The accurate simulation of cardiac excitation is vital for understanding arrhythmias [1]. However, traditional high-fidelity solvers, fail to provide the real-time results required for clinical applications. While Physics-Informed Neural Networks (PINNs) have addressed some of these challenges, they are limited to single PDE solutions, requiring expensive retraining whenever tissue parameters or the boundary conditions of the problem change [2]. This work presents a model based on Deep Operator Networks (DeepONets) [3] that shifts the focus from individual PDE solutions to learning the solution operator of the Aliev-Panfilov model. To enable patient-specific customization, we utilize 1D Convolutional Neural Networks (CNNs) to process Action Potential (AP) curves as input functions, capturing unique tissue propagation characteristics and preventing information loss during discretization. To mitigate the black-box nature of neural surrogates, we introduce a regularized latent space through contrastive loss mechanisms and β-Variational Autoencoders (β-VAE). This framework disentangles the manifold of solutions into three distinct physical regimes: wave blockage, non-interactive propagation, and spiral wave generation. Our findings demonstrate that the variables within this learned latent space exhibit strong correlations with clinical biomarkers, such as the APD90. Such correlations establish a link between deep learning features and myocardial restitution properties. The resulting surrogate model achieves a 363x computational speedup, performing dense-solution inference in just 0.135 seconds compared to 49 seconds for a traditional solver. This framework offers a scalable and customizable tool for cardiac simulation, enabling real-time evaluation for different scenarios without the need for model retraining. REFERENCES [1]​ R. Aliev and A. Panfilov, “A simple two-variable model of cardiac excitation”, Chaos, Solitons & Fractals, 7(3), 293–301 (1996). [2]​ C. Martin, A. Oved et al., “EP-PINNs: Cardiac Electrophysiology Characterisation Using Physics-Informed Neural Networks”, Frontiers in Cardiovascular Medicine, 8:768419 (2022). [3] ​L. Lu, P. Jin, et al., “Leaning nonlinear operators via DeepONet based on the universal approximation theorem of operators”, Nature Machine Intelligence, 3(3), 218–229 (2021).