Most of our knowledge about the Universe comes from the careful analysis of light that reaches us. Spectroscopy, which is the most detailed method of spectrum analysis, when applied to stars provides information on the parameters of their atmospheres, including effective temperature, acceleration, velocity fields, and their chemical composition. Stellar classification brought forth the understanding of what physical parameters are critical in shaping stellar atmospheres. It is a key element that has linked efforts related to numerical modelling of atmospheres with observations. We present preliminary results on the method of stellar spectra classification based on large-scale unsupervised pre-training. The applied deep neural network of the auto-encoder type, thanks to the use of differentiable elements of physical modelling in the decoder, allows to work with medium to high-resolution spectra, is insensitive to normalization errors, and different radial and rotational velocity, and operates in a wide range of signal-to-noise ratio.