Inorganic Structure Prediction with GRINSP Armel Le Bail Université du Maine, Laboratoire des oxydes et Fluorures, CNRS UMR 6010, Avenue O. Messiaen, Le Mans Cedex 9, France.

CONTENT Introduction GRINSP algorithm GRINSP predictions Opened doors, and limitations Prediction confirmation Conclusion

I- INTRODUCTION To predict a crystal structure is to be able to announce it before any confirmation by chemical synthesis or discovery in nature. A predicted structure should be sufficiently accurate for the calculation of a predicted powder pattern that would further be used with success in the identification of a real compound not yet characterized.

If the state of the art had dramatically evolved, we should have huge databases of predicted compounds. Not any new crystal structure would surprise us since it would correspond already to an entry in that database. Moreover, we would have obtained in advance the physical properties and we would have preferably synthesized those interesting compounds. Of course, this is absolutely not the case. Where are we with inorganic structure prediction?

Things are changing, maybe : Two databases of hypothetical compounds were built in > hypothetical zeolites at : >2000 inorganic compounds in PCOD (zeolites as well as other oxides and fluorides) at : However, inorganic prediction software and methods remain scarce: CASTEP, GULP, G42, SPuDS, AASBU, CERIUS2… Hence the development of a new one : GRINSP

II- GRINSP Algorithm Geometrically Restrained INorganic Structure Prediction Applies the knowledge of the common geometrical characteristics of a well defined group of crystal structures (N-connected 3D nets with N = 3, 4, 5, 6 and combinations of two N values), in a Monte Carlo algorithm, In GRINSP, the quality of a model is established by a cost function depending on the weighted differences between calculated and ideal interatomic first neighbour distances M-X, X-X and M-M in binary M a X b or ternary M a M' b X c compounds. J. Appl. Cryst. 38, 2005,

Comparison of a few GRINSP-predicted cell parameters with observed ones Predicted (Å)Observed or idealized (Å) Dense SiO 2 abcRabc Quartz Tridymite Cristobalite Zeolites ABW EAB EDI GIS GME JBW LTA RHO Aluminum fluorides  -AlF Na 4 Ca 4 Al 7 F AlF 3 -pyrochl

More details on the GRINSP algorithm Two steps : 1- Generation of structure candidates First the M/M’ atoms are placed in a box whose dimensions are selected at random, and the model should exactly correspond to the geometrical specifications (exact coordinations, but some tolerance on distances). The cell is progressively filled with M/M’ atoms, up to completely respect the geometrical restraints, if possible. The number of M/M' atoms placed is not predetermined. In this first step, atoms do not move, their possible positions are tested and checked, then they are retained or not.

2- Local optimization The X atoms are added at the midpoints of the (M/M')-(M/M') first neighbours. It is verified by distance and cell improvements (Monte Carlo moves) that regular (M/M’)X n polyhedra can really be built. The cost function is based on the verification of the provided ideal distances M- M, M-X and X-X first neighbours. A total R factor is defined as : R =  [(R 1 +R 2 +R 3 )/ (R 01 +R 02 +R 03 )], where R n and R 0n for n = 1, 2, 3 are defined as : R n =  [w n (d 0n -d n )] 2, R 0n =  [w n d 0n ] 2, where d 0n are the ideal first interatomic distances M-X (n=1), X-X (n=2) and M-M (n=3), whereas d n are the observed distances in the structural model. The weights retained (w n ) are those used in the DLS software for calculating idealized zeolite framework data (w 1 = 2.0, w 2 = 0.61 and w 3 = 0.23).

The ideal distances are to be provided by the user for pairs of atoms supposed to form polyhedra (for instance in the case of SiO 4 tetrahedra, one expects to have d 1 = 1.61 Å, d 2 = Å and d 3 = 3.07 Å). For ternary compounds, the M-M' ideal distances are calculated by GRINSP as being the average of the M-M and M'-M' distances. It is clear that this R factor considers only the X-X intra-polyhedra distances, neglecting any X-X inter-polyhedra distances This cost function R could possibly be better defined differently, for instance by using the bond valence sum rules (this is in project for the next GRINSP version). Minimizing the Difference of Distances with Ideal distances is a very basic approach… This basic approach can work only for regular polyhedra

During this second step, the atoms are moving, but no jump is allowed because a jump would break the coordinations established at the first step. This is a simple routine for local optimization. The change in the cell parameters from the structure candidate to the final model may be quite considerable (up to 30%), During the optimization, the original space group used for placing the M/M' atoms may change after adding the X atoms, so that the final structure is always proposed in the P1 space group, and presented in a CIF. The final choice of the real symmetry has to be done by using a program like PLATON. More on the optimization second step

How GRINSP works : 1- Create a small datafile corresponding to your desire Example for such a datafile: TiO6/VO5 Pbam - 55 ! Title P B A M ! Space group ! Nsym (symmetry code), Npol, etc 6 5 ! Coordinations of these npol polyhedron-type Ti O ! Definition of the elements for the first polyhedron V O ! Definition of the elements for the second polyhedron ! Min and max a, b, c ! Min and max angles ! Min and max framework density ! Nruns, MCmax, Rmax saving, optimizing ! number of MC optimization cycles, refinement code ! first filename (will be cif,.xtl,.dat, etc)

2 – Verify if your atom-pairs are already defined : See into the distgrinsp.txt file : V O Ti O These are minimal, maximal and ideal distances for V-V, V-O and O-O in VO 5 square pyramids, and for Ti-Ti, Ti-O and O-O in TiO 6 octahedra.

