Migrated from GitLab software/mapPDF#1, opened by Soham Banerjee (sbanerjee) on 2017-03-08. Assigned to chliu.

@chliu @akovyakh

The first type of reduced data maps that we need to make sure are easily constructed from the DB are pearson correlation maps. We sort of overlooked this because we focused on refined variables from the fits, but this functionality must exist for this kind of software to be useful. It already exists in ctPDF which has a more mature/specific implementation, but that code is not easy to extend.

Once the migration and upgrading of the code is done from here to mapPDF we should make sure we can accomplish a simple UC like: user wants to see where diffraction patterns are most different from the background, user selects a bkgd position as reference, user generates correlation maps vs. reference position (I(q) and G(r)). I think this should be easy to do, since we've essentially done the more advanced UCs, but we need to take a step back and think about what will happen for different kinds of grids/films/arrays. I was doing a bit of reading, and these kinds of correlation maps will be particularly useful for screening, and that is the direction I anticipate this technique/protocol will be most useful. Situations where chemists have no idea what they deposited, so we have no candidate structures, and the zeroth order analysis would be to tell them how many structurally different things they made, which will be "contrast" limited as a function of signal quality. Phrased differently, how many unique diffraction patterns did we measure for the full catalyst array? We have no criteria for giving that answer right now, because we unnecessarily limit our "contrast" by using agreement factors from candidate models. Well, it isn't unnecessary, but it is a different, potentially less ubiquitous UC.

Summary: Combinatorial arrays are useful for materials discovery, and less useful/interesting for materials characterization. So, the concept of screening is very important (sorta obvious). The machinery we are building is becoming too focused on characterization (mostly my fault), instead, we need to build screening methods first. I would guess that this is where tools like PCA might become powerful, but we should think about it carefully, and research what else has been done. These could also become diffpy independent methods to use in xpdAn.

Let's try and make sure we can generate pearson maps by next week. I can try and make some milestones later to make sure we keep pushing on this stuff. I'm assigning this to @chliu but let's all work on it.

Comments from GitLab

Soham Banerjee (sbanerjee) on 2017-03-08:

If we need to do a full overhaul of the existing code structure, either for optimizing speed (Tim, the nested arrays issue) or in anticipation of how this will work with databroker etc., I think we should do that now and don't think it will be wasted effort.