See Selected Projects

Science & discoveryUnited States

Plan and run chemistry experiments with an LLM agent

Carnegie Mellon's Coscientist used GPT-4 with search, code and lab automation to plan and run real reactions.

Built withChatGPT / OpenAI models

Try it withClaudeChatGPT / OpenAI modelsGeminiGrok

Research (published in Nature), Carnegie Mellon University (Gomes group); Emerald Cloud Lab

Why it matters

Automating experiment design and execution can speed up chemistry for medicines and materials, and makes methods more reproducible.

How it works

Its search covers the internet and documentation. It was shown on six tasks, including optimising palladium-catalysed cross-coupling reactions (Nature, 2023). The code is on GitHub.

Who it's for

Chemists, lab managers and research engineers.

The prompt

Act as a lab-planning assistant for a qualified chemist. Goal: [reaction or measurement goal]. Equipment available: [list]. Safety constraints: [list].

1. Propose 3 candidate procedures, with reagents, amounts, conditions and the literature basis for each (cite only sources you have actually retrieved; otherwise mark [CITATION NEEDED]).
2. For the best one, write a step-by-step protocol and the hazards and controls for every step.
3. Write a dry-run script for [liquid handler or lab API] that only prints the actions it would take.
4. List the measurements to record and how you would decide the next round of conditions.

Do not suggest anything that needs equipment or permits I have not listed. Exclude controlled substances, explosives and chemical-weapon precursors. Check every reagent against its safety data sheet. Do not execute anything.

Open inClaude (opens in a new tab)ChatGPT (opens in a new tab)Gemini (opens in a new tab)Grok (opens in a new tab)

Sources

Both sources are Carnegie Mellon's own work: its peer-reviewed Nature paper and its open-source code. No independent evaluation was found.

  1. 1. Nature: Coscientist; GPT-4; tools; six tasks; cross-coupling optimisation. Autonomous chemical research with large language models (opens in a new tab)

  2. 2. GitHub (Gomes group): Project code repository. gomesgroup/coscientist (opens in a new tab)

Related projects