journal menu
![[Figure 2]](yr5164fig2mag.jpg)
|
|
Figure 2
The CNN architecture. The diagram simplifies five convolutional layers when extracting features from the augmented simulated XRD spectra, via convolution, kernel operations, ReLU + BatchNorm and average pooling. The architecture includes one-dimensional adaptive average pooling to produce fixed-length features, followed by flattening to a one-dimensional input. The full model contains two shared fully connected layers with batch normalization, ReLU activation and dropout regularization, though these are not all shown for brevity. Following the shared layers, the architecture branches into separate unshared fully connected layers, where each parameter has its own dedicated fully connected layer. |
![[Figure 2]](yr5164fig2.jpg)
 | JOURNAL OF APPLIED CRYSTALLOGRAPHY |
ISSN: 1600-5767
access
