摘要
Modern AI and cloud data centers rely on high-density magnetic storage systems whose performance is increasingly influenced by coupled thermal, airflow, structural-vibration, and workload interactions. Although recent machine-learning techniques have demonstrated excellent predictive capability for storage-performance modeling, most existing approaches are optimized for prediction rather than interpretation. Consequently, they provide limited understanding of the underlying coupled thermo-vibro mechanisms responsible for location-dependent performance degradation, making it difficult to derive effective engineering interventions. This work presents a physics-informed dependency-learning framework based on a Graphical Gaussian Model (GGM) to characterize conditional dependencies among coupled operational responses in high-density magnetic storage systems. Physics-based simulations, experimental measurements, and a digital twin are integrated to construct synchronized multi-physics operational states, from which slot-wise conditional dependency networks are inferred through partial-correlation analysis. A quantitative dependency score is further developed to identify dependency-critical storage regions and guide localized engineering interventions. Engineering evaluation using localized ventilation and vibration-isolation modifications further demonstrates the practical utility of dependency-guided spatial prioritization for storage-performance assessment. The proposed framework provides an interpretable and computationally efficient methodology for dependency-aware analysis and reliability-oriented optimization, establishing a foundation for explainable machine learning and physics-informed design of future high-density storage infrastructures.