Researchers from the Hong Kong Polytechnic University (PolyU), École Polytechnique Fédérale de Lausanne (EPFL), Wenzhou Institute of Technology (WIT), University of Nottingham Ningbo China, Shenzhen University of Advanced Technology, North China Electric Power University, Zhejiang University, Peking University and the University of Oxford have developed an advanced AI-robotics framework that redefines how perovskite solar cells (PSCs) are synthesized, fabricated, and analyzed.
The study introduces a domain-specific recipe language model (RLM) integrated with 11 interconnected robotic boxes to achieve fully enclosed, automated, and feedback-driven experimentation for PSC research. At the heart of this system lies a seven-layer artificial intelligence (AI) architecture encompassing learning, generating, RecipeQA, fine-tuning, reasoning, evaluation, and optimization. This structure allows both numerical and semantic recipes - formulas and parameters derived from over 60,000 PSC-related studies - to be encoded into machine-readable formats, optimized by the language model, and translated into robotic instructions. Each robotic box contributes to a closed-loop workflow that connects recipe recommendation, fabrication, characterization, and semantic mechanistic analysis.
The 11 robotic boxes collectively executed 50,764 full-device PSC experiments, achieving a maximum power conversion efficiency (PCE) of 27.0%, with a certified value of 26.5%. During this large-scale experimentation, high-throughput characterization generated more than 578 million machine tokens, pairing experimental data with their corresponding mechanisms. This continuous data augmentation refined the RLM’s capacity for mechanism-grounded reasoning, pushing its performance to about 80% on dedicated evaluation metrics.
The hardware comprises 101 functional modules, over 1,500 mechanical and sensing components, and 4,300 controllable parameters - reconstructing traditionally fragmented manual operations inside interconnected robotic enclosures. A digital twin interface bridges the software and hardware, ensuring real-time translation of AI-generated recipes into robotic commands and returning immediate feedback for iterative fine-tuning.
According to the researchers, the key innovation lies in unifying three dimensions into one agentic framework: controllable device-scale fabrication, robotic characterization transforming raw data into mechanistic insights, and a continuously evolving RLM that integrates both numerical and semantic learning. This combination creates a powerful feedback loop capable of adaptive optimization and mechanistic reasoning at a scale and precision previously unattainable in perovskite research.
Beyond PSCs, this development signifies a shift toward materials intelligence - where AI-guided robotics bridge the gap between data-driven optimization and scientific understanding. The integrated framework demonstrated here lays the groundwork for autonomous materials discovery and in the future could be deployed in remote or extreme environments for on-site intelligent manufacturing.
