Failed Jobs and Possible Integrator Problems in MadWeight

We have been using MadWeight to calculate the event probability densities for use in an EPD (Event Probability Discriminant) in the H->WW->lvlv + 2-jet channel MVA analysis. Our signal hypothesis is VBF decaying via this channel, and these calculations have been working perfectly (i.e. P_VBF calculated using the ME generation >> generate u d > h u d , h > e- ve~ mu+ vm /z and similarly for the other possible final state we are interested in mu- vm~ e+ ve). However, we have noticed that there are a significant number of failed jobs for our background hypothesis (P_ttbar generated using >> generate p p > t t~, (t > w+ b, w+ > e+ ve), (t~ > w- b~, w- > mu- vm~)). We suspect this might have to do with integration settings; however, we are at a bit of a loss on how to diagnose what exactly needs to be done with the integrator (or how to verify if in fact it is the integrator). It is also important to note that about 25% of these jobs have also failed P_WW and P_ggf calculations in our own standalone that uses the Divonne integrator instead of Vegas; this has cast a little bit of doubt onto whether it is truly the integrator giving us trouble.

***Kinematic distributions removed as they are ATLAS internal*** ; as you will see, it appears as though most of these failed events are highly boosted (low DPhill and higher Njets). Possibly some indication that it may be a physical issue with the events themselves that are causing the ME calculations to not converge?

Thank you for your time, and please let me know if you need any more information in order to help in your diagnosis.