Outline

GSAS Parameters & Controls

What to Refine when?  Definitions of terms

(Refinement Recipes)  Powder peak shapes
– Fitting functions
– Size vs. strain broadening
– Low angle asymmetry
– Bragg-Brentano issues
 GSAS control values
 Setting up a refinement
Brian H. Toby  Profile functions and parameter definitions
 Preferred orientation
 Reducing model complexity (constraints/restraints)
 What to refine when (refinement recipe)

Parameters/Controls/Constraints

Parameters:
To model a “crystal structure” with Rietveld, we must fit a large
number of “experimental” parameters in addition to the
“crystallographic” parameters.
Controls:
Controls dictate how the programs operate. Profile parameters
Constraints:
We do not always have enough data to fit all the parameters robustly.
If so, we must approach the minimum with care and sometimes
introduce constraints to reduce degrees of freedom.

3 4

1
 Profile Models Approaches to Profile Models

Rietveld refinement requires that the shape of diffraction peaks (profile) as Three different approaches to reproducing peak shapes have been used:
well as their intensity be reproduced.  Empirical functions
Functions are chosen simply because they are computationally simple
Intensities: determined by structural parameters and fit peaks well. The parameters cannot be interpreted because
they have no relationship to the underlying physics of diffraction.
Profile: determined by instrument & by sample  Physically-based parameters
– Instrument response function Functions are based on the physical phenomena. Parameters are
– Residual stress usually found empirically, but often have a physical meaning.
– Crystallite (Scherer) broadening  “Fundamental Parameters”
Functions and where possible parameter values are determined from
diffraction physics. The only adjustable parameters are those related
to sample properties.

5 6

The physics that determine peak profiles Lorentz (Cauchy) and Gaussian Broadening Functions

Common factors Less common factors Most instrument & sample broadening contributions are Lorentzian or Gaussian
 Instrumental Resolution  Stacking faults*
Normalized Gaussian
 Axial Divergence (Low Angle  Modulated Structures*
Asymmetry)  Coherence differing by atom type* & ) 4 ln 2((T ) 2 #
4 ln 2
 Sample placement/transparency  Compositional inhomogeneity! G ((T , 'G ) = 2
exp $ !
 Crystallite Broadening* * 'G %$ 'G 2 "!
* Hard to model
 Strain Broadening* ! Sometimes impossible to model
* Note that these effects can vary for
Normalized Lorentzian
different classes of reflections
2 1
L( 'T , ( L ) = 2
(anisotropic peak broadening) )( L & 2'T #
1 + $$ !!
% (L "
Gaussian & Lorentzian functions compared. Both
Note that peak widths vary so curves have same FWHM & area, but note the
Results of these factors are convoluted to produce the observed peak shape. ΓG and γL are both functions of Q much longer tails for the Lorentzian.
Lorentzian.

7 8

2
Voigt vs. Pseudo-Voigt CW: Variation of FWHM with Q

A Gaussian convoluted with a Lorentzian function is a Voigt function, Gaussian Lorentzian

however the Voigt is slow to compute and the derivatives are messy. Few  U, V & W are Cagliotti terms, derived  X is crystallite broadening
Rietveld programs implement a Voigt. as instrument response function for  Y is strain broadening
CW neutron diffraction. Incomplete for
x-rays.
The “pseudo-Voigt” is the weighted sum of a Gaussian & Lorentzian function  P is a crystallite broadening X
– approximation is normally pretty good contribution " = + Y tan !
cos !
Fractions of each function depend on the relative widths of each [see mixing
factor (η) in GSAS manual, η=0 is Gaussian, η=1 is Lorentzian] N.B. no instrumental broadening term

"g2 = U tan 2 ! + V tan ! +

W+P
cos 2 !

9 10

Crystallite Size Broadening Crystallite Size Broadening

The Fourier transform (FT) from GSAS fits crystallite broadening

an infinite array of regularly Δd*=constant with two profile terms:
spaced objects is an array of • LX -> Lorentzian
delta functions. "d "! cot ! • GP -> Gaussian
"d* = =
The FT from a finite length array d2 d
is broadened. b* Relation between avg. size (p) and
"2#cot #sin#
= GSAS terms:
The finite sizes of crystallites will $
a*
broaden all orders of reflections
equally in units of Q (∝ d*)
Crystallite Size Broadening can $"d 18000K" 18000K"
– differing reciprocal space produce Lorentzian peak shapes
"2# = 2 p= p=
directions may have differing (common) or Gaussian peak shapes ! d cos# #LX # GP
amounts of broadening, if (uncommon) or a combination of both. K ≅ 1 (Scherrer constant, related to
crystallites dimensions are See GSAS Manual, pp 158-167.
crystal shape)
not isotropic on average
!
! !

