In Situ Analog In-Memory Computing Under Ionizing Radiation Exposure

We experimentally performed in situ analog in-memory computing (IMC) under ionizing radiation, using a 40-nm silicon-oxide–nitride-oxide–silicon (SONOS) charge-trap memory array with peripheral circuits that support analog matrix-vector multiplication (MVM) operations. The SONOS array used analog MVMs to process the last layer of a convolutional neural network (CNN) for TinyImageNet image classification while being irradiated by gamma rays from a Co-60 source. We experimentally characterized how the following quantities were gradually degraded by increasing the total ionizing dose (TID), up to 3.2 Mrad(Si): neural network weights that were mapped to SONOS states, dot products that were computed by analog MVMs, and the resulting image classification accuracy of the neural network. Using multiscale modeling, we confirmed that the experimentally observed accuracy loss originates almost entirely from the state-dependent current shifts induced by ionizing radiation in the SONOS memory cells. Our experimentally validated model of radiation effects in SONOS analog computing can be used to guide the design of reliable space-grade analog IMC accelerators.

@article{xiao2025radiation,
  author    = {T. Patrick Xiao and Maximilian Siath and Matthew Spear and Donald Wilson and Christopher H. Bennett and Ben Feinberg and David R. Hughart and Jereme Neuendank and William E. Brown and Hugh Barnaby and Vineet Agrawal and Helmut Puchner and Sapan Agarwal and Matthew J. Marinella},
  title     = {{In Situ Analog In-Memory Computing Under Ionizing Radiation Exposure}},
  journal   = {IEEE Transactions on Nuclear Science},
  year      = {2025},
  volume    = {72},
  number    = {4},
  pages     = {1243--1251},
  doi       = {10.1109/TNS.2025.3537985}
}