Application of multi-agent in energy materials research

As energy materials research—a critical field covering batteries, photovoltaics, thermoelectrics, piezoelectrics, and superconductors—confronts an efficiency bottleneck from escalating data and complexity, the traditional research paradigm reliant on manual experience and theoretical deduction is proving insufficient. This necessitates a strategic evolution toward intelligence-driven methodologies. While single agents empowered by large language models (LLMs) offer preliminary support, they exhibit critical performance limitations. A single agent struggles to manage multi-task workflows, integrate disparate multi-source data, and maintain logical coherence through complex, long-chain reasoning, ultimately failing to meet the demands of rigorous scientific discovery. To overcome these structural limitations, multi-agent systems (MAS) have emerged as a pivotal solution. MAS leverages a distributed architecture that enables both vertical role specialization—assigning distinct agents to tasks such as planning, execution, and review—and horizontal parallel collaboration for tasks like debating and consensus-building. This structure directly addresses single-model bottlenecks by facilitating a “hypothesis-practice-analysis” research closed-loop. This approach offers superior flexibility, fault tolerance, and scalability. Crucially, it significantly enhances the credibility and reproducibility of outcomes by generating auditable, traceable decision-making pathways, a non-negotiable requirement for high-stakes scientific research. This review systematically analyzes the application landscape and introduces a novel three-level classification framework for MAS in energy materials, based on two key criteria: the integrity of the scientific research closed-loop and the degree of human intervention. Level 1 (tool-level) systems function as sophisticated auxiliary tools, handling discrete sub-tasks like data retrieval and analysis, and remain fully dependent on human instructions and evaluation. Level 2 (assistant-level) systems represent a collaborative partner, connecting multiple research stages autonomously but requiring human feedback at critical decision nodes, thus operating in a “human-in-the-loop” model where the human guides and validates. Level 3 (system-level) represents a significant leap, capable of autonomously executing the complete research loop from hypothesis generation to experimental execution and iterative self-optimization. In this “human-out-of-the-loop” paradigm, the human role evolves from a hands-on operator or collaborator to a high-level strategic auditor who reviews principles and outcomes. Despite this progress, the field faces significant and persistent challenges. A primary obstacle is the scarcity of high-quality, reproducible experimental data, particularly the negative samples and failure cases, which inhibit generalization and exacerbate the “sim-to-real” gap. Furthermore, systemic issues of reliability and interpretability remain. This creates the risk of systemic hallucinations, where multiple agents might reach a consensus on a factually incorrect conclusion. This is further compounded by the lack of unified standards for data formats, interfaces, and application programming interfaces (APIs) in materials science, which constrains MAS scalability when integrating diverse databases, simulation tools, and experimental platforms. Looking ahead, MAS is poised to drive a fundamental paradigm shift. Future systems will move toward flexible, end-to-end autonomous iteration, creating a continuous cycle of discovery and refinement. This will facilitate true multi-scale information integration, finally connecting atomic-level simulations with macro-scale device performance in a predictive, holistic model. Perhaps most profoundly, this evolution will enable interdisciplinary cognitive fusion, allowing the system to bridge concepts from disparate scientific domains. This evolution will transition the research model from “human-machine collaboration”, where the machine is a tool, toward “system co-creation”, where the autonomous system acts as a second cognitive entity. By building robust, verifiable, and safe autonomous research platforms, MAS can provide critical technological support to accelerate innovation and help achieve global sustainability and carbon-neutrality goals.