Convolutional Neural Network-Based Regression Model for Distribution Data from X-Ray Radiographs of Metallic Foams

Abstract

The difficult determination of morphological properties in metal foams stands behind the reasons why metal foams are not widely used in industry, since quality control of the batches produced is limited to destructive methods. To approach this challenge, a new method of analysis of morphological properties based on 2D X-Ray radiograms and the employment of a new Convolutional Neural Network architecture is proposed. The training of this model is based on a combined approach of simulating simplified foams as pretraining data and the acquisition of real experimental data, extracted from X-Ray computer tomographies. The network is trained successfully with 41 foams to obtain predictions for cell size distribution between 0.3 and 5 mm, as well as sphericities in ranges from 0.4 to 1. In addition, tests are carried out to get an insight into the robustness of the model when confronted with similar data that are not included in the training process. It is found that the effectiveness of the neural network increases with a larger number of cells in the observed volume where above 500 cells per volume 92.5% of sphericity predictions and 99.4% of cell size predictions passed the Kolmogorov-Smirnov test.