The translation of laboratory-scale photovoltaic devices to commercially viable manufacturing remains a critical bottleneck, particularly for organic and hybrid perovskite solar cells where vast, interdependent parameter spaces complicate optimization. We present our group's comprehensive self-driving laboratory (SDL) ecosystem for accelerating photovoltaic innovation, integrating purpose-built hardware platforms, critically evaluated machine learning software, and targeted materials development to create closed-loop autonomous discovery systems spanning the full research-to-manufacturing pipeline.

Our high-throughput hardware centres on low cost. modified 3D printer systems such as the MicroFactory platform, a roll-to-roll fabrication and characterization system producing over 11,800 organic photovoltaic devices within 24 hours across 64.8 metres of flexible substrate. This throughput transforms intractable optimization problems involving 36 manufacturing variables requiring 68.7 billion experiments through factorial design into tractable exhaustive searches. Using printing-inspired digital twins and inverse parameter generation, we achieved 1.1% absolute PCE improvement in a single optimization iteration for PBF-QxF:Y6 devices exceeding 8.5% efficiency. Recognizing that commercial automation costs restrict SDL adoption, we further demonstrate that consumer-grade 3D printers modified with custom syringe attachments can serve as automated liquid handlers, reducing costs by two orders of magnitude while enabling distributed robotic architectures for coordinated multi-station perovskite device fabrication.

Our software development critically examines assumptions underlying SDL optimization strategies. Through systematic evaluation of 11,587 devices across 25 optimization iterations, we find that Bayesian optimization achieves only marginally higher PCE (7.69%) compared to random search (7.66%), with environmental factors, particularly humidity, showing stronger performance correlation than algorithm selection. This suggests that robust process control supersedes algorithmic sophistication for manufacturing-scale systems. We extend our software capabilities through large language model integration, developing retrieval-augmented generation (RAG) chatbots for research assistance and establishing rigorous benchmarking frameworks using RAGAS evaluation metrics to assess factual correctness, context recall, and faithfulness against expert baselines. Such validation becomes essential as LLM agents increasingly support autonomous experimentation.

Materials innovation within our SDL framework addresses the stability limitations hindering perovskite commercialization. We introduce PBDF-DFC, a biomass-derived furan-based conjugated polymer enabling simplified one-pot precursor integration into hybrid perovskite solar cells. Unlike petroleum-derived thiophene polymers with limited precursor solubility, PBDF-DFC dissolves directly in perovskite solutions, streamlining fabrication while achieving 21.39% PCE, a 7.8% relative improvement over controls. Crucially, devices retain 90% initial efficiency after 1100 hours under environmental stress, compared to 52% for unmodified controls. Transmission electron microscopy reveals polymer accumulation at grain boundaries, passivating defects through oxygen-lead coordination. Complementing this work, we demonstrate non-fullerene acceptor interlayers for SnO₂ electron transport layer modification, where cyano, carbonyl, and halogen functional groups bind electrostatically to oxygen vacancies, reducing trap-state density and enhancing charge extraction in hybrid organic-perovskite architectures.

Looking forward, we position current SDLs within an evolution from basic automation toward fully autonomous discovery systems. Present implementations achieve approximately 71% recipe success rates with single-domain focus. Near-term integration phases will enable cross-domain optimization and manufacturing awareness. Ultimately, agentic discovery systems incorporating goal-directed reasoning and integrated knowledge representations will autonomously navigate from hypothesis generation through experimental validation to scale-up. Our work demonstrates that combining high-throughput hardware, rigorously validated software, and targeted materials expertise creates a powerful engine for next-generation photovoltaic development.