3
Microstrain Broadening Microstrain Broadening

!d
= cons tan t GSAS fits strain broadening with two
When a material has residual stresses d
profile terms:
present, some crystallites are "d "d *
compressed. This must be balanced = = "! cot ! • LY -> Lorentzian (most common)
by other crystallites that are stretched d d* • GU -> Gaussian
(because ∑F=ma=0) 2 "d (note that GU also has an instrumental
" 2! = tan !
d contribution)
This leads to a range of lattice b*
constants. See GSAS Manual, pp 158-167.
a*
The spread between diffraction Relation between strain (as percentage) and GSAS terms:
locations for the maximum and
minimum lattice constant increases " "
linearly with Q (∆Q/Q or ∆d/d = S = 100% LY S = 100% GU # GU I
constant) 18000 18000
where GUI accounts for the instrumental contribution

! !

Where Strain Differs by Reflection Class: Anisotropic strain broadening terms

“Anisotropic peak broadening”

Strain may be anisotropic

– think of a layered material where the layers can be pulled apart
without much effort, but the layers themselves are quite “hard”
(resistant to applied forces).
– Such a material will be “squishy” in the layer direction and rigid in the
other two (more broadening in the squishy direction.)

Canonical anisotropic strain model: P. W. Stephens, Journal of Applied

Crystallography 32, 281 (1999).
– Restricts strain components in terms of 1st & 2nd-order terms allowed
by lattice symmetry

15 16

4
Anisotropic strain broadening terms

Axial Divergence: Low-angle peak asymmetry

17 18

Bragg cones in powder diffraction Axial Divergence (Low Angle Asymmetry)

Work of Finger, Cox & Jephcoat, based on derivations by van Laar & Yelon

FCJ: Convolute
profile with this
Beam curve

19 20

5
F-C-J: Example

 The Finger-Cox-Jephcoat correctly models the effective shift of the peak

due to axial divergence.

Note: the “competition,”

the split Pearson VII
(empirical), does not
Other instrumental effects
model this effect at all!

21 22

Sample Displacement & Transparency

In Bragg-Brentano geometry, samples are ideally placed exactly at

rotation axis and all diffraction occurs from sample surface (highly
absorbing sample). Neither is commonly true.
 Peak centers are shifted by "T ' = "T + S s cos ! + Ts sin 2!
– Sample Displacement (SHFT), Ss " !RS s " 9000
displacement = µ eff =
– Sample transparency (TRNS), Ts 36000 ! RTs GSAS Controls
R is diffractometer radius
These corrections correlate very highly with the zero correction for 2θ,
ZERO. Do not refine this too.

Parallel-Beam instruments (neutron or synchrotron) are very tolerant of

displacement and transparency. Never refine SHFT or TRNS, but do
refine ZERO (correction to 2θ).

23 24

6
Least Squares Controls Intensity Extraction Controls

These settings determine if and how reflection intensities will be estimated.

Title There is an overall title & one for each phase
# of cycles 0 to compute pattern with no refinement, N.B.
Le Bail extraction is performed even with 0 Extract Fobs* Compute Fobs values. Needed for Le Bail
cycles extraction & Fourier computation (I
Print options I recommend always using “summary of recommend: always on)
shifts” Extraction Normal extraction + two LeBail options:
Convergence GSAS considers the refinement to be Method F(calc) ! Start LeBail using Fobs = Fcalc from
Criterion 2
converged when ! ( shift / " ) is less than this. structure or
Increase with large numbers of parameters. Equal Weighted !start LeBail with Fcalc = 1
Marquardt Values >1 (1.2 recommended?) decrease the * note there is a Extract Fobs flag for each histogram and a
Damping sensitivity of the refinement to correlation, but extraction methods set for each phase in each histogram
slows convergence.

25 26

Damping Controls

Damping is available throughout GSAS to decrease the changes made to

parameters.
 More damping:
– smaller changes, Setting up GSAS Models
– more least-squares cycles required for convergence,
– but divergence is less likely. What parameters should be used?
 Damping values:
– 0: apply full shift
– 1: apply 90% of shift…
– 9: apply 10% of shift

27 28

7
Phase (structural) Parameters
Histogram Parameters
(one set for each histogram)

Unit cell & atom parameters  Phase flags

– GSAS allows up to 9 phases – select which phases are present
– Occupancy of 1 => fully occupied site, even for special positions – run POWPREF after changing
(unlike SHELX)  Background
 GENLES constrains cell, atom coordinates, anisotropic displacement – I recommend use of type 1 (Chebyshev polynomial) with as many
parameters according to space group terms as needed
– N.B. EXPGUI does not enforce symmetry in site/cell edits CW Diffractometer constants
(note data type & defaults determined by instrument parameter file)
Wave -- refine if cell is fixed or if multiple histograms of different types
are used
POLA -- don’t refine
Zero -- refine for neutron & synchrotron, never for flat-plate Bragg-
Brentano

