<p>Deep mineral prospectivity prediction faces the challenge of effectively utilizing high-dimensional geological data, particularly in capturing the complex temporal evolution of ore-forming processes. Conventional machine learning methods predominantly rely on static geological features or single-snapshot physical fields, thereby overlooking the critical temporal dimension of mineralization. To address this gap, this study proposes a novel four-dimensional mineral prospectivity prediction approach that integrates long short-term memory (LSTM) networks with time-series data derived from coupled mechano-thermo-hydrological numerical simulations. Taking the Tongshan copper deposit in Anhui, China, as a case study, the dynamic evolution of volumetric strain, temperature, and pore pressure was reconstructed over a 660,000-year cooling history. These spatiotemporal dynamics features were structured as time-series sequences and fed into the LSTM model. Comparative analysis demonstrates that the LSTM model significantly outperforms standard artificial neural networks—including a benchmark model utilizing time-averaged dynamics features. This result confirms that learning temporal dependencies and evolutionary patterns (e.g., sequential fracturing and cooling) yields superior predictive accuracy compared to relying solely on the magnitude of physical parameters. Furthermore, the LSTM model identified a high-probability mineralization temporal window of approximately 110,000&#xa0;years during the early cooling stage, providing quantitative insight into the duration of the active hydrothermal system. Finally, the model delineated high-potential exploration targets in the undrilled southwestern deep zones of the Tongshan mine. This study demonstrates that fusing physics-based dynamics simulations with temporal deep learning offers a robust pathway for process-driven mineral prediction.</p>