This study investigates the use of neural networks as surrogate models for electrophysiology simulations, focusing on their application to the FitzHugh-Nagumo model. It compares Data-Driven Neural Networks (DDNNs), Physics-Informed Neural Networks (PINNs), and Iterative Neural Networks (ITNNs) to identify the most effective approaches for real-time, scalable, and accurate simulations in clinical and research contexts.
Three neural network architectures were trained using numerical solutions derived from the FitzHugh-Nagumo model under different problem scenarios. DDNNs relied on data-driven optimization, PINNs incorporated physical laws into loss functions, and ITNNs iteratively approximated time-step transitions. Extensive architecture optimization and GPU-accelerated inference via TensorRT were employed to benchmark model performance against traditional Euler solvers, focusing on accuracy, computational efficiency, and scalability.
DDNNs demonstrated superior accuracy in data-rich environments, outperforming PINNs and ITNNs. PINNs proved advantageous in data-scarce scenarios but incurred higher computational costs. ITNNs struggled with error accumulation in long-term predictions. Small, optimized neural networks achieved comparable or faster inference speeds than numerical solvers, while retaining differentiability. Architectural choices, such as neuron count and activation functions, significantly influenced performance in smaller models, whereas larger models achieved consistent results but required higher computational resources.