## Wednesday, December 01, 2010

### Routine Prediction

Not everybody remembers how easy it is and how effective it can be to add (or correct) a few points at the beginning of the FID. Two years ago I explained how Linear Prediction works and how we can extrapolate the FID in both directions. This time I will show a simple practical application.
I have observed that, in recent years, C-13 spectra acquired on Varian instruments require a much-larger-then-it-used-to-be phase correction. When I say correction, I mean first-order phase correction, because the zero-order correction is merely a different way of representing the same thing (a different perspective).
A large first-order phase correction can be substituted with linear prediction. I will show the advantage with a F-19 example, yet the concept is general.

The spectrum, after FT and before phase correction, looks well acquired. Now we apply the needed correction, which amounts to no less than 1073 degrees.

Have you noticed what happened to the baseline? It's all predictable. When you increase the phase correction, the baseline starts rolling. The higher the phase correction, the shorter the waves. With modern method for correcting the baseline we eliminate all the waves, yet there are two potential problems: 1) The common methods for automatic phase correction will have an hard time. 2) If you prefer manual phase correction, you need an expert eye to assess the symmetry of the peaks over such a rolling baseline. Anyway, just to show you that linear prediciton is not a necessity, here is the spectrum after applying standard baseline correction:

Now let's start from the FID again, this time applying linear prediction. one way to use it is to add the 3 missing points at the beginning. The result, after a mild phase correction (<10°) and before the baseline correction is:

The lesson is: by adding the missing points we correct the phase.
Alternatively we can both add the 3 missing points and recalculate the next 4 points. In this way the baseline improves a lot:

The lesson is: by recalculating the first few points of the FID we can correct the baseline of the frequency domain spectrum.

At 5:53 PM,  Anonymous said...

Hi, why there are two humps in the processed spectrum? I cann't see them in the original one.
Sergey

At 6:01 PM,  old swan said...

Neither do I (in the first picture), but the humps are clearly visible in the second picture (before baseline correction). Also consider that the intensity has been enhanced in the last picture, as you can see by the higher noise level. If you click the title of the post and follow the links you can download the original data set and reprocess it by yourself.

At 7:44 PM,  old swan said...

I have rewritten the article. Linear prediction has been applied in two different ways.

At 9:59 AM,  Anonymous said...

"I have observed that, in recent years, C-13 spectra acquired on Varian instruments require a much-larger-then-it-used-to-be phase correction."

I am not sure what that means.The latest Varians do predict first few points 'automatically' - no first-order corrections should be needed at all.

At 10:06 AM,  old swan said...

"The latest Varians do predict first few points 'automatically".
You are confirming that something has changed (compared to 20 years ago). If Varian software predicts the missing points, it does it on the fly, because when you open the files with another software those points are still missing.
Different is the case with the now ubiquitous digital filters: the saved FIDs are already filtered.

At 10:12 AM,  Anonymous said...

"Different is the case with the now ubiquitous digital filters: the saved FIDs are already filtered."

This is what I mean - you have to do something special to switch this feature off.

At 10:21 AM,  old swan said...

Now you are really cryptic. I don't get what you mean.
Anyway, the original Varian file, as it arrived to me, says: "lp = -1049.58420782". How do you explain this?

At 10:54 AM,  Anonymous said...

First of all I am talking about VNMRS. On older system it can be different. The 'dead' time of the reciever is determined by two parameters: 'rof2' and 'alpha'.
rof2 is a receiver gating time following pulse, alfa is a delay before acquisition. There is a digital reciever parameter 'ddrtc' which specify the time of the backwards prediction. I've been told it's not really a linear prediction but some fancy algorithm executed on DSP chip before data is transfered from the console to the PC. But the idea is similar. The value of ddrtc must be set. Usually it's done automatically when variable ddrpm is set. There is a description of ddrpm from the Varian manual:

ddrpm Set ddr precession mode (P)

Applicability: VNMRS systems

Syntax: ddrpm=<'mode’>
Values: mode can be either of following:

p Pulse — default if no argument is supplied.
The value is calculated as follows if ddrpm does not exist or ddrpm='p':
ddrtc = alfa + rof2 + 2 * pw[1] / Pi
e Echo — The value is calculated as follows: ddrtc = alfa.

The ddrpm should be set correctly for standard Varian experiments.

So, if everything is correct, your lp should be zero. On other hand, it's not impossible to break the correction in custom experiments/pulse sequences.