This study proposes a geologically constrained convolutional neural network (CNN) for mineral prospectivity mapping (MPM) in western Henan Province, China, to improve the accuracy and interpretability of mineral exploration. The CNN integrates both soft and hard geological constraints to guide the model training. The soft constraint is based on a penalty term derived from the spatial coupling relationship between ore-controlling strata and gold deposits, incorporating additional prior geological knowledge. The hard constraints are derived from domain knowledge related to mineralization processes and geochemical indicators, embedded in the feature extractor and classifier of the CNN, respectively. These constraints ensure that the model training is based on the mineralization mechanism. Comparative experiments show that the geologically constrained CNN outperforms other models, demonstrating the effectiveness of integrating data and domain knowledge for MPM. This approach enhances the rationality and interpretability of the results. Mineral prospectivity mapping is a key task in mineral exploration, aiming to identify undiscovered mineral deposits by delineating potential zones. Recent advances in artificial intelligence (AI) have enabled data-driven MPM using various AI algorithms. Deep learning algorithms (DLAs), including pixel-based, image-based, and graph-based methods, have been applied to MPM. However, purely data-driven models often overlook geological domain knowledge, leading to poor interpretability and inconsistency in mineralization patterns. To address this, researchers have integrated domain knowledge into DLAs, enhancing model interpretability. The integration of data and domain knowledge is crucial for accurately characterizing complex and dynamic mineral systems. This study demonstrates that incorporating geological constraints into CNNs improves the effectiveness of MPM, providing a more reliable and interpretable approach for mineral exploration.This study proposes a geologically constrained convolutional neural network (CNN) for mineral prospectivity mapping (MPM) in western Henan Province, China, to improve the accuracy and interpretability of mineral exploration. The CNN integrates both soft and hard geological constraints to guide the model training. The soft constraint is based on a penalty term derived from the spatial coupling relationship between ore-controlling strata and gold deposits, incorporating additional prior geological knowledge. The hard constraints are derived from domain knowledge related to mineralization processes and geochemical indicators, embedded in the feature extractor and classifier of the CNN, respectively. These constraints ensure that the model training is based on the mineralization mechanism. Comparative experiments show that the geologically constrained CNN outperforms other models, demonstrating the effectiveness of integrating data and domain knowledge for MPM. This approach enhances the rationality and interpretability of the results. Mineral prospectivity mapping is a key task in mineral exploration, aiming to identify undiscovered mineral deposits by delineating potential zones. Recent advances in artificial intelligence (AI) have enabled data-driven MPM using various AI algorithms. Deep learning algorithms (DLAs), including pixel-based, image-based, and graph-based methods, have been applied to MPM. However, purely data-driven models often overlook geological domain knowledge, leading to poor interpretability and inconsistency in mineralization patterns. To address this, researchers have integrated domain knowledge into DLAs, enhancing model interpretability. The integration of data and domain knowledge is crucial for accurately characterizing complex and dynamic mineral systems. This study demonstrates that incorporating geological constraints into CNNs improves the effectiveness of MPM, providing a more reliable and interpretable approach for mineral exploration.