Group additivity needs improvement (Dah)

[alongd]

We should think of how to improve our groups.

Here's an interesting example: H298 of vinoxy radical (C=C[O]) is 18.5 kJ/mol according to the DFT_QCI_thermo library. If we ask GAV to estimate the thermo of each resonance structure, we get:

H298 = 18.101 kJ/mol S298 = 264.78 J/(mol*K) Thermo group additivity estimation: group(Cds-OdCsH) + group(Cs-(Cds-O2d)HHH) + radical(CJC=O)

H298 = -13.3136 kJ/mol S298 = 251.061 J/(mol*K) Thermo group additivity estimation: group(O2s-(Cds-Cd)H) + group(Cds-CdsHH) + group(Cds-CdsOsH) + radical(C=COJ)

A very minor deviation for [CH2]C=O, but a significant deviation for C=C[O]. However, since RMG chooses the thermo of most stable resonance structure (lowest H298), we get a very wrong value.

It seems that we're hitting the correct groups for C=C[O]. Also, the radical correction of C=COJ is in accordance to literature (https://doi.org/10.1002/chem.201301381).

We should understand what causes this large deviation.

[alongd]

Our GAV (and also the relevant literature) tries to estimate the thermo of a species (i.e., resonance hybrid) based on a given representative localized structure. I think we should add as many group corrections as we can for different resonance transformations (like the C=COJ rad group) to get better values.

Continuing the above example, the radical correction of C=C[O] doesn't seem to be the source of the problem. The HBI form C=CO has H298 of -124.9 kJ/mol according to the DFT_QCI_thermo library, but -153.4 kJ/mol according to GAV. This is the same ~30 kJ/mol deviation we saw above for the C=C[O] structure.

[alongd]

@laitcl, I understood you have some interesting CBS-QB3 thermo calcs for species with an allyl shift resonance and relatively large delocalization (something like [CH2]C=CC=C). Are these available online?

[cgrambow]

So we're actually planning to have a more extensive version of the neural net estimator fairly soon, which will be incorporated into the current PR before it is merged. I've actually been meaning to talk to you about that; will come see you tomorrow. Although regardless of that, if we have calculations available, we might as well add them, since both group additivity and neural net estimator will still be available, and neural net will most likely not be the default choice at first.

[laitcl]

Guys, I'm stuck in Hong Kong due to a Visa issue that everybody saw coming.

Short answer to your question, my new groups are in the branch NewGroups_Lai_03162018. I am still waiting to update a few of these numbers soon, but you could see my preliminary work there.

I expect a finalized version to be ready two weeks from now, with all my ongoing calculations turned in. If you want to talk to me, I could show you the data and the preliminary conclusions I made about the groups on Slack or some sort of voice/video communication.

[alongd]

Another case where GAV fails: It assigns 0 H298 and Cp's for CH3NO: https://rmg.mit.edu/database/thermo/molecule/1%20C%20u0%20p0%20c0%20%7B2,S%7D%20%7B4,S%7D%20%7B5,S%7D%20%7B6,S%7D%0A2%20N%20u0%20p1%20c0%20%7B1,S%7D%20%7B3,D%7D%0A3%20O%20u0%20p2%20c0%20%7B2,D%7D%0A4%20H%20u0%20p0%20c0%20%7B1,S%7D%0A5%20H%20u0%20p0%20c0%20%7B1,S%7D%0A6%20H%20u0%20p0%20c0%20%7B1,S%7D%0A