Fig. 9: Artificial neural network architecture. | npj Climate and Atmospheric Science

Fig. 9: Artificial neural network architecture.

From: Deriving ozone and PM pollution vertical profiles using lightweight, cost-effective sensors and deep learning

Fig. 9: Artificial neural network architecture.

The network is composed of an autoencoder with a prediction head over the latent space (green). The input is first masked by replacing random features with learnable parameters, signaling the network to ignore these values. The mask itself is concatenated to the input to yield an input of [B = 1024,2*fin = 2*165], which is then fed into the autoencoder. The hidden dimensions of the encoder and prediction head are 32, while the auxiliary decoder head has a dimension of 64. Each layer is separated by a LeakyReLU activation layer for non-linearity and a dropout of 0.164. The sizes and layers were optimized for the task of predicting ozone profiles.

Back to article page