PeroMAS: A Multi-agent System of Perovskite Material Discovery

arXiv:2602.13312 · cs.MA, cs.AI · Submitted 2026-02-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "PeroMAS: A Multi-agent System of Perovskite Material Discovery".

Jane: The paper was written by Yishu Wang, Wei Liu, Yifan Li, Shengxiang Xu, Xujie Yuan et al. from Southeast University (Nanjing, China) and Sun Yat-Sen University (Zhuhai, China) and Hong Kong University of Science and Technology (Guangzhou, Guangzhou, China) and Zhejiang Normal University (Jinhua, China).

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: So, we’ve established that PeroMAS: A Multi-agent System of Perovskite Material Discovery is built on the concept of collaboration. But let's dig a little deeper into what that means for the authors and the field. The title itself suggests something quite powerful about how these agents are working together to solve problems.

Jane: It’s a big leap from simple machine learning, Tom. We’re not just asking an AI to predict if a material is good; we are creating an entire autonomous discovery pipeline that moves past the concept of discrete models.

Lu: The fact that they are using a multi-agent system means they believe the solution is in the coordination, allowing different specialized "brains" to handle their specific parts of the workflow.

Meng: I'm interested in how this architecture allows us to scale up; does it require a lot of resources or can we deploy these multi-agent systems across various computational platforms?

Lalam: It’s about building an intellectual structure, Meng, where the ideas flow between different roles like a team meeting, making PeroMAS: A Multi-agent System of Perovskite Material Discovery much more intuitive for humans to understand.

Tom: That makes sense. We're moving away from a single monolithic AI and towards this division of labor.

Jane: It’s about creating specialized agents that handle the specific nuances of the material science, which is exactly what we need in complex fields like photovoltaics.

Lu: The potential for finding new combinations that haven't been thought of before is huge when you can coordinate multiple specialized AI tools like this.

Meng: I just hope that this structure doesn's efficiency matches the complexity of the real-world optimization problems we have to solve in a lab.

Lalam: We need to make sure this system allows for better decision-making in a way that improves our scientific culture by encouraging structured, repeatable inquiry.

Summary: Tom: Now, looking at the summary of PeroMAS: A Multi-agent System of Perovskite Material Discovery, we see how they tackle the entire process from start to finish. It seems like they've mapped out every single step a scientist takes.

Jane: That’s right, Tom. They’ve taken that messy, multi-step scientific workflow—from finding papers to testing chemicals—and organized it into this structured framework using Model Context Protocols.

Lu: The Model Context Protocols are the glue here; they allow the different agents to pass information seamlessly, which is crucial for maintaining context across a large-scale project.

Meng: I’m curious about how this translates into actual data management; how do we ensure that all the disparate data points from literature and lab tests stay cohesive within these structured protocols?

Lalam: It feels like they’ve solved the problem of fragmentation, allowing us to see a complete picture of material discovery rather than just disconnected tasks.

Tom: Fragmentation is the word, Jane. The current methods were doing small parts well but failing to link them together into one coherent whole for PeroMAS: A Multi-agent System of Perovskite Material Discovery.

Jane: It’s about having a system that can actually handle both the theoretical knowledge gathering and the physical process, not just separating them into two different tools.

Lu: The ability to cover everything from literature retrieval right through property prediction is a huge win for algorithmic design in this domain.

Meng: I worry that if it's too complex, it could be hard to maintain or debug, so ensuring the modularity of the MCP is key for practical scaling.

Lalam: This framework should allow us to reduce cognitive load on human researchers by providing them with a comprehensive overview of all possible paths.

Improvements: Tom: The paper highlights several ways PeroMAS: A Multi-agent System of Perovskite Material Discovery improves upon existing approaches. We are moving past just single-objective optimization, which is a massive step forward.

Jane: It’s the multi-objective constraints, Tom, that are so important. Instead of just asking for the most efficient cell, we're asking for something that balances efficiency against stability and even toxicity simultaneously.

