Anomaly detection

We have trained convolutional autoencoders to on 2D temperature maps from simulations with fixed cosmology and astrophysics. 
We then pass to the autoencoder temperature fields from simulations with different cosmologies and astrophysics as those it has been trained on. 
The autoencoder is able to reconstruct these images with the same accuracy as the images it was trained on. This is illustrated in the left image, where the left panels display the original temperature images while the right panels show the reconstructed ones.
We then pass to the autoencoder images whose structure completely differ from the ones it was trained on: the CAMELS logo. The original and the reconstructed image is shown below. 
Although the reconstructed image looks really good, the error associated to the reconstructed image is (with the exception of the background) much larger than the error associated to any temperature image. Thus, the autoencoder has identified the camel body and the letters as an anomaly with respect to the temperature fields.