hot stars for SDSS-V (addendum 2)

Following up on the original post, I did the following variant:

The goal is to find massive stars, essentially by their two characteristics:
 they are luminous (abs(K)<-1, and they are hot, i.e. blue in unreddened colors.
The selection here is focussed on stars with some Gaia detection; even if the parallax
measurement is seemingly "marginal", it is powerful at ruling out all the low velocity stars.
The remaining challenge then is to separate luminous giant from luminous hot stars in the the presence of severe reddening. This can be done ad-hoc by devising self-dereddening colors, or (YST has done that with me) by taking BP,G,RP,J,H,K,W1 for stars with Apogee T_eff and training a neural net to predict T_eff in the presence of severe reddening; this seems to work beautifully, except that the training set is limited to 3500K<T_eff<8000K; this can be overcome.

Upshot of all of this: HWR's  naive view is that this works well (completeness and purity), as long as there is some Gaia information. Clean selection of massive stars that are H<12 or H<13, yet undetected in Gaia (say, G>19) remains a challenge, even conceptually.

Inspired by the hot-cool star separation (previous post, among absK<0),

I ran the query:

SELECT  g.*, tm.*, sqrt(g.phot_g_n_obs)/g.phot_g_mean_flux_over_error as variability
FROM gaiadr2.gaia_source AS g
INNER JOIN gaiadr1.tmass_best_neighbour AS xmatch
 ON g.source_id = xmatch.source_id
INNER JOIN gaiadr1.tmass_original_valid AS tm
 ON tm.tmass_oid = xmatch.tmass_oid
WHERE
tm.h_m < 12.
and
g.phot_g_mean_mag < 18.
and
( g.bp_rp < 1.  or  ( tm.j_m - tm.ks_m - 0.25*(g.phot_g_mean_mag - tm.ks_m) < 0.) )
and 
g.bp_rp < 2.
and
tm.j_m - tm.ks_m - 0.25*(g.phot_g_mean_mag - tm.ks_m) > -0.2
and
parallax < power(10.,(10.-tm.ks_m-1.)/5.)

which yielded:

I made a cut at bp_rp < 2, to cut out reddened variable sources.

Then I took the "massive stars regime"  at

then we get on sky for 40.000 objects:

those with 3sig parallaxes are:

Aside? Should I be worried that I don's see Orion?

Now let me just show what the plot looks like with the Teff estimates, derived from
BP,G,RP,J,H,K,W1 (trained to predict APOGEE Teff; courtesy Yuan-Sen Ting)

Here is a plot of NN-inferred (from colors) T_eff in the color-color plane:

The limiting (maximal) Teff is a consequence of the limited training set. The on-sky distibution of stars >6000K (16.000) of them at H<11.5 is here: (not bad?)

Addendum: November 2019

After the initial cuts, if one j_m - h_m - 0.13*(phot_g_mean_mag-ks_m) < 0.06, i.e.
makes another cut on on of these self-dereddened colors, one gets very "clean" samples.


In position space, they look like this:

and color coded by their G-K color (reddening), like this: