SSRESF-OPT: A Sensitivity-Aware Framework for Single-Particle Radiation Effects Simulation and Optimization in SoC Platforms Using Machine Learning

Abstract

The increasing scale of System-on-Chip (SoC) architectures exacerbates the reliability risks of single-particle radiation effects, rendering traditional exhaustive gate-level fault injection prohibitively slow. To address this, we propose SSRESF-OPT, an end-to-end simulation and optimization framework combining accelerated dynamic fault injection with machine learning-driven prediction. The framework first introduces a cross-layer static binding mechanism to eliminate runtime string resolution, reducing per-injection simulation overhead to \(\mathcal {O}(1) \) complexity. Building on this optimized engine, an SVM-based classification pipeline, guided by structural clustering, accurately predicts the radiation vulnerability of unsimulated circuit nodes using netlist-derived features. To maximize resource efficiency, an OPT-based top-k hardening strategy is then employed to prioritize critical architectural bottlenecks. Evaluations across 15 gate-level netlists, including massive AES subsystems and RISC-V PULP SoCs, demonstrate the framework’s efficacy. The model-assisted approach achieves an average end-to-end runtime speedup of 6.35 × over the highly optimized simulation baseline while maintaining a 94.58% prediction accuracy. Furthermore, hardening only the top 10% of ranked nodes reduces system Soft Error Rate (SER) by 85%, providing a fast and actionable methodology for radiation-hardened microprocessor design.