Rapid mechanical property prediction and <i>de novo</i> design of three-dimensional spider webs through graph and GraphPerceiver neural networks

Spider webs feature advanced structural performance due to the evolutionary success of over more than 3 × 109 years, including lightweight design and exceptional mechanical properties. are appealing for bio-inspired since web designs serve multiple functions protection prey catching. However, high computational cost limited quantified properties render extensive spider studies challenging in part complexity randomness fiber arrangements 3D webs. Here, we report a method relate graph microstructures effective properties, focusing on strength toughness, upscaling from microscopic mesoscale level. The new framework uses deep neural networks, trained graph-structured Cyrtophora citricola data, order capture complex cross-scale relationships. Three different models developed compared. First, two Graph Neural Network (GNN) models, Convolutional Network, Principal Neighborhood Aggregation method. Second, GraphPerceiver transformer model that is fed similar input data as provided GNN approach but within natural language modeling context using self-attention mechanisms. can achieve model, offering added flexibility building learning diverse hierarchical biological materials. As an application propose optimization tool synthetic used generate synthetic, de novo architectures. Finally, multi-objective enables us discover structures meet specific objectives.