Lu: The integration of mechanistic analysis means PeroMAS isn't just generating random ideas; it is generating scientifically grounded hypotheses based on existing knowledge.

Meng: From an engineering standpoint, I’m impressed by the adaptive pruning mechanism, which suggests we are minimizing unnecessary computation by skipping redundant tasks in the design process.

Lalam: This improvement allows us to have a more holistic view of what's possible, making our approach to science much more comprehensive and less likely to overlook critical trade-offs.

Tom: Trade-offs are where things get interesting, Jane. The old models often ignored these conflicts, but PeroMAS: A Multi-agent System of Perovskite Material Discovery is designed to solve them head on.

Jane: It’s about having a system that can handle the complexity of multiple targets and achieving the desired performance across different physical constraints simultaneously.

Lu: It's like ensuring that every single step, from initial design to final analysis, contributes positively to a holistic scientific outcome.

Meng: I just hope this ability leads to faster iteration cycles in real-world labs without sacrificing quality control or rigor.

Lalam: We can improve our entire research culture by adopting systems that understand these multi-faceted constraints, moving beyond single-focus experiments.

Conclusion: Tom: So, we've covered the structure, the workflow, and how PeroMAS: A Multi-agent System of Perovskite Material Discovery improves upon previous methods. We've seen that it’s robust and handles complex scenarios.

Jane: It seems like a powerful combination of theory and practice, Tom. The fact that it actually worked in the wet-lab validates everything we've discussed about its promise a real-world results are truly encouraging.

Lu: I think the release of the PeroMAS Benchmark is a huge contribution, providing a standardized tool for future AI scientists to measure their own system capabilities.

Meng: The experimental results, especially achieving seventeen percent efficiency while reducing lead content by fifty percent, show that practical impact is exactly what this architecture can deliver.

Lalam: I just hope this will encourage more people to use multi-agent systems in other fields, not just that we're solving the perovskite problem.

Tom: It’s a proof of concept for a much bigger idea, Lalam. We have to recognize the power of PeroMAS: A Multi-agent System of Perovskite Material Discovery.

Jane: It’s exciting to wrap up this discussion and share that everything from the initial literature search to is validated by real-world science.

Lu: I am excited about what comes next, seeing how this opens up even more ambitious research goals for the community.

Meng: I'm looking forward to seeing how these designs translate into a full, autonomous self-driving lab operation.

Lalam: We can be confident that this work will contribute significantly to a smarter, more efficient scientific future.

Southeast University (Nanjing, China) · Sun Yat-Sen University (Zhuhai, China) · Hong Kong University of Science and Technology (Guangzhou, Guangzhou, China) · Zhejiang Normal University (Jinhua, China)

cs.MA, cs.AI

Submitted: 2026-02-10

Updated: 2026-09-03

Code: https://github.com/rdkit/rdkit

Importance score: 91/100

The gist: PeroMAS is presented as a sophisticated Multi-agent System designed for advanced perovskite material discovery, addressing complex scientific challenges such as balancing conflicting objectives like

Key concepts

Multi-agent System
This architecture uses multiple specialized AI agents that collaborate, rather than relying on a single monolithic model. The system's power comes from the coordination of these different agents, allowing them to handle specific parts of the complex material discovery workflow.
Model Context Protocols (MCP)
These protocols function as the structural glue within PeroMAS. They allow different specialized agents to pass information seamlessly, which is crucial for maintaining context across a large-scale project and linking disparate data from literature and lab tests.
Multi-objective Constraints
This improvement allows the system to handle multiple goals simultaneously, moving past single-focus optimization. Instead of optimizing only for efficiency, it can balance conflicting targets like efficiency against stability and toxicity.

