Interpretable machine learning and active learning-driven design of graphene quantum dot-containing transparent thermal emitters for photovoltaic encapsulation applications

Designing transparent thermal emitters requires a balance between high visible light transmittance and excellent radiative cooling performance. However, optimising performance across multiple objectives usually involves introducing various functional materials. This significantly increases the complexity of material preparation and performance regulation, making inefficient trial-and-error experimental methods inadequate to meet current demands. Furthermore, when it comes to multi-objective material design, the introduction of artificial intelligence (AI) technology necessitates a primary focus on clarifying whether the design logic of the algorithm aligns with the physicochemical mechanisms of the materials, as well as improving the efficiency of material design overall. In this study, a transparent thermal emitter incorporating graphene quantum dots (TTEG) is developed using interpretable machine learning (PXML) and an active learning framework with prior-knowledge constraints (ALL-PV). Six predictive models are established to estimate the fluorescence intensity of graphene quantum dots (GQDs). PXML is applied to identify the key factors governing model decisions and to clarify the associated physicochemical mechanisms. The results show that reaction temperature and catalyst molecular polarity play decisive roles in determining the fluorescence intensity of GQDs in solvothermal growth systems. Based on this understanding, we further optimized the reaction conditions and catalyst systems. The photoluminescence (PL) intensity of the resulting optimized sample, E-81, exceeded the highest value in the training dataset, and its photoluminescence quantum yield (PLQY) reached 38.9%. Next, the TTEG is fabricated from E-81 and tricyclodecanedimethanol diacrylate (DCPDA), with ALL-PV guiding the optimization of the preparation process. In ALL-PV, prior-knowledge constraints, the surrogate model, and the acquisition function work together to direct each round of experimental selection. Within a design space of more than 390,000 possible process combinations, this strategy locates the optimal preparation conditions after only five rounds of iteration and 60 experiments. The resulting fluorescent-particle-containing TTEG shows a transmittance of 95.6%. While exhibiting outstanding thermal management performance. Furthermore, we employed the finite-difference time-domain (FDTD) method to simulate the TTEG, revealing the physical mechanism by which GQDs induce the enhancement of TTEG transmittance. Finally, the energy-saving performance and stability of the TTEG are further evaluated. In Beijing, Kunming, and Haikou, the annual cooling energy consumption of the TTEG is 9.8, 8.75, and 11.45 MJ·m-2 lower than that of ordinary glass, respectively. Simultaneously, the TTEG exhibits excellent chemical stability (ΔTvis < 1%), thermal stability, and mechanical stability. When the TTEG is applied to solar cell encapsulation, the power conversion efficiency of a 5 cm × 5 cm crystalline silicon solar cell increased by 17.6% (AM 1.5, 100 mW·cm-2). For a 41 cm × 30 cm commercial crystalline silicon solar cell under all-day illumination conditions (8:00 AM to 6:30 PM), its average operating temperature decreased by 3.4 °C, with a maximum temperature drop of up to 7 °C, while the maximum output power increased by approximately 10.8%. This study lays the foundation for the engineering development of transparent thermal emitters and highlights the potential of AI-guided design for multi-step, multi-property materials.