Development of a machine-learning approach for predicting the corrosion rate of steel in aqueous electrolytes

Abstract

The object of research is the corrosion of St20 steel in aqueous electrolytes of different salt compositions under aerated and deoxygenated conditions. The problem addressed is the development and comparative evaluation of corrosion rate prediction models based on a limited experimental dataset, with the possibility of interpreting the influence of individual factors. The effects of exposure time, NaCl, Na2SO4, and Na2CO3 concentrations, and aeration and deoxygenation conditions on the corrosion rate were investigated experimentally. Three models were considered: a physics-informed LSTM model, an LSTM model with an attention mechanism, and a Random Forest model. The training dataset was formed by pseudo-temporal expansion of the experimental data using a sliding window. This approach transforms discrete measurements into sequences for model training without replacing experimental corrosion rate determination with a purely computational procedure. The models generally reproduce the experimental dependencies within the dataset. The R2 values for the physics-informed LSTM model, the LSTM model with the attention mechanism, and the Random Forest model were 0.9926, 0.9925, and approximately 1. The physics-informed LSTM model accounts for differences between aerated and deoxygenated conditions, which are important for describing corrosion processes in aqueous media. The attention mechanism provides model interpretability by estimating the contribution of individual factors to the prediction. The Random Forest model was used for comparison with neural network approaches and to assess the reproduction of dependencies in the prepared experimental data. The results confirm the feasibility of combining experimental data and machine learning methods to model complex nonlinear corrosion processes. The approach can be used for preliminary assessment of the corrosion activity of aqueous media and for planning further laboratory and industrial researches.