Deep earthquakes, 300 to 700 km deep, have been observed for decades and shown to originate from major mineral transformations occurring at these depths, including phase transitions of olivine and pyroxenes. Yet, we still do not fully grasp their mechanism. Although transformational faulting in the rim of the metastable olivine wedge (MOW) is hypothesized as a triggering mechanism of deep-focus earthquakes, there is no direct seismic evidence of such rim. Variations of b-value – slope of the Gutenberg-Richter distribution – have been used to decipher triggering and rupture mechanisms of earthquakes. However, regarding deep-focus earthquakes the detection limit prevents full understanding of rupture nucleation at all sizes. With one of the most complete catalogs, the Japan Meteorological Agency (JMA) catalog, we estimate the b values of deep-focus earthquakes (> 300 km) of four clusters in the NW Pacific Plate based on unsupervised machine learning. The applied K-means algorithm divides the events into four clusters. For the first time, we observe kinks in the b values with abrupt reductions from 1.5–1.8 down to 0.7–1.0 at a threshold Mw of 3.7–3.8 for the Honshu and Izu clusters, while normal constant b values (0.9–1.0) are observed for the Bonin and Kuril clusters. The four clusters found by the algorithm actually correspond to events within four different segments of the sinking Pacific lithosphere, characterized by significant differences in hydration state prior to subduction. High b values (1.5–1.8) at low magnitudes (Mw < 3.7–3.8) correlate with highly hydrated slab portions. The hydrous defects would enhance the nucleation of small earthquakes via transformational faulting within the rim. Such mechanism operates for small events with a rupture length of less than 1 km, which would correspond to the thickness of the MOW rim. Combining with the b-value analysis from the latest CMT catalog, the kink at Mw 6.7 suggests that the thermal runaway mechanism operates for larger earthquakes rupturing through and possibly propagating outside the MOW, with increased heterogeneity in the new rupture domain. The changes of controlling mechanism and rupture domain heterogeneity due to the slab hydrous state and thermal state can explain the spatially varying b values.