Radiation behavior of analog neural network chips

The paper summarizes the neural network experiment conducted for the Space Technology Research Vehicle (STRV) l-b launched in June 1994. A set of identical analog feedforward neural network chips was used to study and compare the effects of space and ground radiation on the device characteristics. Reconfigurable synapses as multiplying digital-analog converters with weights stored in digital memory and neurons as variable gain amplifiers with sigmoidal output were used in a feedforward architecture. The network and device characteristics were monitored in ground-based testing as well as during geotransfer orbits in flight using shielded and exposed identical VLSI chips. Ground based degradation depended on whether the chips were electrically biased during test or not. Unbiased chips were operational to about 100 krads of total dose. Space based radiation effects were mostly with the chips unbiased, with a dose of about 11 krads during 228 orbits. The slope of the linear part of a typical synapse-neuron input/output curve was found to become steeper with increasing dose. A correlation of the slope degradation was obtained between ground and space radiations. This paper provides a summary of the experiments and accumulated and analyzed data, and points out three failure mechanisms.