Prediction performance on each interpolation and extrapolation for corrosion rates with laws using machine learning

Corrosion problems affects widely such as strength reduction, collapse of objects due to damage, damage to neighborhoods due to falling damaged objects. In order to prevent accidents before they are happen, it is necessary to be able to predict limitation of durability loss due to corrosion and to accurately calculate maintenance cycles. So the correct prediction for corrosion change is important significance. We are attempting to predict corrosion by deep learning, machine learning and statistical regression methods. This research aims to reveal prediction performance for corrosion change using their methods, specially by comparing with approximation functions based on supervised learning and approximation functions following physical law’s change based on statistical method. In addition to that we reveal the difference of the predict score between interpolation and extrapolation. Because the prediction on interpolation and extrapolation shows completely difference. The result showed that the best prediction method on extrapolation was logarithmic approximation (Feature: Day, MSE: 4.8, $R^{2}: 0.7$), Neural Network(Feature: A-D, MSE: 35.1, $R^{2}: 1.19)$. The best method on interpolation was exponential approximation (Feature: Day, MSE: $19.7, R^{2}: 0.98$), Neural Network(Feature: AD, MSE: $.23.9, R^{2}: 0.97)$.