For over 35 years of research, the debate about the systematic compositionality of neural networks remains unchanged, arguing that existing artificial neural networks are inadequate cognitive models. Recent advancements in deep learning have significantly shaped the landscape of popular domains, however, the systematic combination of previously trained neural networks remains an open challenge. This study presents how to dynamically synthesize a neural network for the design of broadband electromagnetic metasurfaces. The underlying mechanism relies on an assembly network to adaptively integrate pre-trained inherited networks in a transparent manner that corresponds to the metasurface assembly in physical space. This framework is poised to curtail data requirements and augment network flexibility, promising heightened practical utility in complex composition-based tasks. Importantly, the intricate coupling effects between different metasurface segments are accurately captured. The approach for two broadband metasurface inverse design problems is exemplified, reaching accuracies of 96.7% and 95.5%. Along the way, the importance of suitably formatting the spectral data is highlighted to capture sharp spectral features. This study marks a significant leap forward in inheriting pre-existing knowledge in neural-network-based inverse design, improving its adaptability for applications involving dynamically evolving tasks. An eco-conscious synthetical neural network for metasurfaces is proposed. This approach employs knowledge inheritance from previous designs, akin to passing knowledge from "parent" to "offspring". Proper spectral data formatting is emphasized to capture precise features. Two broadband metasurface inverse design examples show accuracies of 96.7% and 95.5%. This method simplifies metasurface design, minimizing modeling intricacies while enhancing sustainability and efficiency. image