Cardiac mechanics tools can be used to enhance medical diagnosis and treatment, and assessment of risk of cardiovascular diseases. Still, the computational cost to solve cardiac models restricts their use for online applications and routine clinical practice. This work presents a surrogate model obtained by training a set of Siamese networks over a physiological representation of the left ventricle. Our model allows us to modify the geometry, loading conditions, and material properties without needing of retraining. Additionally, we propose the novel concept of intrinsic domain that improves the accuracy of the network predictions by one order of magnitude. The neural networks were trained and tested with numerical predictions from a previously published finite element model of the left ventricle. Different loading conditions, material properties and geometrical definitions of the domain were simulated by the model leading to a dataset of 5, 670 cases. In terms of accuracy and performance, the surrogate model approximates the displacement field of the finite element model with an error of 4.4 ± 2.9% (with respect to the _{2}-norm of the true displacement field) across all cases while performing computations 62 times faster. Hence, the trained model is capable of computing a passive cardiac filling of the chamber at 10 different time points in just ~0.7 s. These outcomes prove usability of training surrogate models for efficient simulations to facilitate the use of complex mechanical models in clinical practice for therapeutic planning and online diagnosis.