Re: refinement with two phases

[Pavol Juhas]

On Wed, Jun 7, 2017 at 1:22 AM, ANUPAM KUMAR SINGH School of Materials
Sci. & Tech., IIT (BHU) Varanasi <anupamks...@itbhu.ac.in>
wrote:
> Dear sir,
> I am unable to find out how to get phase fraction of two phase system after
> the refinement through PDFgui.
> It will be very helpful to me if you could suggest me something about it.

Dear Anupam,

Please, see the Results tab after clicking the top fit-node of your
two-phase refinement.
Scroll to the very end and it will report the phase fractions in terms of atoms,
unit cells and mass ratios.

Hope this helps,

Pavol

[ANUPAM KUMAR SINGH School of Materials Sci. & Tech]

Dear sir,

 I am working on two phase system PDFgui refinement.It will be very helpful to me know that once first phase refinement is done then after that when I insert second phase and start to refine it then while refining the second phase parameter should I refine  the first phase parameter also or not? I mean, can we refine the both phase simultaneous or first refine the first phase and then second phase while fixing the first phase? 

---

with best regards

 Anupam

Anupam Kumar Singh

Ph.D student

[Chenyang Shi]

Hi Anupam,

In a two phase fit, the fit equation is something like following:

G(r)_total = scale_1 * G(r)_phase_1 + (1 - scale_2)*Gr_phase_2

For each G(r)_phase_1 or G(r)_phase_2, you will have a set of parameters to simulate it (such as lattice parameters, atomic positions or atomic displacement parameters/ADPs). The end goal is to minimize the error between calculated and measured PDFs. 

So if were you, I would start to fit two phases simultaneously. If there are too many parameters, you can leave atomic positions or ADP untouched at default values. Once the first attempt using a two-phase fit captures the overall features, you can then start to fine tune other parameters. It will converge eventually. 

Before performing a two phase fit, I would suggest take a look at peak features in each simulated phase: does the peak position ROUGHLY match experimental peak? Is there any chance that a combination of these two PDFs can match the measured one? Visually can you see the simulated peak positions off by a large margin. If so, you might consider replacing the structural model for one of the phase. 

Make sure the resolution parameters (Qdamp/Qbroad) are fixed for each phase. 

PDFgui uses a downhill optimizer which requires roughly correct starting parameters; if after many attempts, you cannot get a good converge, you may need to reconsider your models.

Hope this helps.

Chenyang

[ANUPAM KUMAR SINGH School of Materials Sci. & Tech]

Dear sir, thank you so much for your valuable reply.It really helps me a lot.Sir, one more thing I want to know that what is the meaning of NaN and the value written in bracket in phase fraction for example:

atoms                             unit cells                            mass  Phase 1 :0.709432 (NaN)          0.785513 (NaN)          0.709432 (NaN)

 Phase 2 :0.290568 (NaN)          0.214487 (NaN)          0.290568 (NaN)

 

atoms                       unit cells                     mass  Phase 1 :0.826508 (5.3e+004)     0.877239 (4e+004)       0.826508 (5.3e+004)

 Phase 2 :0.173492 (5.3e+004)     0.122761 (4e+004)       0.173492 (5.3e+004)

And for two phase, could I refine the scale factor of data set and scale factor of phase? If I do, there will be total four scale factor for two phase. 

Please give your valuable suggestions.

with best regards

  Anuapm

[Pavol Juhas]

On Wed, Jun 21, 2017 at 11:08:06AM -0700, ANUPAM KUMAR SINGH wrote:
...

> Sir, one more thing I want to know that what is the meaning of NaN and
> the value written in bracket in phase fraction for example:
>

> atoms cells mass
>
> Phase 1 :0.709432 (NaN) 513 (NaN) 0.709432 (NaN)
> Phase 2 :0.290568 (NaN) 487 (NaN) 0.290568 (NaN)

Hi Anupam,

The values in parentheses are estimated standard deviations (ESD)
on the fitted values. NaN stands for Not-a-Number and is a result
of some undefined math operation, for example, division by zero.
Getting NaN-s or excessively large values for your ESD-s means
there is some instability in your fit; this usually happens
if some of the fitted parameters have similar or
linearly-dependent effects on the simulation.
A simple example would be fitting both A, B in the
"A + B = 7" model.

> Phase 1 :0.826508 (5.3e+004)

Also note that ESD-s are correctly scaled only if your data
provide standard deviations for the observed G(r) curve.
Even if such dG values are absent, the ESD-s are usually
smaller or comparable to the associated results. Getting
ESD in the order of 1e4 suggests that something is
incorrect in the fit.

...

> And for two phase, could I refine the scale factor of data set and
> scale factor of phase? If I do, there will be total four scale
> factor for two phase.

PDFgui uses the same phase-scale factors for each dataset, so
that would be at most 3 scale factors. The data and phase scale
factors are multiplied in the PDF simulation so you need to make
sure to refine only as many as meaningful. The last email had
a bit of a typo in the formula, I think it should have read

G_total = dscale * (pscale_1 * G_1 + (1 - pscale_1) * G_2)

If you are getting troubles with pscale_1 converging to values
larger than 1 you can effectively setup the same fit by fixing
data-scale at 1 and fitting the 2 phase-scale factors instead:

G_total = pscale_1 * G_1 + pscale_2 * G_2

With this latter form, you can constrain pscale_1, pscale_2
as a square of some parameter (e.g., "@5 * @5") and thus prevent
them from reaching negative values.

If you have 2 datasets, the first form is physically more
descriptive. In such case you should refine dscale_1, pscale_1
and dscale_2. Note that fitting phase ratios for 2 datasets
works well only if both of those PDFs are of the same kind,
i.e., both are X-rays or both neutron PDFs.
For a mixed X-ray & PDF refinement the ratios of pscale_1,
pscale_2 should be different for each dataset, but this is
not supported in PDFgui. Such refinement can be correctly
done with DiffPy-CMI, as the phase scales there can be
different per each fitted dataset.

Hope this helps,

Pavol

--
Dr. Pavol Juhas
Computational Science Initiative
Brookhaven National Laboratory
P.O. Box 5000
Upton, NY 11973-5000

[ANUPAM KUMAR SINGH School of Materials Sci. & Tech., IIT (BHU) Varanasi]

Dear sir,

Thank you so much for your kind attention.It is really important for me.

with kind regards

  Anupam

[ANUPAM KUMAR SINGH School of Materials Sci. & Tech]

Hello sir,

I am working in the field of alloy. Recently I am trying to refine my PDF data by PDFgui with two phase as the way suggested by you. All the parameters looks good after refinements except the phase fractions of both phases. Please give your valuable suggestions to resolve my problem.

with kind regards,

Anupam