Coscientist (CMU)

GPT-4 planner that orchestrates web search, documentation retrieval, Python code execution, and Symbolic Lab Language commands sent to a physical robotic chemistry platform. Fully autonomous on the experimental loop.

   
Affiliation Carnegie Mellon University, Gomes Lab (Boiko et al.)
First introduced 2023-12 (Nature)
Lifecycle stages Experiment design, Analysis
Autonomy level Fully autonomous (within a closed-loop chemistry workstation)
Domain focus Chemistry (catalysis, palladium-catalyzed cross-couplings)
Availability Code on request (per Boiko et al. supplementary)

Approach

GPT-4 acts as a planner that orchestrates web search, documentation retrieval, Python code execution, and Symbolic Lab Language (SLL) commands sent to a physical robotic platform. The agent autonomously plans, generates SLL, transfers code to the device, and executes experiments.

Validation

Wet-lab execution of optimized palladium-catalyzed Suzuki and Sonogashira cross-couplings on a physical platform.

Notable results

Designed and executed cross-coupling reaction optimizations end-to-end with no human intervention on the experimental loop.

Primary paper

Boiko et al., “Autonomous chemical research with large language models,” Nature 624, 570–578 (2023).

Other references

Code

Repository