Basic objective function that can be used as a starting point for developing XRD full-pattern matching strategies. [details pending...]
f.noise(inits, pure.patterns, sample.pattern, eps.total = 0.05)
vector of initial guesses for mineral fractions, last item is a noise component
a matrix of XRD patterns of pure samples, resampled to the same twotheta resolution and rescaled according to an external standard
the unkown or composite pattern, aligned to the same twotheta axis as the pure patterns and rescaled to an external standard
precision of comparisons; currently not used
the sum of absolute differences between the unkown pattern and combination of pure patterns for the current set of mixture proportions
This is similar to the work of Chipera and Bish (2002), using the methods described in (Bish, 1994). If the flexibility of a custom objective function is not required, the linear model framework should be sufficient for pattern fitting. GLS should be used if realistic standard errors are needed.
Chipera, S.J., & Bish, D.L. (2002) FULLPAT: A full-pattern quantitative analysis program for X-ray powder diffraction using measured and calculated patterns. J. Applied Crystallography, 35, 744-749.
Bish, D. 1994. Quantitative Methods in Soil Mineralogy, in Quantitative X-Ray Diffraction Analysis of Soil. Amonette, J. & Zelazny, L. (ed.) Soil Science Society of America, pp 267-295.
# NOT RUN {
# sample data
data(rruff.sample)
# get number of measurements
n <- nrow(rruff.sample)
# number of components
n.components <- 6
# mineral fractions, normally we don't know these
w <- c(0.346, 0.232, 0.153, 0.096, 0.049, 0.065)
# make synthetic combined pattern
# scale the pure substances by the known proportions
rruff.sample$synthetic_pat <- apply(sweep(rruff.sample[,2:7], 2, w, '*'), 1, sum)
# add 1 more substance that will be unknown to the fitting process
rruff.sample$synthetic_pat <- rruff.sample$synthetic_pat +
(1 - sum(w)) * rruff.sample[,8]
# try adding some nasty noise
# rruff.sample$synthetic_pat <- apply(sweep(rruff.sample[,2:7], 2, w, '*'), 1, sum) +
# runif(n, min=0, max=100)
# look at components and combined pattern
par(mfcol=c(7,1), mar=c(0,0,0,0))
plot(1:n, rruff.sample$synthetic_pat, type='l', axes=FALSE)
legend('topright', bty='n', legend='combined pattern', cex=2)
for(i in 2:7)
{
plot(1:n, rruff.sample[, i], type='l', axes=FALSE)
legend('topright', bty='n',
legend=paste(names(rruff.sample)[i], ' (', w[i-1], ')', sep=''), cex=2)
}
## fit pattern mixtures with a linear model
l <- lm(synthetic_pat ~ nontronite + montmorillonite + clinochlore
+ antigorite + chamosite + hematite, data=rruff.sample)
summary(l)
par(mfcol=c(2,1), mar=c(0,3,0,0))
plot(1:n, rruff.sample$synthetic_pat, type='l', lwd=2, lty=2, axes=FALSE,
xlab='', ylab='')
lines(1:n, predict(l), col=2)
axis(2, cex.axis=0.75, las=2)
legend('topright', legend=c('original','fitted'), col=c(1,2), lty=c(2,1),
lwd=c(2,1), bty='n', cex=1.25)
plot(1:n, resid(l), type='l', axes=FALSE, xlab='', ylab='', col='blue')
abline(h=0, col=grey(0.5), lty=2)
axis(2, cex.axis=0.75, las=2)
legend('topright', legend=c('residuals'), bty='n', cex=1.25)
## fitting by minimizing an objective function (not run)
# SANN is a slower algorithm, sometimes gives strange results
# default Nelder-Mead is most robust
# CG is fastest --> 2.5 minutes max
# component proportions (fractions), and noise component (intensity units)
# initial guesses may affect the stability / time of the fit
## this takes a while to run
# # synthetic pattern
# o <- optim(par=c(0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1), f.noise,
# method='CG', pure.patterns=rruff.sample[,2:7],
# sample.pattern=rruff.sample$synthetic_pat)
#
#
# # estimated mixture proportions
# o$par
#
# # compare with starting proportions
# rbind(o$par[1:n.components], w)
#
# # if we had an unkown pattern we were trying to match, compare fitted here
# # compute R value 0.1 - 0.2 considered good
# # sum(D^2) / sum(s)
# # o$value / sum(rruff.sample$sample)
#
# # plot estimated mixture vs sample
# # combine pure substances
# pure.mixture <- apply(sweep(rruff.sample[, 2:7], 2, o$par[1:n.components], '*'), 1, sum)
#
# # add in noise
# noise.component <- o$par[n.components+1]
# est.pattern <- pure.mixture + noise.component
#
#
# # plot results
# par(mfcol=c(2,1), mar=c(0,3,0,0))
# plot(1:n, rruff.sample$synthetic_pat, type='l', lwd=2, lty=2, axes=FALSE,
# xlab='', ylab='')
# lines(1:n, est.pattern, col=2)
# lines(1:n, rep(noise.component, n), col=3)
# axis(2, cex.axis=0.75, las=2)
# legend('topright', legend=c('original','fitted','noise'), col=c(1,2,3), lty=c(2,1,1),
# lwd=c(2,1,1), bty='n', cex=1.25)
#
# plot(1:n, rruff.sample$synthetic_pat - est.pattern, type='l', axes=FALSE,
# xlab='', ylab='')
# abline(h=0, col=grey(0.5), lty=2)
# axis(2, cex.axis=0.75, las=2)
# legend('topright', legend=c('difference'), bty='n', cex=1.25)
#
# }
Run the code above in your browser using DataLab