Key Moments
Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms
Stanford OnlineKey Moments
AI is being used to accelerate scientific discovery, but the path to autonomous labs is more challenging than anticipated, requiring hands-on, iterative development.
Key Insights
Periodic Labs, founded by former OpenAI and DeepMind researchers, aims to apply AI to physical science, specifically in material discovery, with a 40,000 sq ft facility in Menlo Park.
The initial assumption that the first year would be spent in silico (computer simulation) for material design was incorrect; building semi-manual, semi-autonomous labs allowed for faster feedback loops.
AI is crucial not only for predicting new materials but also for handling routine tasks like proper powder mixing and impurity detection, which are essential for scientific progress.
While large language models excel at tasks with clear verification (e.g., '2+2=4'), scientific discovery involves ambiguity, requiring more time to understand problem nuances before defining evaluation metrics.
The project's focus is on semiconductors and superconductors, driven by the immense demand for more advanced computing hardware and the material science bottlenecks hindering Moore's Law.
Active learning is highlighted as a critical technique for scientific discovery, allowing models to gradually expand into unknown areas and learn from real-world experiments, unlike static academic datasets.
The unexpected path to autonomous labs
Periodic Labs, co-founded by Liam Ferriss (OpenAI) and Dorje Chubak (DeepMind), is applying AI to the physical world, aiming to accelerate scientific discovery, particularly in materials science. Their initial vision for the first year involved purely in silico (computer simulation) design of new materials, followed by scaling up to a high-throughput, autonomous lab. However, this assumption was quickly revised. The team found that building smaller, semi-manual, semi-autonomous labs was more effective. This approach allowed them to guide the research program, understand which equipment to scale, and, crucially, close the feedback loop much faster. This iterative, hands-on development proved essential for navigating the complexities of physical experimentation and has been a significant update to their initial strategy.
AI's dual role in material discovery
The application of AI, particularly large language models, extends beyond just predicting promising candidates for material properties like superconductivity. AI systems are proving invaluable in managing the tedious, routine aspects of scientific experimentation that are critical for progress. These include tasks such as ensuring accurate powder mixing, identifying and correcting sample contamination, and optimizing experimental procedures. These seemingly minor, routine steps, when automated and optimized by AI, collectively form the backbone of efficient scientific research. This demonstrates that AI's impact in science is not solely about high-level prediction but also about refining and streamlining the fundamental experimental processes.
Focus on semiconductors and superconductors
Periodic Labs is concentrating its efforts on semiconductors and superconductors due to their profound impact on technology and the significant material science challenges they present. The relentless demand for more powerful computing hardware, as outlined by the principle of Moore's Law, is increasingly hitting material engineering bottlenecks. Developing new superconductors and improving semiconductor properties are seen as key to overcoming these limitations and enabling future technological advancements. The project's hypothesis is that by focusing AI's capabilities on the fundamental interactions between atoms and electrons – governing superconductivity and semiconductor interfaces – they can unlock significant progress. This focus is strategic, aiming to address critical bottlenecks that affect computation, a physical process at its core.
The challenge of scientific ambiguity
While AI has made tremendous strides in tasks with clear verification, such as solving mathematical problems, scientific discovery inherently involves ambiguity and uncertainty. Unlike a simple calculation where the answer is definitively known, scientific research often involves interpreting fragmented evidence from various tools and experiments. Periodic Labs emphasizes the importance of spending considerable time understanding the nature of a scientific problem before committing to specific evaluation metrics or rushing into experimentation. This nuanced approach acknowledges that AI must be adept at navigating these gray areas, piecing together disparate information, and guiding research in complex, ill-defined domains.
The 'ONES' system and its inspiration
The AI system developed by Periodic Labs is named 'ONES', inspired by Heike Kamerlingh Onnes, the physicist who first liquefied helium and discovered superconductivity. The name also reflects Onnes's pioneering approach to scientific research, advocating for industrial-scale laboratories even in the early 20th century. This ethos of conducting science with seriousness, at scale, and with a sense of urgency resonates with Periodic Labs' mission. Compared to systems like ChatGPT, which focus on language and code, ONES possesses a much deeper understanding of the physical world. It's designed to engineer new systems, predict their stability, and guide their fabrication, moving beyond abstract concepts to tangible material outcomes.
Active learning: a key to real-world AI
Active learning is presented as a crucial methodology for scientific AI, particularly in contrast to the static datasets typically used in academic machine learning. While standard training and testing splits on datasets like ImageNet might not benefit from active learning, real-world applications, such as training self-driving car systems like Waymo, rely heavily on it. Active learning allows AI models to identify their weaknesses and focus on gathering new data in those specific areas. For Periodic Labs, this means using new experiments daily to probe the boundaries of their current models, continuously improving their generalization capabilities in the complex, less predictable environment of scientific research.
Defining 'new material discovery'
The discovery of new materials at Periodic Labs encompasses several dimensions. It can refer to identifying entirely novel crystal structures not previously cataloged. It also includes finding materials with unprecedented properties, such as superconductivity at higher temperatures (e.g., above 133 Kelvin at ambient pressure). Furthermore, the discovery of new manufacturing 'recipes' is valued, even for previously known materials, if they can be synthesized more efficiently or with improved characteristics. Ultimately, material discovery is framed as advancing human control over the arrangement of atoms, pushing the boundaries of what can be engineered to build advanced technologies.
The future of AI in the physical world
The overarching goal is to make AI systems more intelligent in their interaction with the physical reality, bringing science fiction futures into the present. While AI has revolutionized digital domains, the application to the physical world is seen as the next frontier. The belief is that as AI and robotic systems improve, the barriers to creating new labs and infrastructure will decrease. The current work at Periodic Labs is considered among the most difficult, with the expectation that future iterations will be faster and less capital-intensive. The drive is to move beyond the limitations of digital tasks and apply AI's power to tangible scientific progress, with an endless scope for discovery.
Mentioned in This Episode
●People Referenced
Common Questions
Periodic Labs is a company co-founded by Liam Ferris and Doug Chip that uses AI to discover new materials. Their primary focus is on finding materials for high-temperature superconductors, aiming to accelerate scientific discovery and technological advancement.
Topics
Mentioned in this video
More from Stanford Online
View all 142 summaries
60 minWebinar: What AI Can and Cannot Do: Intelligence Augmentation in Practice with Michael Bernstein
61 minWebinar: AI Agent Simulation of Human Behavior with Michael Bernstein
105 minStanford CME295 Transformers & LLMs | Autumn 2026 | Lecture 1 - Transformers
41 minAI in Healthcare Series: Have We Already Bent the Healthcare Cost Curve?
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free