Groundwater depletion threatens water security, agriculture, and livelihoods in semi-arid regions such as Bannu Division, Pakistan, where reliable local-scale groundwater storage information remains limited. Existing GRACE/GRACE-FO studies estimate regional groundwater changes, but coarse resolution and sparse in-situ data mask sub-regional variability and reduce management usefulness. This study addressed this gap by downscaling GRACEderived groundwater storage from 55 km to 5 km using a physics-informed machine-learning framework, followed by k-means clustering, STL decomposition, heat-map analysis, and frequency distribution. Results showed groundwater storage anomalies from +120 to -180 mm, with 61% of grid cells persistently negative and central depletion hotspots =-140 mm covering 28-32% of the division. Downscaling increased spatial variance by 2.3 times while preserving long-term trends. Four groundwater regimes were identified; two depleted clusters occupied 56% of the area and showed mean anomalies of -90 to -150 mm. Seasonal variability explained 55-70% of variance, while depleted clusters declined by 4-7 mm yr?¹, causing cumulative losses of 80-100 mm. Negative anomalies increased from 38% of months in 2003-2009 to 72-76% after 2015. The study concludes that depletion is structurally persistent, mainly driven by abstraction exceeding episodic recharge and recommends high-resolution monitoring, localized abstraction control, efficient irrigation, stronger in-situ validation and predictive climate-abstraction modeling.