Prediction with Spectrograms

Question

Hi All,   With strong inspiration from prior work by Bill, Pete, and Emmanuel,  I've just put together a new script for creating spectrogram images.  Spectrograms are heatmaps computed via Fast Fourier Transform (FFT) on sliding windows across a series of spectra.

The primary idea is to use these images as inputs in Torch to predict properties of the spectra.   This appears to be a good complement / addition to spectral FDE workflows and is easy to set up.

A first example from the NMR alcohol data is attached--big thanks to Emmanuel for the full data, extracted from his excellent journal.    It works!  Example output:

Also attached is an initial JSL script for computing spectrograms.

Feedback welcome.

Smiling,

Russ

P.S. Planning to convene a "Machine Learning for Chemistry" interest group at Reunion next week--please attend if you are able.

spectrogram.jsl

3b - Mixture NMR Wide 1.jmp

Florian_Vogt · Answer

Thanks @russ_wolfinger !

For those interested I have a slightly updated the version which uses the "wrap" variable and saves a script to the main table for the data table and gb to investigate individual spectrograms.

spectrogram_rev.1.jsl

russ_wolfinger · Answer

Yes @Florian_Vogt , thanks for mentioning--I encountered this also.   To remedy:  File > Preferences > Graph Builder and increase Page Limit from 200 to some number >= the number of spectra.

Florian_Vogt · Answer

Hi Russ,

great work as always - Thanks!

I tested the script and it works nice, except the fact when using the "page" function in GB it seems to limit my display to only 200. This results in only 200 out of 219 spectrograms to be brought back into the original table.

It works fine by using :g as the "wrap" variable.

Not sure if this is a local setting for me only.

Best

Florian