<p>Predicting and understanding quench in high-temperature superconducting (HTS) materials pose significant challenges due to strong nonlinearity from extreme multifield coupling at cryogenic temperatures, as well the phase transition from the superconducting state to the normal one. Accurately determining the occurrence and location of a quench in HTS tapes, triggered by thermal disturbances, is crucial for advancing superconducting applications, as it requires solving a complex inversion problem. This study develops a deep learning framework to solve this inverse problem for predicting quench phenomena in YBCO tapes. We first construct a coupled forward model based on thermoelastic quench theory to simulate the quench onset and generate a high-fidelity dataset. A convolutional neural network is then designed to directly map temperature and strain distributions to the characteristics of the thermal disturbance, including its location, average power, and duration. The model, trained solely on physics-based simulation data, achieves exceptional performance with a mean relative error of less than 3% and demonstrates strong robustness. Through a transfer learning strategy, the trained network effectively generalizes to challenging scenarios involving noisy and time-series data while maintaining high accuracy, highlighting its superior fault tolerance and feature extraction capability. By integrating the predictive power of deep learning with physically grounded training data, this work provides an accurate and efficient tool for real-time quench prediction, offering a promising pathway to enhance the reliability and performance of superconducting systems.</p>