Terminology

Summary

PeroMAS is presented as a sophisticated Multi-agent System designed for advanced perovskite material discovery, addressing complex scientific challenges such as balancing conflicting objectives like "High Efficiency (PCE > 20%), High Stability (T80 > 1000h), and Reduced Toxicity." This framework models the entire research lifecycle, from initial literature retrieval to final physical device fabrication, providing a comprehensive platform for accelerating materials science breakthroughs.

Architecture and Protocol Standardization

The system's operational backbone is the Model Context Protocol (MCP). This protocol was developed to standardize the interaction between agents and external knowledge, ensuring that specialized algorithms and data sources can be reliably integrated into the workflow. The full-chain MCP tool training and runtime environment is deployed on a high-performance computing cluster equipped with 4× NVIDIA H100 GPUs, indicating its computational demands for complex simulations.

Agent Roles and Multi-Step Workflow

PeroMAS utilizes a sequence of specialized agents to guide the discovery process, as demonstrated in the dry-lab case of designing Sn-Pb Mixed Perovskite solar cells. The workflow proceeds through distinct roles:

  • Miner Agent (Retrieval): Initiates the process by providing foundational scientific strategies, such as proposing Sn-Pb alloying (1:1) to 1.23 eV bandgap.

  • Designer Agent (Recipe Design): Translates the retrieved knowledge into actionable experimental plans, generating a detailed Bill of Materials (BOM) and Standard Operating Procedure (SOP). For example, it specifies the target composition as (FASnI3)0.6 (MAPbI3)0.4 + 10% SnF2 + 3% EDAI2.

  • Emulator Agent (Performance Prediction): Predicts key device metrics based on the proposed design, yielding values for PCE, Voc, Jsc, FF, and T80.

  • Analyst Agent (Critical Trade-off Analysis): Provides a critical evaluation of the formulation's viability. It highlights trade-offs—such as noting that Lower toxicity comes at the cost of oxidation stability—and identifies necessary mitigation strategies.

  • Meta Agent (Final Conclusion Report): Synthesizes all preceding outputs into a final, executive-level conclusion, culminating in a formal recommendation such as PROCEED TO EXPERIMENT.

Scope of Scientific Tasks and Data Integration

The system is designed to support multiple functional modules essential for materials research. The task assignment scope includes:

  1. Synthesis Feasibility Assessment: Determining if a compound is synthesizable, requiring both positive (synthesizable) and negative (non-synthesizable) samples for training.

  2. Synthesis Route Planning: Providing Method Recommendation and Precursor Identification, utilizing large datasets of 4,500/500 samples.

  3. Crystal Structure Generation: Supporting both a Core Set (Unique Compositions) and an Expanded Set (Data Augmentation).

  4. Property Prediction: Handling multi-target regression tasks, which involve predicting various physical attributes critical for device performance.

Physical Implementation and Validation Details

The system’s outputs are validated through rigorous wet-lab procedures conducted in a standardized ISO Class 7 cleanroom environment. The fabrication process is highly detailed, including:

  • Substrate Preparation: ITO glass substrates undergo sequential cleaning followed by UV-Ozone treatment to improve wettability. A Nickel Oxide (NiO x) nanoparticle dispersion is then spin-coated to form the hole transport layer.

  • Perovskite Deposition: The precursor solution, prepared according to the Designer Agent's stoichiometry, is spin-coated using a one-step method within a nitrogen-filled glovebox (O 2 < 0.1 ppm, H 2 O < 0.1 ppm).

  • ETL & Electrode Evaporation: The final layers, including C 60 and Bathocuproine (BCP), followed by a Silver (Ag) electrode, are deposited sequentially via thermal evaporation under a high vacuum of 5 times 10-4 Pa.

Improvements for AI systems

The provided framework, PeroMAS, is highly advanced but operates within a specialized domain (perovskites) and relies on discrete functional modules. To elevate this into a universally powerful AI platform capable of handling broader scientific challenges—and to mitigate inherent risks like data bias or generalization failure—I propose the following improvements:

The current system provides correlation (e.g., Low-Lead correlates with lower stability). We must move beyond correlation to establish causality.

