HPIDF: A Hybrid SDAERandom Forest for Exascale Interconnect Prognostics

The transition to exascale computing has exposed vulnerabilities in high-radix interconnects where soft errors and transient performance fluctuations can substantially degrade throughput. Traditional threshold-based monitoring cannot reliably separate stochastic telemetry noise from incipient hardware degradation. We present a hybrid, physics-informed predictive maintenance framework that combines an Arrhenius-based degradation feature with a lightweight two-stage ML pipeline. The architecture uses a proposed Stacked Denoising Autoencoder (SDAE) for denoising and a Cost-Sensitive Random Forest (RF) for failure classification; the experiments in this submission report results from a localized, failure-injection pilot. On synthetic telemetry (N=10,000; test set = 2,500) the RandomForest pilot achieved Precision = 0.730, Recall = 0.703, F1 = 0.716, and ROC-AUC = 0.916; detected failures had mean lead time 32.8 h (median 29.9 h). These controlled results are illustrative and further large-scale, real-world validation is required before deployment-grade claims.