Discovering key meteorological variables in atmospheric corrosion through an artificial neural network model

Abstract

This paper presents a deterministic model for the damage function of carbon steel, expressed in μm of corrosion penetration as a function of cumulated values of environmental variables. Instead of the traditional linear model, we designed an Artificial Neural Network (ANN) to fit the data. The ANN numerical model shows good results regarding goodness of fit and residual distributions. It achieves a RMSE value of 0.8 μm and a R 2 of 0.9988 while the classical linear regression model produces 2.6 μm and 0.9805 respectively. Besides, F LOF for the ANN model were next to the critical value. The improved accuracy provides a chance to identify the most relevant variables of the problem. The procedure was to add/remove one after the other the variables and perform from scratch the corresponding training of the ANN. After some trial and error as well as phenomenological arguments, we were able to show that some popular meteorological variables like mean relative humidity and mean temperature shown no relevance while the results were clearly improved by including the hours with RH < 40%. The results as such might be valid for a limited geographical region, but the procedure is completely general and applicable to other regions.