17:15 〜 18:45
[MGI26-P02] Disentanglement of the features in Rankine vortices using VAE
キーワード：深層学習、生成モデル、変分オートエンコーダ、台風、機械学習、人工データ
In recent years, Japan has faced a significant increase in human casualties and economic losses due to intense typhoons. Developing more accurate numerical weather predictions has become an urgent task for mitigating such typhoon-derived disasters. For that purpose, accurate atmospheric state estimates based on data assimilation are essential to improve subsequent numerical weather predictions. However, it is still challenging to implement highly accurate data assimilation because there is spatiotemporally insufficient observational data, such as central pressure and maximum wind speed. One effective method to solve this challenge is estimating atmospheric states from sparse observations using machine learning. Here, we focus on Variational Autoencoder (VAE), which has been developed in the deep learning field to generate pseudo data from latent space such as for drawing fake images. Since the VAE model could be suitable for creating artificial data, we aim at estimating atmospheric states from limited observation data. If we obtain the appropriate latent space, we can reconstruct atmospheric states based on VAE.
This paper aims to train the VAE for the Rankine vortex structures and disentangle the structures into four latent-space variables. A Rankine vortex is an idealized vortex model in fluid dynamics, and it is characterized by four parameters: the center coordinates Cx and Cy , the maximum radius R, and flow velocity U. The VAE consists of two parts: an encoder and a decoder. The encoder compresses input data into latent-space variables, learning to fit its probability distribution into a standard normal distribution. The decoder is trained to generate reconstructed data from latent variables that resembles the input data. The disentangled vortex features generated by the encoder could be used as input for the decoder to generate artificial vortices. The encoder part of the VAE consists of four convolutional layers and two linear layers. Also, the decoder part consists of two linear layers and four transposed convolutional layers. Batch Normalizations were added to each layer. We set the latent space dimension to be four because the four parameters can define a Rankine vortex. Here, we expected that each dimension could represent each parameter independently. The VAE is trained with Mean Square Error (MSE) for the reconstruction error of the loss function as well as Kullback-Leibler (KL) divergence as the regularization term.
The data for training and testing was generated as follows: 48,600 Rankine vortices on 91x91 grids were generated with R sampled from a Gaussian distribution (GD) with a mean of 12 and a standard deviation of 3 , U sampled from a GD with a mean of 30 and a standard deviation of 3 , and Cx and Cy sampled from a GD with a mean of 45 and a standard deviation of R (i.e., extracted radius). Training and testing processes used 48,000 and 600 samples, respectively. In the testing process, the latent-space variables of the test data were computed using the encoder, and their correlation coefficients with the true four parameters were calculated.
As a result, we succeeded in expressing Cx and Cy clearly as single variables. Two of the dimensions of the latent space had a correlation coefficient of -0.91 with Cx and 0.88 with Cy respectively. However, getting the latent-space variables for R and U as a single parameter turned out to be difficult. Although we did not successfully isolate latent-space variables of R and U, we obtained latent-space variables representing Cx and Cy . This result suggests that the VAE was trained to get features mainly of the central position because it is the domestic feature of vortices.
This paper aims to train the VAE for the Rankine vortex structures and disentangle the structures into four latent-space variables. A Rankine vortex is an idealized vortex model in fluid dynamics, and it is characterized by four parameters: the center coordinates Cx and Cy , the maximum radius R, and flow velocity U. The VAE consists of two parts: an encoder and a decoder. The encoder compresses input data into latent-space variables, learning to fit its probability distribution into a standard normal distribution. The decoder is trained to generate reconstructed data from latent variables that resembles the input data. The disentangled vortex features generated by the encoder could be used as input for the decoder to generate artificial vortices. The encoder part of the VAE consists of four convolutional layers and two linear layers. Also, the decoder part consists of two linear layers and four transposed convolutional layers. Batch Normalizations were added to each layer. We set the latent space dimension to be four because the four parameters can define a Rankine vortex. Here, we expected that each dimension could represent each parameter independently. The VAE is trained with Mean Square Error (MSE) for the reconstruction error of the loss function as well as Kullback-Leibler (KL) divergence as the regularization term.
The data for training and testing was generated as follows: 48,600 Rankine vortices on 91x91 grids were generated with R sampled from a Gaussian distribution (GD) with a mean of 12 and a standard deviation of 3 , U sampled from a GD with a mean of 30 and a standard deviation of 3 , and Cx and Cy sampled from a GD with a mean of 45 and a standard deviation of R (i.e., extracted radius). Training and testing processes used 48,000 and 600 samples, respectively. In the testing process, the latent-space variables of the test data were computed using the encoder, and their correlation coefficients with the true four parameters were calculated.
As a result, we succeeded in expressing Cx and Cy clearly as single variables. Two of the dimensions of the latent space had a correlation coefficient of -0.91 with Cx and 0.88 with Cy respectively. However, getting the latent-space variables for R and U as a single parameter turned out to be difficult. Although we did not successfully isolate latent-space variables of R and U, we obtained latent-space variables representing Cx and Cy . This result suggests that the VAE was trained to get features mainly of the central position because it is the domestic feature of vortices.