Bridging Human Intent and Hardware

The core of our autonomous framework relies on Agentic Workflows — utilizing Large Language Models (LLMs) as cognitive agents to manage complex scientific tasks. Rather than relying on rigid, pre-programmed scripts, we deploy agents capable of parsing high-level user goals, reasoning through experimental parameters, and seamlessly orchestrating machine-specific instruments in real-time. By acting as the "brain" of the laboratory, these agents enable dynamic, closed-loop discovery across scanning probe microscopy, electron microscopy, and combinatorial synthesis.

LLM agent controlled STEM

Contributors: Dominick Pelaia, Levi Dunn, Austin Houston, Utkarsh Pratiush, Gerd Duscher, Sergei Kalinin

We built a custom LLM agent called "Angstrom" to autonomously operate the STEM. It is able to write its own code to move the stage, acquire HAADF images and EDS spectra, and perform other functions.

AtomGPT

Contributors: Utkarsh Pratiush, Austin Houston

A demonstration of AtomGPT seamlessly translating raw HAADF images directly into structured atomic data. See how this LLM-driven agent automates the extraction of atomic positions to predict complex supercell structures in real-time.

Literature agent for perovskite materials discovery

Contributors: Jordan Marshall, Mahshid Ahmadi

We built a literature-mining agent to read perovskite solar-cell papers and extract structured information for machine learning. The agent collects device structure, composition, processing, performance, and stability data so it can be used in our PCE and stability prediction workflow.

Small Science Agentic Models

We have developed Small Science Agentic Models. The core idea is that rather than building one large, generalist AI agent that tries to do everything — or relying solely on a single optimization algorithm like Bayesian optimization — we distribute intelligence across a set of narrow, physically grounded agents, each responsible for one well-defined stage of the experimental workflow:

  • 1. Hypothesis Agent: The scientific reasoning layer.
  • 2. Protocol Agent: The translation layer.
  • 3. Characterization Agent: The state-generation layer. Critically, this agent does not treat measurements as endpoints.
  • 4. ML/Decision Agent: The optimization layer.
  • 5. Orchestrator Agent: The connective tissue.

The Hypothesis Agent

Contributors: Sheryl Sanchez, Ariel Thompson

A demonstration of the scientific reasoning layer, which leverages Socratic reasoning to clarify research questions, surface hidden assumptions, and generate testable hypotheses for experimental design.

The Protocol Agent

Contributors: Sheryl Sanchez, Ariel Thompson

A showcase of the translation layer in action, autonomously interpreting high-level hypotheses to formulate executable laboratory workflows and precise automation protocols for robotic liquid handling.

Automated Hardware Orchestration

Contributors: Sheryl Sanchez, Ariel Thompson

The physical execution layer: automated robotic hardware orchestrated by the agentic framework, seamlessly carrying out high-throughput fabrication and characterization of combinatorial material libraries.