Bridging the Gap: Physics-Driven Deep Learning for Heat Transfer Model

Abstract

The potential of Physics-Informed Neural Networks (PINNs) in addressing intricate real-world challenges exceeds the capabilities of traditional deep learning methods by merging data-driven and physics-driven approaches. PINNs offer advantages in handling computationally and temporally demanding problems that were conventionally approached using techniques like Finite Difference Method (FDM). This study focuses on analyzing heat transfer within cardiac tissue to deepen our understanding of the mechanism from ablation catheters or implants to surrounding blood and tissue layers. Accurately estimating temperature increases is crucial for designing reliable catheters and mitigating overheating risks. By defining the geometry and boundary conditions to simulate catheter-induced heating, we compare the results with state-of-the-art solutions such as FDM. PINN models demonstrate high accuracy, validated against FDM, with notable benefits including reduced reliance on handcrafted features, adaptability to complex domains, and efficient training and inference. These results can guide future investigations of the behavior of cardiac tissue under varied conditions.