Neural Network Prediction - Intelligence
Examples of the intelligence added to the input layer of the NMRPredict Neural Network:
- It is well known that para-substituents in aromatic systems have a large influence
on the carbon shift. Within the HOSE code the para-substituent is 4 bonds
away from the focus atom and therefore has a low priority during the selection
of similar HOSE codes from the reference data collection. With the design of NMRPredict's Network
a dedicated network for aromatic carbons artificially moves all 6 possible
substituents on an aromtaic ring system into the first sphere, which reflects
the real-world situation much better than the HOSE code.
- When comparing the series of 'chloromethanes' starting at CH4 (Cl=0) and
going to CCL4 the increase of the chemical shift values is not linearly
coupled to the number of chlorines present. Therefore the number of chlorines
is not a good measure for the effect to be expected on the chemical shift value.
In other words, the pure number of a certain element has to be substituted
by a specific non-linear input value.
- A more complicated situation is present in a molecule like CHBr3 where the
electron withdrawing effect of bromine is severly overlapped by the
'heavy-atom' effect, leading to totally unexpected values.
The examples below give a small insight into the clever design of the
Neural Network within NMRPredict:
The larger difference between HOSE and Neural Network predicted values for I compounds
is based on the strong solvent dependency of this particular shiftvalue. The
Br/Cl/F substitued derivatives fit together within less than 1.5 ppm.