a Reduced Search Space (opens in a new window)
  7. Millions of new materials discovered with deep learning (opens in a new window)
  8. Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery (opens in a new window)
  9. Investing in Periodic Labs (opens in a new window)
  10. Building Labs that Learn (opens in a new window)
  11. Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk (opens in a new window)