Agentic robotic boxes for perovskite solar cell fabrication with recipe language model

Abstract

Perovskite solar cells (PSCs) have emerged as a promising photovoltaic technology owing to their high efficiency, compositional tunability and compat ibility with solution processing. However, the fabrication of high-performance PSCs still suffers from time-consuming trial-and-error synthesis and labour-in tensive fabrication, because precursor formulas, additive selection, interfacial engineering and process parameters are highly coupled and collectively deter mine crystallisation behaviour, defect formation and device performance. Alt hough robotic experimentation and artificial intelligence (AI) are increasingly being introduced into materials research, existing approaches often remain lim ited to isolated fabrication, characterisation or numerical optimisation, and an integrated route that connects robotic fabrication, mechanistic characterisation and language-model-based learning for PSCs is still lacking. This thesis ad dresses this challenge by establishing an AI–robotics research framework based on robotic boxes for controllable fabrication, robotic characterisation augmented mechanism, and robotic recipes for language model training.

Firstly, an agentic AI–robotics framework was developed for PSCs fab rication and characterisation. Fabrication knowledge was encoded into struc tured formula–parameter (FP), enabling experimental procedures to be ex pressed as machine-readable and robot-executable units. On this basis, eleven robotic boxes and seven AI layers architecture were integrated to connect recipe generation, robotic execution, experimental feedback and model training. This framework establishes a closed recommendation–synthesis–fabrication–charac terisation–mechanism workflow, providing the methodological foundation for systematic PSCs research beyond conventional manual experimentation.

Secondly, controllable fabrication of PSCs was established by recon structing conventional manual procedures into modular robotic operations coor dinated through a digital twin. More than 50,000 robotic experiments were car ried out to explore perovskite composition, process parameters, additives, self assembled molecules (SAMs), passivators and interface regulation in a progres sive manner. Through this robotic optimisation workflow, device performance evolved from broad and stochastic distributions towards concentrated high-effi ciency regimes, ultimately achieving a champion power conversion efficiency (PCE) of 27.0%, with a certified PCE of 26.5%. These results demonstrate that robotic fabrication can transform PSCs development from empirical trial-and error into a controllable optimisation process.

Thirdly, robotic characterisation was incorporated to strengthen the mechanistic interpretation of fabrication outcomes. High-throughput imaging, X-ray diffraction (XRD), in-situ photoluminescence spectroscopy (PL), graz ing-incidence X-ray diffraction (GIXRD) and photovoltaic measurements were integrated to generate multimodal descriptors of film morphology, crystallisa tion dynamics, residual phases, interfacial structure, residual stress and device performance. The results show that different stages of optimisation contributed differently to photovoltaic improvement: early-stage optimisation of perovskite FP primarily enhanced short-circuit current density (JSC) through improved film formation, whereas subsequent introduction of SAMs, additives, surface pas sivation and buried-interface regulation mainly improved open-circuit voltage (VOC) and fill factor (FF) through reduced defects, modified crystallisation be haviour and improved interfacial selectivity. Robotic characterisation therefore extended the system beyond automated screening towards mechanism-relevant understanding.

Finally, the results generated by robotic fabrication and characterisation were transformed into robotic recipes for language model training. Robotic rec ipe reports were converted into both numerical and semantic recipe forms, and further organised into mechanism-aware corpora. With the corpus scaling to 578 million tokens, the recipe language model (RLM) achieved substantial gains in both recipe recommendation and mechanistic reasoning, reaching a final overall score of 80.9%. This shows that robotic recipes function not only as records of experiments, but also as a structured medium for converting robotic experimen tation into trainable domain knowledge.

Overall, this thesis establishes an integrated route for PSCs research that connects robotic fabrication, robotic characterisation and RLM training within a unified AI–robotics framework. By transforming experiments into structured, trainable and mechanism-relevant knowledge, this work provides a new meth odological basis for more controllable, interpretable and scalable research on PSCs, and offers a new paradigm for the convergence of robotics and language models in materials science.

Date of Award	15 Nov 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Cheng Heng Pang (Supervisor)