Chemical sample classification using autoencoder-based spectroscopy

We present a Proof of Concept of a chemical sample classifier. Instead of a spectrometer, we use autoencoders, machine learning models that try to reconstruct their output images presented at their input. The autoencoder is trained with spectrograms generated from optical signals detected after propagating through the sample. The signal shape is specifically designed, and the emitted wavelength varies as a function of time. Over propagation through the sample, the signal is partially absorbed. Since absorption depends on the sample's transmittance, the detected signal represents a signature. Hence, based on reconstruction error, the autoencoder performs classification with 100% accuracy.