Automated Model Discovery for Skin: Discovering the Best Model, Data, and Experiment

This work introduces a new class of Constitutive Artificial Neural Networks (CANNs) that autonomously discover the optimal constitutive model and material parameters from experimental data. The authors trained the network using biaxial extension datasets from rabbit and pig skin, acquired through mechanical tests performed on a CellScale BioTester 5000. The CANN framework embeds kinematic, thermodynamic, and polyconvexity constraints into its architecture, enabling the network to identify meaningful constitutive equations directly from data. The discovered models revealed that skin mechanics can be represented by a two-term exponential function based on the first and fourth strain invariants, corresponding to isotropic matrix and anisotropic collagen fiber contributions. The network also identified the optimal experimental protocols for maximizing model information, highlighting off-axis biaxial tests as the most informative for skin mechanics.