Gaia24azi investigation

[rachel3834]

Why was Gaia24azi not flagged for observation earlier?

[rachel3834]

At first glance it appears as though the modeling for this event hit a bug: the last_fit date given was suspiciously old (2446756.5) and the fitted values (e.g. tE, u0, t0) were nans. The classification given was Unclassified poor fit with category Unclassified. This may be an accurate result but causes MOP to stop evaluating the event as a possible microlens, even if more data arrive.

So my working hypothesis is that for some reason, the early modeling of this event resulted in an erroneous fit for some reason. The fitting process itself (fit_need_events_PSPL) doesn't reevaluate the classification but would have stored the resulting erroneous fit parameters.

After that the gaia_classifier would have picked up the event. This uses the following statement to select events with valid fits to evaluate:

if mulens.extras['u0'].value != 0.0 \
                        and mulens.extras['t0'].value != 0.0 \
                        and mulens.extras['tE'].value != 0.0 \
                        and event.ra != None and event.dec != None:

All of these conditions evaluate to True for an event with nan entries for the model parameters. So this event would have passed this test.

At this point, the gaia_classifier evaluates the blend_mag, u0, dmag and chisq of the fit to judge whether or not the fit is reliable. Unfortunately, I didn't record what values these parameters had before I re-fitted the event. However, the outcome was that it was classified as 'Unclassified poor fit', which suggests that the red_chi2 value may also have been nan, causing it to fail this test, and be classified as 'unclassified'.

The drawback to this classification is that MOP will no longer evaluate the event as potentially microlensing if its marked as 'unclassified'.

Furthermore, all of the available data that we have suggests that the event was fainter than 18mag. This would have ruled out this event from observations regardless of its classification.

[rachel3834]

So, what can we do to improve our handling of cases like this?

• [x] The selection clause identified above in Gaia_classifier should catch events with nan values for the model parameters as well as the default zero values. This would have caused MOP to leave the classification as (potential) microlensing, and would have meant that it would have been remodeled later on, perhaps with more data and a more robust fit.
• [ ] The cause that tests for a valid red_chi2 could also be removed to stop such an event from being set to 'unclassified' and removed from MOP's automatic procedures. However, I am less convinced this step is necessary as this clause was introduced to weed out an excessive number of variables.

[rachel3834]

I've implemented and deployed better handling for invalid fit parameters as described above.