3- Run GRINSP

4- Wait a bit (one day…) and, when finished, see the summary file :

5 – See the results (here using Diamond from a CIF) :

GRINSP is Open Source, GNU Public Licence Download it at :

III- GRINSP Predictions Binary compounds Formulations M 2 X 3, MX 2, M 2 X 5 and MX 3 were examined Zeolites More than a thousand models (not >100000) were built with R < 0.01 and cell parameters < 16 Å and placed into the PCOD database. The way GRINSP recognizes a zeotype is by comparing the coordination sequence (CS) of any model with a list of previously established ones (as well as with the other CS already stored during the current run). The CIFs can be obtained by consulting the PCOD database, giving the entry number provided with the figure caption (for instance PCOD , etc).

Hypothetical zeolite PCOD SG : P432, a = Å, FD = 11.51

Hypothetical zeolite PCOD SG : P6/m, a = Å, c = 7.13Å, FD = Estimated number of zeolite models proposed by GRINSP : > 2000

Hypothetical aluminosilicate PCOD SG : P432, a = Å - FD = formulation : [Si 2 AlO 6 ] -1 Estimated number of aluminosilicates proposed by GRINSP : > 2000

Hypothetical aluminophosphate SG : Pma2, a = Å, b = 8.06 Å, c = 5.64 Å - FD = 13.9 formulation : [Al 4 PO 10 ] -3 Estimated number of aluminophosphates proposed by GRINSP : > 2000

Can GRINSP predict > zeolites as well ? Yes, if R max fixed at 0.02 instead of 0.01, if the cell parameters maximum limits (16Å) are enlarged, and if multi-redundant solutions in various space groups are all kept. I prefer not. Is there any sense to predict > zeolites when less than 200 are known ?

B 2 O 3 polymorphs predicted by GRINSP Not a lot of crystalline varieties are known for this B 2 O 3 composition. Too many are proposed by GRINSP. Hypothetical B 2 O 3 PCOD

Hypothetical B 2 O 3 PCOD Estimated number of B 2 O 3 models proposed by GRINSP : > 3000

M 2 X 5 compounds Example : unknown V 2 O 5, SG: Pbam, a = Å, b = Å, c = 7.25 Å, FD = 16.5, R = , VO 5 square pyramids : Estimated number of V 2 O 5 models proposed by GRINSP : > 200

AlF 3 polymorphs yet to be synthesized, predicted by GRINSP All the known structure-types (5) were retrieved, Two other structure types existing with stuffed MX 3 formulations were proposed. Five unknown, “yet to be synthesized" AlF 3 polymorphs were predicted That time, the total number is small : 12 models only with R < 0.02.

Classification of the 12 AlF 3 polymorphs proposed by GRINSP (identified as known or unknown) according to increasing values of the distance quality factor R < 0.02 Structure-type FDabc  SGZNR HTB P6 3 /mmc TlCa 2 Ta 5 O Pmmm U-1 (AlF 3 ) P2 1 /m Pyrochlore Fd-3m U-2 (AlF 3 ) P-4m Perovskite Pm-3m Ba 4 CoTa 10 O Iba TTB P4 2 /mbc U-3 (AlF 3 ) Pnc  -AlF P4/nmm U-4 (AlF 3 ) I4 1 /a U-5 (AlF 3 ) P4 2 /mmc FD = framework density (number of Al atoms for a volume of 1000Å 3 ). SG = higher symmetry spage group in which the initial model of Al-only atoms was obtained (not being necessarily the true final space group obtained after including the F atoms). Z = number of AlF 3 formula per cell. N = number of Al atoms with different coordination sequences. R = quality factor regarding the ideal Al-F, F-F and Al-Al first neighbour interatomic distances.

Yet to be synthesized U-3 (AlF 3 ).

Known :  -AlF 3 - tetrahedra and chains of octahedra

Unknown : U-4 (AlF 3 ), dense packing of tetrahedra of octahedra, exclusively

Model 13 : U-6 (AlF 3 ), R > 0.02, not viable due to a too high level of octahedra distortion and short F-F distances

By-products of the search with GRINSP Irregular polyhedra can be produced… For instance, sixfold polyhedra other than octahedra can be produced: trigonal prisms or pentagonal based pyramids. Since they do not correspond to one unique ideal X-M distance or M-X distance, they are ranked with high R-values.

