🤖 AI

PeroMAS: A Multi-agent System of Perovskite Material Discovery

This paper introduces PeroMAS, a multi-agent system that integrates perovskite-specific tools via Model Context Protocols to enable end-to-end, closed-loop material discovery from literature retrieval to experimental synthesis, demonstrating superior efficiency and effectiveness in identifying candidate materials compared to existing single-model approaches.

Original authors: Yishu Wang, Wei Liu, Yifan Li, Shengxiang Xu, Xujie Yuan, Ran Li, Yuyu Luo, Jia Zhu, Shimin Di, Min-Ling Zhang, Guixiang Li

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yishu Wang, Wei Liu, Yifan Li, Shengxiang Xu, Xujie Yuan, Ran Li, Yuyu Luo, Jia Zhu, Shimin Di, Min-Ling Zhang, Guixiang Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to invent a new, super-efficient recipe for a cake that is also healthy, cheap to make, and doesn't spoil in the fridge. In the world of science, this "cake" is a Perovskite Solar Cell—a type of solar panel that could revolutionize how we capture energy from the sun.

For decades, scientists have been trying to perfect this recipe. But the problem is that there are billions of possible ingredient combinations (chemicals), and testing them one by one in a lab is like trying to find a needle in a haystack by looking at one straw at a time. It's slow, expensive, and frustrating.

Enter PeroMAS: The "Super-Chef" Team

This paper introduces PeroMAS, which isn't just one smart computer program, but a team of specialized AI robots working together to invent this solar cell recipe. Think of it not as a single genius, but as a high-tech kitchen brigade where every member has a specific job.

Here is how the team works, using a simple analogy:

1. The Team Roles (The Agents)

Instead of one robot trying to do everything (which often leads to mistakes), PeroMAS uses four distinct "agents," each with their own superpower:

The Librarian (Miner Agent):
- Job: This agent is the researcher. It dives into millions of scientific papers and databases to find out what other scientists have tried before.
- Analogy: Imagine a librarian who can read every book in the world in seconds and pull out the exact pages about "how to bake a cake that lasts 100 years." It gathers all the existing knowledge so the team doesn't waste time reinventing the wheel.
The Architect (Designer Agent):
- Job: This agent takes the knowledge from the Librarian and starts mixing ingredients. It creates new, hypothetical recipes (chemical formulas) that have never been tried before.
- Analogy: This is the creative chef who says, "Okay, the Librarian told us that adding a pinch of cinnamon helps, but we need to reduce the sugar. Let's try mixing these specific chemicals in this exact order." It designs the blueprint for the new solar cell.
The Simulator (Emulator Agent):
- Job: Before anyone touches a real beaker, this agent runs a "virtual test" on the computer. It predicts: "If we bake this recipe, will it work? Will it generate electricity? Will it break down?"
- Analogy: This is like a flight simulator for a new airplane. The Architect designs the plane, but the Simulator flies it a thousand times in the computer to see if it crashes, without ever building a real one. It filters out the bad ideas instantly.
The Inspector (Analyst Agent):
- Job: This agent looks at the results of the simulation. If the recipe failed, it asks "Why?" and explains the problem. It tells the team, "The cake collapsed because the oven was too hot," or "The solar cell is toxic because of this one ingredient."
- Analogy: This is the quality control manager. If the flight simulator shows a crash, the Inspector points to the exact broken part and tells the Architect how to fix the design for the next try.

2. The "Brain" (The Meta Agent)

Running the show is the Meta Agent.

Job: It's the project manager. It listens to the user's request (e.g., "Make a solar cell that is 20% efficient, lasts 1,000 hours, and uses less lead"), and then directs the other four agents.
Analogy: Imagine a conductor in an orchestra. The Librarian, Architect, Simulator, and Inspector are the musicians. The Conductor (Meta Agent) makes sure they all play the right notes at the right time, ensuring the whole team moves toward the goal together.

3. The "Model Context Protocol" (MCP)

You might wonder, "How do these robots actually do things?"

Analogy: Think of MCP as a universal remote control. Instead of the robots having to learn a new language for every different computer program or database, MCP gives them a standard set of buttons to press. It lets them instantly grab a tool, run a calculation, or search a database without getting confused.

4. The Real-World Test (The "Wet Lab")

The most exciting part of this paper is that they didn't just stop at the computer.

They took the best recipe generated by their AI team and sent it to a real human laboratory.
The humans mixed the chemicals and built the solar cell exactly as the AI instructed.
The Result: The AI-designed solar cell actually worked! It achieved 17% efficiency (very close to the AI's prediction) and successfully reduced the amount of toxic lead by 50%.

Why Does This Matter?

Before PeroMAS, discovering new materials was like trying to solve a puzzle blindfolded. You had to guess, test, fail, and guess again.

PeroMAS is like putting on glasses with night vision. It:

Reads all the history (Librarian).
Dreams up new ideas (Architect).
Tests them instantly in a virtual world (Simulator).
Fixes the mistakes (Inspector).
Builds the real thing (Human Lab).

This system proves that AI can do more than just chat; it can actually discover new scientific materials that are safer, more efficient, and ready to help solve the world's energy crisis. It's the difference between a single person trying to build a house alone versus a fully automated construction crew working in perfect sync.

