Fix tree averaging issue #712

[KEHANG]

This PR tries to fix the issue #712 by

• generating all combinations of cross-level rules
• when calculating the estimated rules, the averaging scheme is only choose distance-1 rules

[mention-bot]

@KEHANG, thanks for your PR! By analyzing the annotation information on this pull request, we identified @jwallen, @connie and @rwest to be potential reviewers

[rwest]

No time to look into it further myself now, but I think this is being killed because it's run out of memory on travis, not because it's run out of time.

[connie]

In addition to kehang's notes on this pull request, I have also updated:

• The storage of kinetics comments for averaged rules is streamlined. Previous storage of fully recursive kinetics comments into the rule entries that are averaged during fillRulesByAveragingUp spent more than 5 GB memory during RMG runs. Therefore, I have simplified it to listing the distance 1 children that contribute to kinetics only. In order to find the actual sources of the original kinetics, one would need to look into the code further. This means that verbose comments are no longer as verbose as before. To turn back on the truly recursive comments, you can go into the code to uncomment a few lines.
• Euclidean distance algorithm implemented for the selection of the best kinetics to use in estimateKinetics. Previously, we went with the Manhattan distance, which meant an entry that was a distance of (3,0) away was equally favorable as an entry of distance (2,1) away. Now, switching to a euclidean distance norm we will choose the (2,1) entry over the (3,0) entry. This is because we saw in the models that higher top level nodes were becoming increasingly incorporated. I think we want the algorithm to prevent top level nodes from appearing as infrequently as possible.

+ = +

C2H3(13) + C2H3(13) ⇔ C2H4(8) + C#C(25)

Before:

! Template reaction: Disproportionation
! Flux pairs: C2H3(13), C2H4(8); C2H3(13), C#C(25); 
! Estimated using average of templates (Y_rad_birad_trirad_quadrad;Cd_Cdrad) + (Y_rad;CH_d_Rrad) + (Cd_pri_rad;XH_Rrad) for rate rule
! (Cd_pri_rad;Cd_Cdrad)
! Multiplied by reaction path degeneracy 2
C2H3(13)+C2H3(13)=C2H4(8)+C#C(25)                   6.005e+11 0.167     1.127    

After:

! Template reaction: Disproportionation
! Flux pairs: C2H3(13), C2H4(8); C2H3(13), C#C(25); 
! Estimated using template (Y_rad;CH_d_Rrad) for rate rule (Cd_pri_rad;Cd_Cdrad)
! Multiplied by reaction path degeneracy 2
C2H3(13)+C2H3(13)=C#C(25)+C2H4(8)                   6.447e+06 1.902     -1.131   

Although this pull request increases the time required to run the minimal example, we believe that the kinetics estimation is in fact better than it was previously. Running the RMG-database/testing/evaluateKinetics.py may help elucidate how the averaging works as well: ReactionMechanismGenerator/RMG-database#117