CCL: The use of NMR spectra computations for verification of computational method



 Sent to CCL by: Susi Lehtola [susi.lehtola()alumni.helsinki.fi]
 Dear Grigoriy,
 if you are asking something, please formulate your message using
 paragraphs. A wall of text is hard to read, and most people will just
 skip over it.
 On 09/21/2015 10:19 AM, Grigoriy Zhurko reg_zhurko^chemcraftprog.com wrote:
 
 The PMR spectrum with B3LYP correlates with the experimental one with
 R=0.997, and with PBE ? R=0.995. These values do not differ very
 much. But the coefficient B in the equation Y=A+B*X (for the linear
 approximation of experiment vs theory graph) is 1.02 for B3LYP, and
 1.14 for PBE. So, with PBE it is far less from 1.
 
 So here you are talking about how well the values reproduced by theory
 agree with experiment, allowing for systematic error in the theory that
 follows a linear model.
 Indeed, the R values are close to one, which indicates the linear model
 is good. The different B (and A?) values just indicate that the
 functionals produce different values for the shielding constants, but as
 both R values are good, you should be able to easily convert a PBE value
 to B3LYP or vice versa.
 
 Does that mean that PBE is much less appropriate method for this
 task?
 
 I'm not an expert on NMR. However, judging from your data, I would say
 that the answer is no. Naturally if your theory reproduced
 experiment exactly you'd have A=0 and B=1. Here, both functionals give
 somewhat incorrect values, since A!=0 and B!=1. However, the error in
 the values seems to be systematical, which you can correct for by
 solving x from y.
 
 I suppose, that the systematic error of absolute values of the NMR
 chemical shifts is caused by unclear physical meaning of these
 chemical shifts and shieldings (maybe the solvent gives some
 additional shielding in experiment).
 
 Naturally, effects that aren't included in your modeling will affect the
 goodness of your model. In general, you might see this as unexpected
 values of A and B, or by a breakdown of your model (small R).
 
 So, my question is, whether the B coefficient in correlation must be
 always equal 1.
 
 No.
 
 If yes, then instead of correlation coefficients R I should use
 another criteria of computation accuracy?
 
 If B were always equal to 1 you wouldn't really need R at all. R is a
 measure of how well y = a + bx fits your data.
 
 the RMS of MAE difference between the computed and experimental
 chemical shifts, if the shielding of the standard (TMS) simply
 fitted for best agreement (not computed quantum-chemically). Is that
 correct?
 
 What?
 
 In my case, these MAE difference must be much bigger for PBE
 functional, than for the B3LYP functional.
 
 
If you took the uncorrected values, sure. But if you applied the correction for the systematical error, both should give similarly accurate results.
 --
 -----------------------------------------------------------------------
 Mr. Susi Lehtola, PhD             Chemist Postdoctoral Fellow
 susi.lehtola : alumni.helsinki.fi   Lawrence Berkeley National Laboratory
 http://www.helsinki.fi/~jzlehtol  USA
 -----------------------------------------------------------------------