1. Problem Statement

The discovery of Perovskite Solar Cells (PSCs) is a complex, multi-step scientific workflow involving knowledge preparation, material design, dry-lab simulation, and wet-lab synthesis. Current AI approaches face three critical limitations:

Fragmentation: Existing Machine Learning (ML) models focus on isolated tasks (e.g., property prediction) and fail to propagate physical constraints across the entire workflow.
Constraint Violation: Large Language Models (LLMs) possess domain knowledge but often generate content that violates physical or chemical constraints (e.g., impossible stoichiometry) and lack the ability to invoke specialized external tools.
Single-Objective Bias: Most methods optimize for a single metric (e.g., Power Conversion Efficiency or PCE), ignoring critical trade-offs such as stability, toxicity, and reproducibility.

The core challenge is to create an autonomous system capable of end-to-end, multi-objective optimization that bridges the gap between digital design and physical realization.

2. Methodology: The PeroMAS Framework

The authors propose PeroMAS, a hierarchical multi-agent system designed to orchestrate the full PSC discovery lifecycle. The system utilizes a Model Context Protocol (MCP) to encapsulate domain-specific tools, allowing agents to interact with external data and simulation engines seamlessly.

System Architecture

The system consists of a Meta Agent and four specialized Functional Agents:

Meta Agent (Orchestrator):
- Acts as the central planner using a Plan-Do-Check-Act (PDCA) workflow.
- Translates user-defined multi-objective constraints (e.g., PCE > 20%, Low Toxicity) into executable task instructions.
- Manages global memory, performs reflection on previous failures, and dynamically prunes redundant sub-tasks to optimize efficiency.
Functional Agents (via MCP):
- Miner Agent: Addresses the "cold-start" problem by retrieving literature and extracting structured data (composition, process, performance) from unstructured texts to build a knowledge base.
- Designer Agent: Performs inverse design within the combinatorial space. It generates candidate compositions, structures, and synthesis parameters, filtering them against hard physical constraints (e.g., synthesizability).
- Emulator Agent: Acts as a high-throughput virtual screening funnel. It invokes Graph Neural Networks (GNNs) and ML regression models to predict intrinsic properties (bandgap) and device metrics (PCE, stability) with uncertainty estimates.
- Analyst Agent: Validates results, performs attribution analysis (using SHAP) to identify key factors affecting performance, and diagnoses discrepancies to guide the next iteration.

Workflow

The system operates in a closed loop:

Knowledge Preparation: Miner extracts data priors.
Design: Designer generates hypotheses based on constraints.
Dry-Lab: Emulator simulates and screens candidates.
Analysis: Analyst critiques results and refines the hypothesis.
Iteration: The Meta Agent updates the strategy based on feedback until targets are met.

3. Key Contributions

First Closed-Loop Perovskite System: PeroMAS is the first specialized multi-agent system to integrate knowledge preparation, design, and dry-lab simulation into a single, iterative workflow, filling a significant research gap.
MCP-Based Tool Orchestration: By standardizing the integration of heterogeneous tools (e.g., arXiv, Materials Project, DFT calculators) via MCP, the system ensures that agents can dynamically retrieve and utilize the most appropriate tools for specific sub-tasks.
Wet-Lab Validation: The system's designs were physically synthesized and tested, demonstrating high consistency between "dry-lab" predictions and real-world performance.
PeroMAS-Bench: The authors constructed the first hierarchical evaluation benchmark for perovskite discovery, comprising 150 high-fidelity tasks (atomic-level and full-process) annotated by human experts to assess agent planning and coordination.

4. Results

The system was evaluated against baselines including Standard LLMs, Single ReAct Agents, and Generalist Autonomous Agents (e.g., Manus).

Performance Metrics:
- Task Completion: PeroMAS achieved a 72.4% task completion rate (using Claude-4.5), significantly outperforming baselines.
- Tool Accuracy: The system achieved 89.5% tool invocation accuracy with the open-source DeepSeek-V3 backbone, proving open-source models can match proprietary ones in complex execution when properly orchestrated.
- Evaluation Consistency: The system's automated scoring showed high alignment (±0.15 points) with human expert judgments, validating the reliability of the evaluation framework.
Wet-Lab Verification (The "Smoking Gun"):
- Objective: Design a low-lead perovskite with PCE $\ge$ 20% and stability $T_{80} > 1000$ h.
- Outcome: PeroMAS designed a Sn-Pb mixed perovskite recipe that reduced lead content by 50%.
- Experimental Result: The synthesized device achieved a measured PCE of 17%, falling strictly within the system's predicted range of 16%–19%. This confirmed the system's ability to generate physically viable, high-performance materials.

5. Significance

Paradigm Shift: PeroMAS moves AI in materials science from "single-task prediction" to "holistic, autonomous discovery." It demonstrates that multi-agent collaboration can effectively resolve the fragmentation of traditional scientific pipelines.
Practical Viability: The successful wet-lab validation proves that AI agents can not only simulate but also guide real-world chemical synthesis, paving the way for self-driving laboratories.
Open Science: The release of PeroMAS-Bench and the MCP toolchain provides a standardized foundation for the community to further develop and evaluate agentic AI in scientific domains.
Scalability: The framework is not limited to perovskites; the hierarchical architecture and MCP integration offer a blueprint for applying agentic AI to other complex scientific fields requiring multi-objective optimization and physical validation.