A Methodology for Identifying Critical Sequential Circuits with Graph Convolutional Networks

The reliability of circuits becomes increasingly important with the transistor scaling. One efficient approach to address this problem is to select the critical components and apply hardening techniques only to them. The question of how to identify such critical parts is one of the key challenges. This paper explores the potential of using Graph Convolutional Networks (GCNs) instead of fault simulation to identify critical sequential circuits at the netlist level. GCNs can learn from both features of individual sequential gates and structural features of the circuit topology. Our proposed methodology converts circuits into graphs by taking each sequential gate as a node and using Breadth-First Search (BFS) to search the shortest path between these sequential gates to create edges. Then GCNs are used to learn from the graphs to predict critical sequential gates. We evaluate two types of GCNs, including GCN with and without combinational logics. Finally, we validate the approach on an open-source RISC-V core. The experiments results indicate that GCN with combinational logics has the best prediction performance, and the achieved accuracy is up to 91.1%.