29 30

Scale Factor

 There is a scale factor for each histogram

 There is a phase fraction for every phase in every histogram.
– Phase fraction multiplies scale factor
– Phase fraction × Z × Molecular Weight (asym unit) ∝ weight fraction Profile functions and profile parameter definitions
of phase (GSAS computes this in .LST file)

Redundant parameters: Don’t refine scale and all phase fractions together

31 32

8
CW Profile functions summarized
CW Profile terms
GU, Gaussian widths as polynomial in tan ; referred to as
GV, Cagliotti terms (U,V,W)
 #1: Gaussian only (rare except for BT-1 @ NIST) GW
 #2: Pseudo-Voigt “standard” GP Gaussian crystallite size (Scherrer) broadening, if GP
is refined, fix GU, GV, GW to instrumental values. I
 #3: same as #2, but with Finger-Cox-Jephcoat asymmetry. Much slower recommend leaving GP=0, but refine GU, GV & GW
than #2, but needed with low angle data LX Lorenzian crystallite size (Scherrer) broadening
 #4: same as #3, but also has Stephens anisotropic strain broadening LY Lorenzian strain broadening
parameters •Use care when initially refining terms, correlation is usually high
•GSAS constrains GU,GW,GP,LX & LY > 0 & GV < 0; Bad things
happen, if terms will refine out of bounds.
Start with “reasonable” GU,GV,GW, then refine them together. I
then fix them & try LX & LY one at a time & retain if >> 0.
See FitWidths in CMPR for a way to find starting UVW values

33 34

CW Profile terms (cont.) CW Profile terms (cont.)

GSAS offers two ways to treat low-angle asymmetry Peak cutoff -- where to stop computing peak tails (smaller ⇒ GENLES
slower)
 poor: asym (profile functions #1 & #2)
– set to match experimental signal-to-noise
– works only if small or no low-angle data
 excellent: Finger-Cox-Jephcoat (profile #3 & #4)
For use in Flat-plate Bragg-Brentano only:
– two terms: S/L & H/L shft -- sample displacement
• refine with care, if at all (I never do) trns -- sample transparency
• never refine both together – prone to correlation. N.B fix histogram zero=0 if these are used.
– S = sample height
– H = detector height,
– L = diffractometer diameter
(N.B. FCJ use radius & half-heights, equivalent, but messy)

35 36

9
Other CW Profile terms (cont.) Profile: Asymmetric Broadening

I seldom, if ever, use these:  # 4 profile function

stec – anisotropic strain broadening (profile type 4 does this better) – The number of Sxxx terms dictated by unit cell symmetry
ptec – anisotropic crystallite size broadening – Eta: 0 to 1
sfec – sublattice anisotropic broadening (typically due to stacking faults)
0 Gaussian; 1 Lorentzian
L11… – don’t use (use type 4 profile instead)
If you suspect asymmetric broadening, test:
 Change to function #4, fix all other profile terms, refine the Sxxx terms
Before using any of these terms, spend a while reading the GSAS and see if they refine away from 1.
Manual (pages 158-167) and make sure you understand how these terms  If the fit improves significantly, then refine Eta, then relax original terms,
are implemented and the physics behind them (none in the case of Lxx). a few (or one) at a time.

37 38

Preferred Orientation: March-Dollase

concept:
– define one (or more) hkl vectors that are over- or under-represented
relative to random distribution
– abundance ratio for that direction
– if 2+ directions, relative amounts for each
Preferred Orientation Corrections – orientation flag (not implemented in EXPGUI)
Note: there is a M-D parameter set for each phase and each histogram

39 40

10
Preferred Orientation: ODF

 Orientational Distribution Function (ODF) aka Spherical Harmonics

 concept: define a shape that defines the relative amounts of crystallites in
each direction
– shape constrained by material (space group) and by specimen
symmetry
– Set of terms per phase (not histogram)
– easy to use Reducing Model Complexity
 Use Cylindrical symmetry, increase order slowly

41 42

“Hard” Constraints Hard Constraint Limitation

 Constraints reduce the # of parameters

 Two common uses:  A parameter cannot appear in more than one constraint equation
– where data provide no information – Problem: A2B3 structure where A is divided between three sites: 1, 2
&3
– model must be simplified to match data
 Starting model Occ1= Occ2= Occ3= 1/3
 EXPGUI implement profile & atom parameter constraints
– define two constraints: δ1 = -δ2 & δ1 = -δ3
– (many more constraint types in EXPEDT)
Does not work! (N.B. no error message in EXPEDT)
 Note: constraints fix shifts not values
– Exception: rigid bodies
Trick solution: define two atoms that share a site, 1A & 1B: Occ1A & Occ1B
 Occ1A+ Occ1B = Occ2= Occ3= 1/3
 and the define two constraints:
– δ1A = -δ2
– δ1B = -δ3
This Works!

