Skip to content

paper.md

Latest commit

 

History

History
77 lines (55 loc) · 5.5 KB

paper.md

File metadata and controls

77 lines (55 loc) · 5.5 KB
Raw
title LabAgent: An Open-Source AI-Guided Framework for Automated Scientific Measurements
tags
Python
laboratory automation
artificial intelligence
instrument control
measurement planning
authors
name orcid affiliation
Anai Guo
0000-0000-0000-0000
1
affiliations
name index
Independent Researcher
1
date 13 April 2026
bibliography paper.bib

Summary

LabAgent is an open-source Python framework that connects large language models (LLMs) to laboratory instruments for automated scientific measurements. The framework provides AI-guided measurement planning, instrument discovery and classification, safety boundary validation, data analysis with domain-specific interpretation, and experiment memory — covering 46 measurement types across 9 scientific disciplines and 68 instrument models across 26 manufacturers.

Statement of Need

High-precision laboratory measurements in physics, chemistry, biology, and materials science require complex instrument orchestration, parameter optimization, and data analysis. Researchers currently face a choice between expensive proprietary software (e.g., LabVIEW) or writing custom measurement scripts from scratch for each experiment. Both approaches are time-consuming, error-prone, and lack safety guarantees.

While instrument control libraries such as PyVISA [@pyvisa] and PyMeasure [@pymeasure] provide foundational drivers, and measurement suites like PICA [@pica] offer hardcoded GUIs for specific instruments, no existing tool combines AI reasoning with instrument control in a unified, extensible framework.

LabAgent addresses this gap by placing AI at every step of the measurement workflow: from literature-informed protocol discovery to adaptive measurement planning with three-tier safety validation, to AI-generated data analysis with domain-specific scientific interpretation.

Architecture

The framework follows an agent harness architecture inspired by OpenHarness [@openharness], with four core subsystems:

  1. Tool Registry: Ten standardized tools (scan, classify, propose, validate, literature search, analyze, memory recall, skill generation, manual lookup, healthcheck), each following a BaseTool abstract interface with Pydantic input validation.

  2. Agent Engine: An asynchronous query loop that processes user messages, invokes LLMs via litellm [@litellm] for model-agnostic routing, detects tool calls, executes them with permission checks, and streams results back to the user.

  3. Permission System: A three-layer safety model with immutable absolute limits (e.g., maximum current 10 A), boundary checks loaded from YAML policies, and interactive user confirmation for potentially dangerous operations.

  4. Multiple Interfaces: A Model Context Protocol (MCP) server for AI IDE integration, a command-line interface with 15 subcommands, a FastAPI web GUI with adaptive measurement forms, and a Textual-based terminal panel for interactive sessions.

Measurement Coverage

The framework includes 46 YAML-based measurement templates — 45 across nine disciplines plus a general-purpose custom sweep:

  • Electrical characterization (7): IV curves, R-T, delta mode, high resistance, FET transfer/output, breakdown voltage
  • Electrochemistry (4): Cyclic voltammetry, EIS, chronoamperometry, potentiometry
  • Semiconductor and optoelectronics (5): Solar cell IV, DLTS, photocurrent spectroscopy, photoresponse transients, tunneling spectroscopy
  • Sensors, materials, and environmental (7): Gas sensor, humidity response, impedance biosensor, cell counting, pH calibration, strain gauge, fatigue
  • Dielectric / ferroelectric (4): C-V, P-E loop, pyroelectric current, capacitance-frequency
  • Thermoelectric (2): Seebeck coefficient, thermal conductivity
  • Superconductivity (2): Tc transition, critical current density
  • Magnetic and transport (condensed-matter specialty) (8): Hall effect, magnetoresistance, anomalous Hall effect, SOT loop shift, FMR, hysteresis loops, magnetostriction, Nernst effect
  • Quantum Design systems (6): PPMS R-T / MR / Hall / heat capacity, MPMS M-H / M-T with MultiPyVu integration

Optical power meters, UV-Vis spectrometers, plate readers, balances, pH meters, and flow/pressure controllers are recognized by the instrument classifier (and several have reference command procedures), but do not yet have dedicated measurement templates.

AI Capabilities

The framework provides eight distinct AI capabilities:

  1. Instrument Classification: Dictionary-based lookup with LLM fallback for unknown instruments
  2. Parameter Optimization: LLM suggests optimal sweep ranges based on sample description and literature
  3. Safety Advisory: LLM explains why limits exist and suggests safer alternatives
  4. Script Generation: LLM creates custom analysis scripts for arbitrary measurement types
  5. Result Interpretation: LLM provides domain-specific scientific insights on extracted values (physics, electrochemistry, spectroscopy, etc.)
  6. Skill Generation: LLM creates new measurement protocols from existing examples
  7. Agent Chat: Multi-turn conversation with autonomous tool calling
  8. Experiment Memory: SQLite with FTS5 full-text search, frozen snapshots injected into agent context

Acknowledgements

Measurement procedures reference patterns from the PICA project [@pica]. Agent architecture is inspired by OpenHarness [@openharness] and Hermes Agent [@hermes]. The framework uses litellm [@litellm] for model routing and PyVISA [@pyvisa] for instrument communication.

References