Update (11/06/2013): the link to the raw data on Francesco Stellacci does not work anymore; here is a mirror of the files that were released last month.
This is a guest post by Philip Moriarty, Professor of Physics at the University of Nottingham
It is to Francesco Stellacci (FS)’s credit that he has now uploaded the data I requested some time ago. I appreciate that this will have been a time-consuming task – I sometimes struggle to find files I saved last week, let alone locate data from almost a decade ago! It’s just a shame that the provision of the data necessitated the involvement of the journal editors (and possibly required prompting from other sources).
It is also worth noting that, over the last week, FS has been very helpful in providing timely responses to my questions regarding the precise relationship of the data files in the archive to the figures published in the papers.
Unfortunately, the data in the archive leave an awful lot to be desired.
In the near future, we will write up a detailed analysis of the data in the archive, combining it with scanning probe data of nanoparticles we have acquired, to show that the conclusions reached by FS et al. on the basis of their STM data, and associated analyses, do not stand up to scrutiny. Like Raphaël, I do not agree with FS that the only appropriate forum for scientific debate is the primary literature. (I’ve written at length I’ve rambled on in my usual loquacious style recently about the importance of embedding ‘Web 2.0’ debate within the literature). Nonetheless, given, for one, the extent to which the stripy nanoparticle papers have been cited, there is clearly significant scope in the primary literature for examining the reliability of FS’ experimental methodology and data analysis.
For now, I would simply like to highlight a number of the most important problems with the data in the archive.
The problems are multi-faceted and arise from a combination of: (i) feedback loop artifacts; (ii) image analysis based around exceptionally low pixel densities (and interpolation, applied apparently unknowingly, to achieve higher pixel densities); (iii) a highly selective choice/sampling of features for analysis in the STM images; and (iv) experimental uncertainties/error bars which are dramatically underestimated.
Predrag Djuranovic showed in 2005 how it was possible for stripes to appear on bare gold and ITO surfaces due to improper feedback loop settings. Julian Stirling, a final year PhD student in the Nanoscience Group here in Nottingham, has written a simulation code which shows that both stripes and “domains” can appear if care is not taken to choose the gains and scan speed appropriately. I’ll curtail my loquacity for now and let the images do the talking…
It’s worth comparing these images against the 3D-rendered experimental data of Fig. 3 of Jackson et al., Nature Materials 2004:
Much more information on the simulations will be provided in due course in the paper I mentioned above but I am sure that Julian will be more than happy to address questions in the comments section below.
Just in case anyone might think that generating stripes in a simulated STM doesn’t quite address the observation of stripes in an ‘ex silico’ environment, we have, of course, produced our own experimental images of stripy nanoparticles.
In (A) the feedback loop gain is increased dramatically at the scan line highlighted by the arrows. Stripes then appear in the upper 2/3 of the image. A low pass filter – which is, in essence, equivalent to an interpolation because it washes out high spatial frequencies – is then applied to a zoom of the image (shown in B), followed by a 3D rendering (C). The similarity with the Nature Materials figures above is striking.
The key aspect of the image above is that the nanoparticles do not have a ligand shell. These are the traditional citrate-stabilised colloidal Au nanoparticles known widely in the nanoscience community (and beyond). They were deposited onto a Au-on-mica sample from water. The stripes arise from feedback loop ringing.
Of course, FS’ argument has always been that his group can distinguish between stripes and features due to improper feedback loop settings and those that are “real”. Unfortunately, the data in the archive really do not support this claim. Again, we’ll provide a detailed analysis in the forthcoming paper, but a single image, entirely representative of the contents of the archive as a whole, is enough to show the limitations of Stellacci et al.’s analysis…
The image above is an uninterpolated digital zoom of one of the images from FS’ archive (yingotmpa2to1011706.006; 196 x 196 nm2; 512 x 512 pixels). Images of this type were used to measure the “periodicity” of ripples in nanoparticles for the statistical analyses described in Jackson et al. JACS 2006 and Hu et al. J. SPM 2009. Quite how one reliably measures a periodicity/ripple spacing for the image above (and all the others like it) is, I’m afraid to say, entirely beyond me.
I’ve noted previously that taking very low resolution STM scans (pixel size = 0.38 nm for the image above) which are analysed via heavily interpolated digital zooms is not, let’s say, the norm in the scanning probe microscopy community. There’s a very good reason for this – why would we be satisfied with low resolution images, where the pixel size is comparable to the features of interest, when we can simply reduce the scan area, and/or increase the pixel density, and “up” the effective resolution?
It is not good experimental practice, to put it mildly, to set the imaging conditions so that the size of a pixel is of the same order as the scale of the features in which you’re interested. This is a bad enough problem for images where the ripple spacing is proposed to be of order 0. 7 nm (i.e. ~ 2 pixels), as for the image above. In Jackson et al., JACS 2006, however, it is claimed that the spacing between head groups for homoligand particles is also measured and is stated to be 0.5 nm. (Unfortunately, these data are not included in the archive).
If this 0.5 nm measurement was reached on the basis of a similar type of imaging approach to that above, then there’s a fundamental problem (even if the image wasn’t generated due to feedback loop artifacts). The Nyquist-Shannon sampling theorem tells us that in order to measure the period of a wave without aliasing artefacts, our sampling frequency must be at least twice that of the highest frequency component. In other words, to measure something with a period of 0.5 nm, we should have, as an absolute maximum, a pixel size of 0.25 nm. I have asked FS whether the same imaging conditions (i.e. 196 nm x 196 nm, 512 x 512 pixel) were also used for the homoligand particles. I cannot tell from the archive because the data are not there. (To be fair to FS, I did not previously ask him to provide these particular data).
Compare and contrast
The contrast of the images in the original Nature Materials 2004 paper is quite saturated for some reason. The data archive allows us to look at the original image before contrast adjustment. This is illuminating…
The box in the image on the right delineates the area used in Fig. 1 of the Jackson et al. Nature Materials paper (which is shown on the left). As was pointed out to me by a researcher in the group here, what is intriguing is that the ripples extend beyond the edges of the particles. An explanation of this was offered by FS on the basis that the ripples arise from particles in a layer underneath the “brighter” particles seen in the image. That’s certainly one explanation. There are others that spring perhaps more readily to mind…