Simulations of cardiac disease remain the cornerstone of discovery and treatment development. Yet current cardiac simulation execution times remain unsuitable for near real-time clinical decisions, as well as limit cardiac model complexity. To address this, simulations need to execute in a tractable clinical timeframe (within minutes). Scientific Machine Learning (SciML) exploits recent advances in machine learning to solve physics-based models. We utilized our SciML approach, neural network finite element (NNFE), to perform realistic high-speed simulations of the normal and infarcted heart using a comprehensive heart model dataset. Herein we demonstrate the ability to generate high-fidelity pressure-volume (PV) and wall stress responses in a matter of seconds. The NNFE code was rewritten to a JAX-based software platform to greatly enhance the performance and flexibility of the code. The method uses a neural network to predict the displacement field based on the physiological range of control variables: pressure and active stress. Finite elements are then used to compute the residual of the PDE, which is the loss function, so no data is required. For the problem, we utilized a high-fidelity ovine heart dataset, including geometry, fiber structure, and pressure-volume from a single heart. The NNFE neural networks for healthy and infarcted models were trained until the errors over the physiological range were within 0.1% error of the standard finite element simulation displacements. This is checked after training and is not used during training. Once trained, the time to evaluate the NNFE cardiac model was approximately 1 second, with a complete pressure-volume loop in 2-4 seconds. In addition to accurate displacement, stress, and strain tensors, the NNFE method was able to accurately predict the effects of acute infarcts. We are currently incorporating B-spline to allow patient-specific geometries and reduce the problem size to reduce training and execution times.