<p>An accurate prediction of the size of an earthquake is vital to successfully prepare for the disaster and mitigate risks. This study enhances machine learning prediction using spatio-temporal seismic data by combining advanced machine learning models with physical properties, such as energy-depth interactions, which utilize explainable AI approaches. The work highlights the dynamic nature of the seismic pattern by examining four dynamic datasets, emphasizing the need for periodic software updates to maintain effectiveness as models age. The explainability methods based on SHAP are used to investigate feature contributions, allowing for model transparency and actionable insights in terms of important drivers like energy, temperature, and space. Moreover, this has strengthened the operational framework for effective model versioning, tracking of experiments, and deployment, all contributing to the scalability for real-time applications. Overall, this study emphasizes the need to integrate explainable AI, machine learning approaches, and domain knowledge to mitigate the hurdles related to the non-stationarity of earthquake datasets, moving towards resilient, interpretable, and adaptable earthquake prediction solutions.</p>