Octahedra + pentagonal based pyramids :

Octahedra + trigonal prisms :

Chimeric compound mixing trigonal prisms with distorted trigonal bipyramids

Two- and one-dimensionnal compounds can be formed. Nanotubes with B 2 O 3 formulation for instance :

Ternary M a M’ b X c compounds with corner- sharing 3D nets M/M’ with same coordination but different ionic radii or different coordination The built ternary compound will not always be electrically neutral.

Borosilicates PCOD , Si 5 B 2 O 13, R = Estimated number of models built by GRINSP : > 3000 SiO 4 tetrahedra and BO 3 triangles

Aluminoborates Estimated number of models built by GRINSP : >2000 Example : [AlB 4 O 9 ] -2, cubic, SG : Pn-3, a = Å, R = : AlO 6 octahedra and BO 3 triangles

Titanosilicates [Si 2 TiO 7 ] 2-, R = , SG : P4 2 /mmc, a = 7.73 Å, c = Å, FD = Estimated number of models built by GRINSP : > 500 TiO 6 octahedra and SiO 4 tetrahedra

Fluoroaluminates Known as Na 4 Ca 4 Al 7 F 33 : PCOD [Ca 4 Al 7 F 33 ] 4-. AlF 6 and CaF 6 octahedra

Unknown : PCOD [Ca 3 Al 4 F 21 ] 3-. Estimated number of fluoroaluminates models built by GRINSP : ???

A satellite software (GRINS) can build isostructural compounds faster than running again GRINSP However, changing the atomic radius may lead to different structures… Automatization is essential for the fast feeding of the PCOD, unfortunately, human eyes looking at the predicted structure is still essential : 5 minutes at least are needed for an evaluation before adding the CIF into the database. With zeolites, identification is easy because the coordination sequences of the known phases helps to recognize if the prediction leads to a new model or is already known But this is less easy with non-zeolites because there is no general extension of structure-types descriptors

IV - Opened doors and limitations Limitation : corner-sharing polyhedra Potentially already > 50 or hypothetical compounds in PCOD (only 2000 added yet) Scheduled improvements Make appear corner-, edge-, and face-sharing polyhedra, altogether. Propose an automatic way to obtain an electrical neutrality by the detection of holes and the filling of these holes by large cations. Use of bond valence rules at the optimization step, or/and energy calculations. Extension to quaternary compounds. Etc.

With a few modifications, GRINSP could Predict structures for ice H 2 O (on the basis of distorted OH 4 tetrahedra): or predict alloys M x M’ y characterized by MM’ 4 and M’M 4 tetrahedra, or predict fullerene structures, or predict structures for series of organic compounds provided they can be described by common geometrical features, etc. You are limited only by your own imagination…

GRINSP can already predict structures deriving from perovskite by oxygen vacancies : Octahedra and square pyramids : > 500 predictions

Brownmillerite A problem with ICSD is the difficulty to identify if a predicted structure-type is already described in the database. Generalized topology descriptors are lacking…

V- Prediction confirmation More difficult even is the prediction of the synthesis conditions for making to appear these predicted crystal structures. However, if the chemical composition involves at least 3 elements or more, one may try the battery of classical synthesis methods. If an interesting model is predicted having the [Ca 3 Al 4 F 21 ] 3- formulation, may be it could be really synthesized as Na 3 Ca 3 Al 4 F 21 or Li 3 Ca 3 Al 4 F 21, or may be not. We can already be sure that most predictions will be vain, never confirmed, because the synthesis route may depend on a precursor (organometallic, hydrate, amorphous compound) which itself is yet unknown, or because the prediction is simply false.

The [Ca 4 Al 7 F 33 ] 4- network proposed by GRINSP really exists with the Na 4 Ca 4 Al 7 F 33 formulation. The more the predicted inorganic formula is complex, the more easy classical and direct synthesis routes can be tested, but metastable compounds will mostly occur from indirect routes.

For the confirmation of the predictions, we will have to wait for decades or centuries, who knows. Anyway, structure (and properties) prediction is an unavoidable part of our future in crystallography and chemistry. Advantages are obvious. We need for searchable databases of predicted compounds, preferably open data on the Web. If we are not able to do that, we cannot pretend having understood and mastered the crystallography rules.

Citation from Frank C. Hawthorne (1994) : "The goals of theoretical crystallography may be summarized as follow: (1) predict the stoichiometry of the stable compounds; (2) predict the bond topology (i.e. the approximate atomic arrangement) of the stable compounds; (3) given the bond topology, calculate accurate bond lengths and angles (i.e. accurate atomic coordinates and cell dimensions); (4) given accurate atomic coordinates, calculate accurate static and dynamic properties of a crystal. For oxides and oxysalts, we are now quite successful at (3) and (4), but fail miserably at (1) and (2)" F. C. Hawthorne, Acta Cryst. B50 (1994) As a conclusion : generalizing GRINSP would be an empirical answer at goals (1) and (2). We have to stop to « fail miserably »!

We need for a database pointing at the future materials. I suggest you to explore your usual crystallography domain, and to help me to feed PCOD with high quality hypothetical compounds either with GRINSP or using any other prediction software. VI - Conclusion This is the future of chemistry and crystallography. Thanks !