Semiconductor manufacturing is characterized by a massive influx of data and high intercorrelation among process variables, making it difficult to extract insights. Therefore, acquiring mechanistic knowledge regarding in-chamber physical phenomena is critical for efficient process control and optimization, going beyond experimental design based on trial and error. To address this need, this paper introduces a novel framework that repurposes eXplainable AI (XAI) from a passive interpretive tool into an active "hypothesis generator" for scientific discovery. We validate this approach using cryogenic plasma etching of silicon dioxide (SiO2) and silicon nitride (SiN), a process governed by complex mechanisms at low temperatures. By analyzing 86 systematic experimental cases characterized by real-time Optical Emission Spectroscopy (OES) and Quadrupole Mass Spectrometry (QMS), we utilized SHapley Additive exPlanations (SHAP) to extract feature interactions. Unlike traditional methods, we translated these XAI insights into explicit mechanistic hypotheses specifically concerning the formation of the ammonium fluorosilicate (AFS) layer and surface reaction pathways. These conjectures were subsequently verified through surface analysis techniques (FT-IR, XPS, SEM) and further confirmed via Molecular Dynamics (MD) simulations to ensure physical validity. Our results resolve the etch selectivity inversion mechanisms at low temperature, proving the effectiveness of the proposed framework in complex environments. Beyond this specific application, the study demonstrates a generalizable workflow where algorithmic explanations actively guide experimental design. By showing how data-driven insights can directly inform and refine physical validation strategies, this work effectively bridges the persistent gap between empirical statistical correlations and deep causal understanding in advanced semiconductor manufacturing systems.