SchNet4AIM is a machine learning model based on the SchNet architecture, designed to predict local quantum chemical descriptors such as atomic charges, delocalization indices, and pairwise interaction energies. This model addresses the challenge of balancing accuracy and interpretability in chemical artificial intelligence (AI). Unlike traditional black-box models, SchNet4AIM provides physically meaningful predictions by leveraging rigorous local quantum chemical properties, enabling explainable chemical AI (XCAI). The model's performance is tested on a wide range of real-space quantities, demonstrating high accuracy and speed, which overcome previous limitations in using real-space chemical descriptors for complex systems. SchNet4AIM is implemented in the SchNetPack (SPK) package and is capable of predicting both one-body (atomic) and two-body (pairwise) terms. It uses a modified SchNet architecture to handle local chemical properties, offering robust models with minimal training data. The model's ability to extrapolate and transfer to new chemical scenarios is highlighted, as it can predict molecular properties in regions not sampled during training, demonstrating its versatility. The model's predictions are validated through various chemical processes, including the CO₂ capture and release by a Calix[4]arene molecule (13P). SchNet4AIM accurately predicts electron delocalization metrics, which are reliable indicators of supramolecular binding events. These predictions are physically coherent and provide insights into complex chemical phenomena, even in scenarios where geometrical features fail. SchNet4AIM's interpretability is further enhanced by its ability to trace predictions back to physically rigorous atomic or pairwise terms, aligning with Coulson's maxim of "giving us insight, not numbers." This approach allows for the development of explainable AI models that are both accurate and interpretable, contributing to the advancement of quantum chemistry and materials science. The model's performance is compared to previous approaches, showing significant improvements in accuracy and efficiency, making it a valuable tool for chemical research and applications.SchNet4AIM is a machine learning model based on the SchNet architecture, designed to predict local quantum chemical descriptors such as atomic charges, delocalization indices, and pairwise interaction energies. This model addresses the challenge of balancing accuracy and interpretability in chemical artificial intelligence (AI). Unlike traditional black-box models, SchNet4AIM provides physically meaningful predictions by leveraging rigorous local quantum chemical properties, enabling explainable chemical AI (XCAI). The model's performance is tested on a wide range of real-space quantities, demonstrating high accuracy and speed, which overcome previous limitations in using real-space chemical descriptors for complex systems. SchNet4AIM is implemented in the SchNetPack (SPK) package and is capable of predicting both one-body (atomic) and two-body (pairwise) terms. It uses a modified SchNet architecture to handle local chemical properties, offering robust models with minimal training data. The model's ability to extrapolate and transfer to new chemical scenarios is highlighted, as it can predict molecular properties in regions not sampled during training, demonstrating its versatility. The model's predictions are validated through various chemical processes, including the CO₂ capture and release by a Calix[4]arene molecule (13P). SchNet4AIM accurately predicts electron delocalization metrics, which are reliable indicators of supramolecular binding events. These predictions are physically coherent and provide insights into complex chemical phenomena, even in scenarios where geometrical features fail. SchNet4AIM's interpretability is further enhanced by its ability to trace predictions back to physically rigorous atomic or pairwise terms, aligning with Coulson's maxim of "giving us insight, not numbers." This approach allows for the development of explainable AI models that are both accurate and interpretable, contributing to the advancement of quantum chemistry and materials science. The model's performance is compared to previous approaches, showing significant improvements in accuracy and efficiency, making it a valuable tool for chemical research and applications.