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
- Gao et al., Cell Perspective 2024 (single-LLM-multi-role case study)