43 44

11
“Soft” Constraints Rigid Body Refinement

 Build a priori knowledge into refinement

 Weighted “cost” function added to total χ2  Rigid Body: group of atoms have a fixed geometry relative to each other
– bond distances & angles – position in cell & orientation can be refined
– unit cell contents – 1 or more “size” parameters can be refined
– misc. geometrical (mostly for macromolecular) – must be defined in EXPEDT (not EXPGUI)
 Can be used to guide a balky refinement  Two steps in rigid body setup:
– relax weight at end to 0 (if possible) 1. Rigid bodies are defined in cartesian coordinates
 Not always possible to remove S-C weight 2. They are then added into the model (up to 9 times):
– be sure to separate derived vs. “driven” results in your paper! • Designate the first atom in the group
• Define the location of the origin in fractional coordinates
• Specify the orientation of the body by setting axes & rotations

 Guides to setting up rigid bodies:

http://www.uni-bayreuth.de/departments/crystal/rietveld/rigid_bodies.pdf
http://www.ccp14.ac.uk/ccp/web-mirrors/ian-swainson/fireside_fuide_to_rigid_bodies.pdf

45 46

What to Refine and When?

If all parameters in a refinement were close to the true minima and

correlation is minimal, then it would fine to start by optimizing all
everything.
Refinement Recipes (Tricks of the trade) Who is that lucky?

Getting GSAS to give good fits In practice parameters must be relaxed slowly, with the order dictated by
which parameters are farthest from the correct values

47 48

12
Basic strategy Refinement Recipe (part 1)

Need reasonable values for unit cell, scale factor, background and profile No single strategy works in all cases
before refining structural parameters, locating missing atoms…
– Le Bail fit provides cell, background and profile
 Fit unit cell with small # of background terms
Think: Are reflections in ~ the correct places?
 Having reasonable profile terms for instrumental resolution can save you
from heartache
 Fit 2theta zero (neutron/synchrotron)
– With this, you only need to fit sample broadening
Think: Are all peaks indexed?
– This is a good justification for refining standards
– If not: check cell & rethink space group
– impurity phases?

49 50

Refinement Recipe (part 2) Refinement Recipe (part 3)

 Is peak shape in the right ball park?  Refine displacement parameters (Uiso)
– If possible, postpone profile refinement to later stages – for complex structures, group Uiso terms
– If profile is way off due to sample broadening, (you do have Think: Is the structure reasonable in terms of distances & angles?
reasonable instrumental terms!) refine only sample terms – missing atoms? Examine Fourier map
– Refine occupancies? (x-rays: beware!)
 Fit atom coordinates
– release “big” atoms first  When fit is pretty good, refine profile terms (damping is often needed).
• big = high multiplicity*b or *f(Q) – To begin, vary sample-dependent terms like GP, LX and/or LY one at
a time.
– Many terms are “clamped” >0. Turn off terms that try to refine “out of
bounds”

51 52

13
Refinement Recipe (part 4) Refinement Recipe (part 5): Finish up

 Refine background, if fixed from LeBail. Add more background terms, if

needed.
Think: Are all lines indexed & reasonably fit?  Ungroup Uiso’s if data allow.
If not, are some peaks wider than others?  Attempt anisotropic refinement of “big” atoms, when data quality permits.
– Consider anisotropic broadening
Are groups of reflections all computing high or low?
– Consider preferred orientation

Think: Is refinement stable and are values chemically plausible?

Wrong space group?
Too few reflections?
– Get better (or more) data, add hard/soft constraints

53 54

Wrap up
Refinement Recipe (part 6)

Common problem: At high Q, peaks run together so that background cannot be Rietveld fits are almost never perfect since materials and instruments are
determined. not perfect. There is almost always something more to try.
– Choice of Background determines average Uiso. Refining Background and
Uiso together results in good fits but unreasonable average Uiso(<0.001 or Complex problems may stretch the limits of what can be learned from the
>0.025 Å2) values. data. Getting better data is always best, but more frequently one must
reduce the complexity of the model (restraints and constraints)
Solution: Need to fix the average Uiso.
– set all Uiso = to something reasonable (say 0.01 to 0.03) GSAS offers many options to address the above -- the GSAS manual is a
– refine background great place to start learning more about the options.
– fix background, & refine Uiso’s
– Document in your paper that background and displacement parameters
had too much correlation for independent refinement & say what was
done.

55 56

14