  • Improvement: Implement a specialized module utilizing Do-Calculus and Structural Causal Models (SCMs). This module must be trained on diverse experimental datasets (not just the curated PeroMAS data) and incorporate known physical laws (e.g., thermodynamic constraints, band alignment rules) as soft constraints into the inference engine.

  • Technical Detail: The Analyst Agent should not merely report a trade-off; it must provide a Causal Pathway Graph showing A to B (e.g., Replacing Pb with Sn Bandgap shift Changes in defect tolerance).

The current MCP tools are static data sources (e.g., Materials Project API). The knowledge base needs to be dynamic, integrating real-time, unstructured scientific advances.

  • Improvement: Upgrade the Sci-mcp tool into a Multi-Modal Graph Neural Network (GNN) that ingests not only text and citation data but also structural images (micrographs), reaction schematics, and spectroscopic signatures (e.g., XRD patterns).

  • Technical Detail: The GNN must continuously update the graph's edges with newly discovered relationships (Material A Property Y) derived from the literature, allowing the system to identify analogous problems or solutions across entirely different material classes (e.g., applying a stable interface layer technique used in superconductors to perovskites).

The final conclusion report currently uses a single Confidence Level (95%). In high-stakes scientific research, failure modes must be quantified rigorously.

  • Improvement: Integrate Bayesian Deep Learning techniques across all prediction modules (Emulator, Designer). Instead of outputting a single value (e.g., PCE = 18%), the system must output a full probability distribution function (PDF) for each predicted metric, accompanied by explicit variance estimates.

  • Technical Detail: The Meta Agent must then perform Risk Propagation Analysis, calculating how the uncertainty in one parameter (e.g., plus or minus 3% in precursor purity) propagates through the entire system to impact the final key performance indicator (KPI), like V oc. This forces a proactive identification of experimental failure points.

The datasets are highly task-specific (e.g., PCE prediction for perovskites). The system must be able to transfer knowledge efficiently between chemically distinct material families (e.g., from metal oxides to chalcogenides).

  • Improvement: Implement a Meta-Learning framework that trains the core models (like the GNN and XGBoost) on a vast, diverse dataset of general chemical bonding principles and crystal lattice motifs. The system should learn how to learn materials properties.

  • Technical Detail: When presented with a new material class (e.g., phosphorene-based thermoelectrics), the system should first use this general knowledge to identify the closest structural and electronic analogues within its training manifold, drastically reducing the need for massive, domain-specific data collection for every new target material.

The improved PeroMAS platform will transition from a powerful Discovery Tool to a robust Predictive Research Engine. It can achieve the following:

  1. Hypothesis Generation with Proven Causality: Instead of simply suggesting Try X because it has similar properties to Y, the system will state: "We hypothesize that adopting Structure S new will increase stability because S new causally alters the defect energy level, thereby reducing the rate of oxidative degradation, as demonstrated by analogous systems in Material Class Z."

  2. Automated Failure Mode Analysis: Given a target device architecture and set of precursors, it will not only predict the optimal recipe but will also generate a comprehensive Failure Mode and Effects Analysis (FMEA) report, detailing the specific experimental steps most likely to fail (e.g., "Annealing too quickly (Rate > 5 C/min) has an 80% probability of creating grain boundaries that reduce J sc below the minimum threshold").

  3. Accelerated Material Portfolio Design: The system can autonomously design and validate an entire portfolio of materials (e.g., ten lead-free perovskite variants) in a single run, providing not only the optimal recipe for each but also a statistical ranking of their predicted performance relative to the required cost/complexity (integrating economic models into the decision matrix).

  4. Bridging Disciplinary Gaps: It can synthesize knowledge from disparate fields—for example, linking advanced computational fluid dynamics (CFD) simulations used in battery cooling systems to optimize the thermal management requirements of a high-power density solar cell, something impossible with current siloed functional agents.

Sources

Related papers