Prediction of Pitting Corrosion in Aqueous Environments via Artifical Neural Network Analysis

摘要

One mission of the Department of Energy's Savannah River Site (SRS), near Aiken, SC, is to store spent nuclear fuel (SNF) and other waste products while permanent storage facilities for such materials are prepared. This extended storage increases the probability of pitting corrosion for both aluminum-based SNF stored in natural (fresh) waters, and in carbon steel waste tanks containing aqueous radioactive waste. The Back Propagation of Error method was used to train and test an Artificial Neural Network (ANN) model using archival pitting data. For aluminum, a database from the British Non- Ferrous Metals Research Association (BNFMRA) was used because it contained the relevant chemical species for pitting. A trained ANN, containing two hidden layers of six and four elements each, provided a good estimate of pit depth as a function of water chemistry after 150000 training cycles. For carbon steel, a set of 'noisy' waste chemistry versus binary pitting state data generated at SRS was used to train and evaluate a feed-forward ANN model. The model contained two hidden layers of five and three elements each, and after 100000 training cycles, successfully predicted the correct pitting state for the carbon-steel tanks 69 percent of the time.