Composing Graph Theory and Deep Neural Networks to Evaluate SEU Type Soft Error Effects

Rapidly shrinking technology node and voltage scaling increase the susceptibility of Soft Errors in digital circuits. Soft Errors are radiation-induced effects while the radiation particles such as Alpha, Neutrons or Heavy Ions, interact with sensitive regions of microelectronic devices/circuits. The particle hit could be a glancing blow or a penetrating strike. A well apprehended and characterized way of analyzing soft error effects is the fault-injection campaign, but that typically acknowledged as time and resource-consuming simulation strategy. As an alternative to traditional fault injection-based methodologies and to explore the applicability of modern graph based neural network algorithms in the field of reliability modeling, this paper proposes a systematic framework that explores gate-level abstractions to extract and exploit relevant feature representations at low-dimensional vector space. The framework allows the extensive prediction analysis of SEU type soft error effects in a given circuit. A scalable and inductive type representation learning algorithm on graphs called GraphSAGE has been utilized for efficiently extracting structural features of the gate-level netlist, providing a valuable database to exercise a downstream machine learning or deep learning algorithm aiming at predicting fault propagation metrics. Functional Failure Rate (FFR): the predicted fault propagating metric of SEU type fault within the gate-level circuit abstraction of the 10-Gigabit Ethernet MAC (IEEE 802.3) standard circuit.