MARS

Hierarchical multi-agent and robotic framework that pairs 19 LLM agents and 16 domain tools with a robotic synthesis platform to autonomously design, synthesize, and optimize perovskite materials.

   
Affiliation Materials Artificial Intelligence Center, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (Yu Xue-Feng group), with collaborators at CityU Hong Kong, NUS, and Chongqing University of Posts and Telecommunications
First introduced 2026-02 (Matter 9, 102577)
Lifecycle stages Multi-stage (literature review + materials design + synthesis planning + robotic execution + analysis)
Autonomy level Fully autonomous within the perovskite synthesis loop; supports natural-language interaction and expert intervention via the orchestrator
Domain focus Materials science / chemistry — perovskite nanocrystal synthesis and water-stable perovskite composites
Availability Open source — tangger2000/MARS

Approach

Five functional groups mirror a human research team:

  • Orchestrator (G1) — central coordinator; receives user requests, plans sub-tasks, dispatches work to administrator agents inside each group, and aggregates results.
  • Scientist (G2) — handles literature review and design proposals using a hybrid retrieval-augmented generation strategy (GraphRAG over a customized knowledge base) intended to reduce LLM hallucination.
  • Engineer (G3) — translates scientific proposals into executable protocols; a structural engineer emits standardized JSON schemas, a software engineer generates code, and a code reviewer checks for safety.
  • Executor (G4) — drives a robotic workstation and mobile-robot controller to perform synthesis and characterization.
  • Analyst (G5) — analyzes experimental results, generates refinement strategies, and feeds them back to the Orchestrator for the next iteration.

The system is built on Microsoft AutoGen with Selector Group Chat and Swarm coordination, runs GPT-4o as the foundation model, and isolates code execution in Docker containers.

Validation

End-to-end experimental demonstration in perovskite nanocrystal chemistry. MARS optimized a synthesis protocol within 10 iterations and designed a biomimetic “core-shell-corona” structure for a water-stable perovskite composite in 3.5 hours of system time, compared with the 4–6 months the authors report for the conventional manual route — a claimed ~60× acceleration.

Notable results

  • Closed-loop perovskite synthesis optimization in 10 robotic iterations.
  • Cross-domain knowledge integration produced a biomimetic perovskite composite design inspired by cell-membrane structure.
  • Reports reproducibility and safety affordances via open-source release and sandboxed code execution.

Primary paper

Shi et al., “Knowledge-driven autonomous materials research via collaborative multi-agent and robotic system,” Matter 9, 102577 (2026).

Other references

None yet.

Code

tangger2000/MARS — core MARS implementation; built on microsoft/autogen.