Radiation exposure effects on the performance of an electrically trainable artificial neural network (ETANN)

The effects of radiation exposure on an analog neural network devices are examined. The neural network implements a fully parallel architecture integrating 10240 analog nonvolatile synapses fabricated in a CMOS process. Graceful degradation of forward propagation performance was observed in units that were exposed to absorbed doses of up to 26 krads (Si) of 10-MeV electrons. The units were exposed without bias, except for that due to the floating gates. Single-chip solutions to two pattern recognition problems representing two levels of difficulty are employed for testing. Over the weeks following exposure, postirradiation effects due to a latent charge trapping mechanism in the oxides of the nonvolatile floating gate structures are observed. It is shown that units with apparently permanent damage can be retrained to 100% recognition performance. >