CNN for training and picking. (a) Three-convolutional-layer CNN architecture. The three Conv2D layers use the indicated number of filters in parentheses. After each Conv2D, the spatial dimensions of the filters are reduced by a factor of two through the pooling process. Two dense layers are used to classify candidate particles. (b)–(e) Training results for the six EMPIAR data sets demonstrate the convergence of the workflow. (b) Training accuracy. (c) Validation accuracy. (d) Training loss function. (e) Validation loss function.