Multiple sensitive volume based soft error rate estimation with machine learning

We propose a new methodology for soft error rate estimation using multiple sensitive volumes and machine learning. The proposed methodology assigns multiple sensitive volumes to a unit circuit (e.g. SRAM cell) and constructs a discriminator from TCAD simulations by machine learning. For each ion reproduced by radiation transport simulation, the discriminator judges whether an upset occurs or not, and consequently we can obtain soft error rate by counting the number of events judged as upset events. Advantages of the proposed methodology are: (1) empirical construction and adjustment of sensitive volume and critical charge is no longer necessary, (2) multiple transistors can be easily considered, and (3) event-wise accuracy can be improved. We confirmed the correlation between irradiation results and simulation results for 65-nm silicon on thin buried oxide (SOTB) SRAM. The estimation error was 7% without any empirical optimization of sensitive volume and critical charge.