To solve the inverse problem of material parameter identification, constitutive models are classically calibrated using iterative approaches such as the Finite Element Model Updating method (FEMU). Most modern approaches that introduce Artificial Neural Networks to address inverse problems in this field of research do not fully exploit their pattern recognition capabilities. Instead, these approaches train Physics-Informed Neural Networks for each new problem, requiring computationally expensive optimisation by analogy with FEMU. This work presents a neural network architecture with separate pathways that resemble different aspects of the expected material response. The proposed design improves interpretability and provides intrinsic uncertainty estimates that aid in the interpretation of the solution obtained. The network is trained on a comprehensive dataset of finite element simulations, learning a direct mapping from full-field displacement and global force data to material parameters. Once trained, predictions are obtained through a single forward pass, eliminating the iterative optimisation required within classical methods and offering significant advantages for rapid parameter identification.