Thanks for the quick response. I’ve been testing a classification model with some over the counter medication as a test and I’m optimistic with the results given a very poor model.
A detection threshold of even 1% would definitely prove to be useful for this use case. The multitude of combinations that exist out there will definitely be problematic (but hey, that’s why we have a community to help!), but a more immediate and feasible use case would be to test if given sample is pure. Any deviation from the well-tested model would throw an alert.
Have you guys made any more progress with returning a confidence threshold from a scan so a yes/no decisions can be made? In the meantime, this application could have an “advanced view” that could show the user what the “ideal” spectra for pure MDMA powder is (we could ignore capsules and pills for now) and let the them decide if they want to further test the sample (with a chemical test, for example).
1) Is a confidence % against a model from a scan even feasible?
2) Also, will there come a point where we can share our models with each other in the community? I’d imagine any such repositories would have rules in place needed to adhere to a strict standardized method of collecting data.