Study of correlation between the steels susceptibility to hydrogen embrittlement and hydrogen thermal desorption spectroscopy using artificial neural network

Abstract

Steels are the most used structural material in the world, and hydrogen content and localization within the microstructure play an important role in its properties, namely inducing some level of embrittlement. The characterization of the steels susceptibility to hydrogen embrittlement (HE) is a complex task requiring always a broad and multidisciplinary approach. The target of the present work is to introduce the artificial neural network (ANN) computing system to predict the hydrogen-induced mechanical properties degradation using the hydrogen thermal desorption spectroscopy (TDS) data of the studied steel. Hydrogen sensitivity parameter (HSP) calculated from the reduction of elongation to fracture caused by hydrogen was linked to the corresponding hydrogen thermal desorption spectra measured for austenitic, ferritic, and ferritic-martensitic steel grades. Correlation between the TDS input data and HSP output data was studied using two ANN models. A correlation of 98% was obtained between the experimentally measured HSP values and HSP values predicted using the developed densely connected layers ANN model. The performance of the developed ANN models is good even for never-before-seen steels. The ANN-coupled system based on the TDS is a powerful tool in steels characterization especially in the analysis of the steels susceptibility to HE.

Evaluation of Steels Susceptibility to Hydrogen Embrittlement: A Thermal Desorption Spectroscopy-Based Approach Coupled with Artificial Neural Network

“…MAE of the HSP prediction using the developed ANN model at the test dataset was calculated to be about 1.4%. The advantage of the proposed measurement data processing is obvious considering our previous study, where MAE was calculated to be about 2.8% and 4.5% for feed-forward and convolutional neural network architectures, respectively [15]. The detailed study of the prediction error calculated at the test dataset for individual steels showed some increase of MAE calculated for steels included in the validation compared to those used in the model learning process (see Figure 7).…”

“…where x i is the descriptor parameter of the artificial spectroscopy data. F(x i ) is a linear function defined by the fitting of the relationship between x i = [1,3] and x i = [4,15] . It is worth noting that the Pearson correlation coefficients between only the same type of descriptor parameters defining the peak height, temperature position, and width were considered.…”

“…Deep feedforward architecture was considered for the development of the regression ANN model [15]. The architecture of the developed model is schematically shown in Figure 3.…”

“…With the use of recent advances in machine learning technologies, we have developed and published a first conceptual approach in the prediction of HE susceptibility of steels correlating the macroscopic mechanical property changes, caused by hydrogen, with hydrogen thermal desorption spectroscopy (TDS) of steels [15]. TDS spectroscopy provides indirect assessment of the material microstructural features controlling the hydrogen diffusivity, solubility, and trapping of hydrogen in the studied steel.…”

“…The parameters of stress state and environmental conditions controlling the hydrogen uptake were defined as constants at this particular stage of the research (see Figure 1). The approach was tested with different ANN model architectures evidencing a strong correlation between TDS of hydrogen and the target value of the hydrogen sensitivity parameter (HSP) standing as a measure of HE susceptibility of steels [15]. The objective of the present paper is to discuss the benefits of the approach and introduce a novel prototype software application for prediction of HSP of steels based on the recently developed ANN model.…”