Catalyst · 6 Nov 2025 · 36m
Ekin Dogus Cubuk, co-founder of Periodic Labs, explains how his company combines frontier AI models with automated physical labs to accelerate materials discovery. Cubuk argues that recent advances in reasoning models like OpenAI's O1 and improved lab automation allow AI to overcome its reliance on limited training data by generating synthetic data and running high-throughput experiments. The company is targeting breakthrough materials, specifically high-temperature superconductors, using a hybrid approach where humans provide initial hypotheses and AI handles execution and iteration. With a $300 million seed round, Periodic Labs aims to build a generalized scientific intelligence platform that serves both its own discovery goals and external partners' R&D needs.
Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
Engineering & simulationOpenAI O1
The team used the O1 reasoning model to overcome AI's limitation in predicting beyond its training data by spending test-time compute to improve results for materials discovery.
“So what O1 showed is if you spend test time compute, you can get better results. So that was very exciting to me because there was one way of investing resources that was beyond the training set.”
Speaker 2 · Hear it at 5:36
Engineering & simulationcustom LLM
The LLM proposes synthesis recipes and simulations, then uses tool-use capabilities to execute them and analyze the resulting experimental or simulation data to tweak the next step.
“The LLM can propose, for example, synthesis recipes, or it can propose simulations to run. And because the LMs are pretty good at tool use. It can actually do it itself.”
Speaker 2 · Hear it at 9:54
Engineering & simulation
The team leverages LLMs' ability to understand both state chemistry and condensed matter physics simultaneously to generate hypotheses and explore interdisciplinary areas that human experts might miss.
“And turns out this is true for many different aspects … And this is one of the reasons I think we are very excited about LLMs because when you talk to them, they seem like they have a pretty good understanding of South-state chemistry and South-state physics at the same time already.”
Speaker 2 · Hear it at 25:14
early-stage companies
Other shows on the same ground, closest first.
Hosts Daniel and Farbode discuss two AI applications in physical engineering: Dean Dang’s use of AI and electron microscopy to reverse-engineer lithium iron phosphate (LFP) battery degradation at Stanford, and Michael Barrens and Dr. Warren Rooter’s training of micro-robots to swim autonomously using a soft actor-critic algorithm at Pitt and Carnegie Mellon. They highlight how AI accelerates material discovery and control system development for medical micro-robots, moving beyond simulation to real-world physical learning.
Hosts Daniel and Farbode discuss two AI applications: Penn State scientists using Generative Adversarial Networks (GANs) to discover new high-temperature alloys, and Carnegie Mellon University researchers developing an AI system to automate process management and replace human managers. They analyze how GANs mimic the creative/critical dynamic of human teams to generate realistic material designs, and how an AI manager trained on communication data outperformed humans in a drone design simulation by avoiding counterproductive micromanagement.
Gabriella Pizzuto, a Lecturer in Robotics and Chemistry Automation at the University of Liverpool, discusses deploying intelligent robotic manipulators to accelerate material discovery in chemistry labs. She explains how her team uses dual-arm mobile manipulators to handle powders, solvents, and glassware, aiming to run experiments around the clock while minimizing human exposure to chemicals. Pizzuto highlights the shift from pre-programmed open-loop systems to robot learning, where AI-driven methods like Bayesian optimization guide chemical reactions based on sensor feedback. She also notes that while data interpretation is easier, the physical challenge of making robots robust enough to handle non-standardized materials in semi-cluttered lab environments remains significant.
Eduardo Torrealba of Lumafield, Vikram Pavate of Tacta Systems, and Pari Singh of Flow Engineering discuss how physical AI is bottlenecked by the lack of specific, non-public data. They argue that while software AI thrives on open internet data, physical products require proprietary datasets from CT scans, tactile sensor gloves, and engineering context to train effective models. The panel highlights the 'missing middle' of manufacturing—products costing $5 to $500—as the key area where automation and AI can enable reshoring to the US. Each founder shares a 'punched in the mouth' moment: Lumafield's scanner durability issues in unconditioned factories, Tacta's glove being rejected for being too soft, and Flow Engineering's decision to burn eight years of code to rebuild for an agentic future.
Matt Puchalski, founder and CEO of Bucket Robotics, explains how his Y Combinator-backed startup uses synthetic data generated from STEP files to automate visual inspection in manufacturing, bypassing the need for physical 'golden samples.' He contrasts this approach with legacy systems from Keyence and Cognex, highlighting how starting from CAD models allows for rapid simulation of good and bad parts. Puchalski also shares his background in autonomous vehicles at Argo AI and Latitude AI, discussing the unique reliability challenges of deploying Level 4 AVs and how those lessons inform his current work in flexible manufacturing.
The summary, chapters and topics are generated from the episode’s audio by this directory, not written by Catalyst. The episode itself belongs to its publisher.