Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms

Company Overview and Team Composition

• Periodic Labs was founded by Liam Fettis and Doge Chubbuck approximately 11 months ago.
• The company operates out of a 40,000-square-foot facility in Menlo Park.
• The team comprises 50% former ML researchers from OpenAI, DeepMind, and similar organizations, and 50% physicists and chemists from Stanford, MIT, and Caltech.

Operational Pivot and Strategy Updates

• Founders initially planned a "silica-only" first year focused on computational design before launching physical labs, but pivoted to a semi-manual, semi-autonomous lab approach within months.
• The pivot allowed the team to direct research programs, identify necessary equipment, and close the feedback loop between AI prediction and physical verification faster than anticipated.
• The team updated their priors regarding LLM capabilities, discovering that AI could impact "real atoms" more effectively than expected, leading to faster material synthesis and computational progress.
• Doge Chubbuck noted that results of automating intelligence affecting physical matter were "crazier" than initially expected.

The "Anas" AI Agent

• The company's primary AI system is named "Anas," honoring Heike Kamerlingh Onnes (often referred to as "Onnes" or "Anas" in colloquial contexts) who discovered superconductivity and pioneered industrial-scale scientific research.
• The agent's name reflects the mission to conduct science at an industrial scale with urgency, moving beyond trial-and-error to intentional engineering of matter.
• Unlike systems focused on coding or abstract reasoning, Anas is designed to orchestrate tool calls for physical world tasks, including synthesis and verification of materials.

The Physical Discovery Pipeline

• Prediction: The AI predicts new materials, specifically targeting high-temperature superconductors and semiconductor materials to alleviate current engineering bottlenecks.
• Synthesis: Robots physically synthesize the predicted materials into specific forms (e.g., powder).
• Verification: Automated machines characterize the materials (e.g., via X-ray diffraction) to verify if they possess the predicted properties.
• Feedback Loop: Verification data is piped back into the training loop to refine the AI's predictions.
• Error Correction: The AI handles mundane but critical tasks such as detecting sample mix-ups, identifying impurities, and ensuring correct powder mixing ratios.
• Scope: While focused on superconductors and semiconductors, the underlying physics of atomistic interactions (phonons, electrons, magnetic properties) is applicable to polymers and other advanced industries like aerospace and energy.

Focus Area: Semiconductors and Superconductors

• The team selected semiconductors/superconductors as their primary evaluation domain due to critical industry bottlenecks: current materials engineering limits the scaling of logic and memory chips.
• The goal is to reduce energy loss in data centers and computing hardware; current systems lose significant energy (estimated double-digit percentages) to heat and resistance.
• Discovery is defined not just by new crystal structures, but by:
  • Materials with superior properties (e.g., ambient pressure superconductivity above the current ~133 Kelvin state of the art).
  • Optimized synthesis recipes that improve efficiency or yield of known materials.

Technical Methodology and Model Architecture

• Active Learning: The team prioritizes active learning over Bayesian optimization to push the frontier of understanding into unknown scientific domains where models lack prior knowledge.
• Sample Efficiency: A core focus is maximizing sample efficiency in reinforcement learning, as physical experiments cannot be scaled arbitrarily like digital rollouts.
• Tools Used: The system utilizes Density Functional Theory (DFT) for ground state properties (formation enthalpy) but acknowledges limitations for bandgaps and excited states.
• Force Fields: Machine learning, particularly graph neural networks, has revolutionized the accuracy of force fields used to approximate atomic interactions.
• Data Sources: The team leverages large-scale open datasets (e.g., Meta's ULMA with 100 million DFT calculations) similar to ImageNet, while generating proprietary data through their autonomous labs.
• Reasoning Limitations: The team observes that current LLMs often possess knowledge in weights that does not surface during inference without direct prompting or tool invocation; combining this knowledge intelligently remains a frontier challenge.

Market and Future Outlook

• Motivation: The team is driven by the "no end" nature of science compared to automation fields like accounting, where tasks eventually dry up.
• Barriers: Physical lab infrastructure barriers are expected to decrease as AI and robotic systems improve, though current autonomous labs remain capital-intensive.
• Expansion: While Periodic Labs is not entering the finance sector (due to latency requirements and lack of full context), the methodology for handling noisy, incomplete data is applicable to other verifiable domains.
• Future Challenges: Key open problems for the field include automating characterization analysis of noisy instrument data, formulating testable hypotheses, and handling scenarios where not all experimental context can be recorded.
• Impact: The technology aims to enable lossless energy transmission, fusion energy, quantum computing, and maglev transportation by unlocking new material capabilities.

Context on AI in Science (Lecture Discussion)

• Physics Majors: A significant portion of the discussion acknowledged that while AI has revolutionized coding and force fields, it has not yet replaced human intuition in formulating hypotheses or handling the "messiness" of scientific decision-making under uncertainty.
• Student Advice: Students are encouraged to apply AI to domains where LLMs have not yet reached a "golden age" of automation, such as analog circuit design or specific material engineering challenges.
• Historical Context: The founders compared the current era to the "Silicon Age," suggesting that the next major leap in civilization depends on mastering the engineering of atoms, similar to the transition from the Bronze Age to the Silicon Age.