Accurate estimation of seismic losses in buildings is crucial for mitigating earthquake disasters. Existing machine learning-based seismic loss prediction methods primarily classify buildings into different damage levels, which limits prediction accuracy. To address this limitation, the wave equation was employed as the activation function in a convolutional neural network (CNN) to directly predict loss rates based on 398 324 actual building damage records from the 2011 Great East Japan Earthquake, thereby enabling refined loss estimation. The results indicate that the CNN model incorporating the wave equation as an activation function outperforms traditional activation functions in three key evaluation metrics: mean squared error, mean absolute error, and the coefficient of determination. Notably, the model demonstrates a significant improvement in predicting loss rates for severely damaged buildings, achieving an accuracy of up to 79.47%. Furthermore, the physical foundation of the wave equation enhances the interpretability of the model, enabling not only effective seismic loss prediction but also providing a theoretical basis for the model itself. The result can be applied to seismic loss assessment and risk pricing in earthquake insurance while also offering new insights into the intersection of earthquake engineering and machine learning.