Dr Toms Rekis

🇪🇺

PhAI can solve the so-called ‘phase problem’ with lower-quality data than is needed for other methods and quickly arrives at high quality solutions. chemistryworld.com/news/ai-tool-o…

Really happy that our paper on solving the crystallographic phase problem using neural networks is out in @ScienceMagazine. Thanks to @mol_crystal_guy and @AndersOMadsen for a great collaboration. science.org/doi/10.1126/sc…

Today we've launched our book "Atlas of Fourier Transforms". It has been a two year passion project led by Miti Shah, a student in our lab. It is uniquely comprehensive and suitable as a coffee table book or advanced research reference. The full pitch is on Kickstarter. kickstarter.com/projects/hlab/…

@ntumanov_Xray I like Z’=8. Will it be published? I wonder what the anion is.

In fact, Z’=2 to Z’=4 to Z’=8 phase transitions. From here (left) to here (right).

Just for scale, it is a 1.5 liter bottle and no microscope. “Obviously, it is already known” compound. (spoiler: no). As a bonus, Z’=2 to Z’=8 phase transition on cooling. #crystallography

Born #OnThisDay in 1903 was crystallographer Kathleen Lonsdale FRS. She was one of the first two female scientists to be elected as a Fellow of the Royal Society, along with Marjory Stephenson. She discovered the structure of benzene, and was a committed pacifist.

Kathleen Lonsdale (neé Yardley), a pioneer in #Xray #crystallography, was born in 1903 #OTD. Famous for proving #benzene core structure; structure factor calculation formulas; serving as editor of Int. Tables for #Xray #crystallography (1952) & first woman @IUCr president (1966).

Hodgkin made huge contributions to #Xray #crystallography, determining the structures of #cholesterol iodide in 1943, #penicillin in 1945 (finding the least expected beta-lactam structure), and #vitamin B12 in 1955—result described by L. Bragg as "breaking the sound barrier".

@stecanossa @MJCliffe @IUCr Don’t forget molecular solid solutions where positions are different for at least some atoms of the two (or more) components. The more we try to define… continuum is hard to accept.

@stecanossa @ETH Congrats Stefano! ;)

@MartinWard_xtal @AndersOMadsen Most likely only the patterns.

@mol_crystal_guy @AndersOMadsen Thanks for the info. I see some hits when I search the ICDD, but these may be later only and no structure or indexing

Q for the diffraction folk on here - Im looking for cif files (or at least unit cell/space group) for phases of magnesium stearate. I have the trihydrate, but looking for the anhydrate and dihydrate forms. I think these may be on the ICDD, but sadly I dont have access :(

Crystallography is science of linear transformation.

@biochem_fan The space is VERY large, so many degrees of freedom.

"42,000,000 structures" for training are very large. I wonder what fraction the training set represents among all possible crystal structures below the given unit cell volume? Did it contain structures very similar to the test case? i.e. naphthalene is a very common motif.

Randomly omitting 15 % of reflections is different from more realistic cases of missing wedges (e.g. insufficient rotation in MicroED). Did the authors test such situations? Can the network be trained on such cases?

@biochem_fan @AndersSL can elaborate here. All indications are there that non-centrosymmetry will not be a problem. Already this network can find all 8 structure semivariants equally well.

What is the network performance when trained on a smaller set? Finally, it would be very interesting to see if this approach can be extended to non-centrosymmetric space groups, where phases are arbitrary.

@biochem_fan No, we did not try anything in real space. Setting everything up in the reciprocal space felt the way to go. I don’t know what exactly was tried, but was the origin problem dealt with?

One could replace traditional charge flipping, solvent flattening, histogram matching etc in dual space iterative algorithms with a density modifying neural network. Was such an approach tested (and failed)? (Actually I tried it 5 years ago for proteins and failed)

Preprint "PhAI: A deep learning approach to solve the crystallographic phase problem" from @AndersSL, @mol_crystal_guy,@AndersOMadsen chemrxiv.org/engage/chemrxi… Impressive and thank you very much for releasing both training codes and model weights.

@biochem_fan That’s just to be able to solve structures where some of the strongest reflections are missing due to whatever reasons. Of course the next step is to generate realistic incomplete data — missing cone/wedge in ED or incompleteness in high-pressure studies. That’s planned!

@ntumanov_Xray It’s there with some examples. :)

@mol_crystal_guy @AndersSL @AndersOMadsen Evil wizardry... I did not downloaded yet full SI, but does it include "how-to-use-for-dummies"? I see now why you were doing some crystallographic archeology some time ago :)

Imagine solving crystal structures from low-resolution SC- and PXRD data with a single click!? Our pilot study was designed for small unit cells and P21/c (+supergroups) only, but the results are striking! Preprint: doi.org/10.26434/chemr… @AndersSL @AndersOMadsen

@MartinWard_xtal @AndersSL @AndersOMadsen If P1 was cracked, structure in any space group would be solvable with that (if the data represent a regular 3D structure). Twinning or some other irregularities would of course pose problems. But ideas for further work are endless. For example, dealing with merohedral twinning.

@mol_crystal_guy @AndersSL @AndersOMadsen Wha if I have weak data but a SCXRD dataset with a strong indexing (P212121) which is also supported by PXRD indexing - might your development help to get a structure from this....?

@MartinWard_xtal @AndersSL @AndersOMadsen The ultimate goal would be P1. @AndersSL is wizarding. We initially chose only one space group for convenience. All supergroups are possible if the data are indexed leading to specific (mostly non-standard) settings that bring the symmetry operators in congruence.

@mol_crystal_guy @AndersSL @AndersOMadsen this sounds amazing - great work and thanks for highlighting the pre-print! what are the challenges on expanding this to consider more spacegroups?