0
$\begingroup$

I have a set of known frequency spectrum data for this set of chemical compounds. Then the unknown Y is the mixture of some of these compounds. The task is to determine what compounds are in this mixture and their strengths. For example, S1 is the unit spectrum of the first compound and S2 is the unit spectrum of the 2nd compound. If the mixture has A units of the first compound and B units of the 2nd compound, the resulting Spectrum = AS1 + BS2.

I have seen examples of multi-class multi-label classification models, and multi-outputs regression models. But I need a model that can do both. Any suggestions?

Improve this question
$\endgroup$
2
  • $\begingroup$ spectrum as image can be classified using any image classification method.. $\endgroup$
    – Nikos M.
    Commented Mar 15 at 10:05
  • 1
    $\begingroup$ Yes, but this task requires the model to predict the strength of "dog" and the strength of "cat". Common image classification models cannot do that. $\endgroup$
    – David293836
    Commented Mar 15 at 16:55

1 Answer 1

Reset to default
1
$\begingroup$

Since your process is additive (emission spectrums?), this can be solved with non-negative least squares. You could try ordinal least squares as well, but it may give you negative concentrations. I don't think you need any fancy neural networks or ML here.

The constrained formulation is $ \arg \min_x || Ax - y ||_2^2$, $x \ge 0$, you "encode" the known spectrum data to $A$ and the measured spectrum is $y$, then $x$ gives you the concentrations.

If you measured the absorption spectrum instead, you'd need a different formulation.

Improve this answer
$\endgroup$
4
  • $\begingroup$ There can be many compounds in the mix. The goal is to identify them and their strengths. Would this method work? $\endgroup$
    – David293836
    Commented Mar 15 at 16:58
  • $\begingroup$ Yes, A is a n x m matrix where n is the number of measured wavelengths and m is the number of known compounds. Vector x has the length m and y has n. x is zero (or close to zero) for compounds not present in the mix, and positive for existing ones. I'm not sure what would happen if there are unknown compounds present though. At least the prediction error would be large. $\endgroup$
    – NikoNyrh
    Commented Mar 15 at 17:07
  • $\begingroup$ To clarify, what do you mean by "the goal is to identify them" [the compounds]? Since we assume that the mixture consists only of known compounds, isn't it enough to determine their strengths (which may be zero for many of them, for a given sample)? For example there is no need to detect whether there are any unknown compounds in the mix. $\endgroup$
    – NikoNyrh
    Commented Mar 16 at 1:06
  • $\begingroup$ Good question. If there is presence of unknown compound(s), this ML/DS/AI module needs to alert such event. $\endgroup$
    – David293836
    Commented Mar 16 at 19:04

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .