scholarly journals Machine-Learnt Fragment-Based Energies for Crystal Structure Prediction

2019 ◽  
Author(s):  
David McDonagh ◽  
Chris-Kriton Skylaris ◽  
Graeme Day
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Force Field ◽  
Structure Prediction ◽  
Alternative Energy ◽  
Force Fields ◽  
Training Data ◽  
Crystal Structure Prediction ◽  
Energy Models ◽  
Lattice Energies

Crystal structure prediction involves a search of a complex configurational space for local minima corresponding to stable crystal structures, which can be performed efficiently using atom-atom force fields for the assessment of intermolecular interactions. However, for challenging systems, the limitations in the accuracy of force fields prevents a reliable assessment of the relative thermodynamic stability of potential structures. Here we present a method to rapidly improve force field lattice energies by correcting two-body interactions with a higher level of theory in a fragment-based approach, and predicting these corrections with machine learning. We find corrected lattice energies with commonly used density functionals and second order perturbation theory (MP2) all significantly improve the ranking of experimentally known polymorphs where the rigid molecule model is applicable. The relative lattice energies of known polymorphs are also found to systematically improve towards experimentally determined values and more comprehensive energy models when using MP2 corrections, despite remaining at the force field geometry. Predicting two-body interactions with atom-centered symmetry functions in a Gaussian process is found to give highly accurate results with as little as 10-20% of the training data, reducing the cost of the energy correction by up to an order of magnitude. The machine learning approach opens up the possibility of using fragment-based methods to a greater degree in crystal structure prediction, providing alternative energy models where standard approaches are insufficient.

scholarly journals Machine-Learnt Fragment-Based Energies for Crystal Structure Prediction

2019 ◽  
Author(s):  
David McDonagh ◽  
Chris-Kriton Skylaris ◽  
Graeme Day
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Force Field ◽  
Structure Prediction ◽  
Alternative Energy ◽  
Force Fields ◽  
Training Data ◽  
Crystal Structure Prediction ◽  
Energy Models ◽  
Lattice Energies

Crystal structure prediction involves a search of a complex configurational space for local minima corresponding to stable crystal structures, which can be performed efficiently using atom-atom force fields for the assessment of intermolecular interactions. However, for challenging systems, the limitations in the accuracy of force fields prevents a reliable assessment of the relative thermodynamic stability of potential structures. Here we present a method to rapidly improve force field lattice energies by correcting two-body interactions with a higher level of theory in a fragment-based approach, and predicting these corrections with machine learning. We find corrected lattice energies with commonly used density functionals and second order perturbation theory (MP2) all significantly improve the ranking of experimentally known polymorphs where the rigid molecule model is applicable. The relative lattice energies of known polymorphs are also found to systematically improve towards experimentally determined values and more comprehensive energy models when using MP2 corrections, despite remaining at the force field geometry. Predicting two-body interactions with atom-centered symmetry functions in a Gaussian process is found to give highly accurate results with as little as 10-20% of the training data, reducing the cost of the energy correction by up to an order of magnitude. The machine learning approach opens up the possibility of using fragment-based methods to a greater degree in crystal structure prediction, providing alternative energy models where standard approaches are insufficient.

scholarly journals Computational Pharmaceutical Materials Science: Beyond Static Structures.

2014 ◽  
Vol 70 (a1) ◽  
pp. C1541-C1541
Author(s):  
Jacco van de Streek ◽  
Kristoffer Johansson ◽  
Xiaozhou Li
Keyword(s):  
Crystal Structure ◽  
Molecular Dynamics ◽  
Force Field ◽  
Crystal Structures ◽  
Structure Prediction ◽  
Force Fields ◽  
Individual Compound ◽  
Crystal Structure Prediction ◽  
Temperature Dependent ◽  
Mechanical Methods

The five Crystal-Structure Prediction (CSP) Blind Tests have shown that molecular-mechanics force fields are not accurate enough for crystal structure prediction[1]. The first--and only--method to successfully predict all four target crystal structures of one of the CSP Blind Tests was dispersion-corrected Density Functional Theory (DFT-D), and this is what we use for our work. However, quantum-mechanical methods (such as DFT-D), are too slow to allow simulations that include the effects of time and temperature, certainly for the size of molecules that are common in pharmaceutical industry. Including the effects of time and temperature therefore still requires molecular dynamics (MD) with less accurate force fields. In order to combine the accuracy of the successful DFT-D method with the speed of a force field to enable molecular dynamics, our group uses Tailor-Made Force Fields (TMFFs) as described by Neumann[2]. In Neumann's TMFF approach, the force field for each chemical compound of interest is parameterised from scratch against reference data from DFT-D calculations; in other words, the TMFF is fitted to mimic the DFT-D energy potential. Parameterising a dedicated force field for each individual compound requires an investment of several weeks, but has the advantage that the resulting force field is more accurate than a transferable force field. Combining crystal-structure prediction with DFT-D followed by molecular dynamics with a tailor-made force field allows us to calculate e.g. the temperature-dependent unit-cell expansion of each predicted polymorph, as well as possible temperature-dependent disorder. This is relevant for example when comparing the calculated X-ray powder diffraction patterns of the predicted crystal structures against experimental data.

scholarly journals Comparing hypothetical structures generated in the third Cambridge blind test of crystal structure prediction

2005 ◽  
Vol 61 (5) ◽  
pp. 528-535 ◽  
Author(s):  
Bouke P. van Eijck
Keyword(s):  
Crystal Structure ◽  
Force Field ◽  
Structure Prediction ◽  
Crystal Structure Prediction ◽  
Blind Test ◽  
Good Correspondence ◽  
The Third ◽  
New Energy ◽  
Order Of Magnitude ◽  
True Energy

In the third Cambridge blind test of crystal structure prediction, participants submitted extended lists of up to 100 hypothetical structures. In this paper these lists are analyzed for the two small semi-rigid molecules, hydantoin and azetidine, by performing a new energy minimization using an accurate force field, and grouping these newly minimized structures into clusters of equivalent structures. Many participants found the same low-energy structures, but no list appeared to be complete even for the structures with one independent molecule in the asymmetric unit. This may well be due to the fact that a cutoff at even 100 structures cannot ensure the presence of a structure that has a relatively high ranking in another force field. Moreover, some structures should have possibly been discarded because they correspond to transition states rather than true energy minima. The r.m.s. deviation between energies in corresponding clusters was calculated to compare the reported relative crystal energies for each pair of participants. Some groups of force fields show a reasonably good correspondence, yet the order of magnitude of their discrepancies is comparable to the energy differences between, say, the first ten structures of lowest energy. Therefore, even if we assume that energy is a sufficient criterion, it is not surprising that crystal structure predictions are still inconsistent and unreliable.

scholarly journals Machine Learning-based Crystal Structure Prediction for X-Ray Microdiffraction

2018 ◽  
Vol 24 (S2) ◽  
pp. 144-145 ◽  
Author(s):  
Yuta Suzuki ◽  
Hideitsu Hino ◽  
Yasuo Takeichi ◽  
Takafumi Hawai ◽  
Masato Kotsugi ◽  
...  
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Structure Prediction ◽  
Crystal Structure Prediction ◽  
X Ray

scholarly journals First-year experience with fully automated crystal structure prediction

2014 ◽  
Vol 70 (a1) ◽  
pp. C1618-C1618
Author(s):  
Marcus Neumann ◽  
Bernd Doser
Keyword(s):  
Crystal Structure ◽  
Structure Prediction ◽  
Density Functional ◽  
Method Development ◽  
Force Fields ◽  
Crystal Structure Prediction ◽  
Blind Test ◽  
Structure Generation ◽  
Statistical Correlations ◽  
Pharmaceutical Molecules

With improving hardware and software performance, usability has become one of the main obstacles to a more widespread use of Crystal Structure Prediction (CSP) with the GRACE program. In terms of method development, important milestones had already been passed by the time of the 5th blind test [1] in 2010, including the parameterization of dispersion-corrected Density Functional Theory (DFT-D) [2], the generation of tailor-made force fields from ab-initio reference data [3], a Monte-Carlo parallel tempering crystal structure generation engine and a DFT-d reranking procedure exploiting statistical correlations. These components have now been incorporated in automated data flow processes that remove the burden of scores of expert decisions from the user. Summarizing the results of CSP studies performed with the new Force Field Factory and CSP Factory modules throughout a year, the current performance of CSP is critically assessed and further method development needs are pinpointed. Studied compounds include 20 small molecules with competing hydrogen bonds motifs, 4 mono-hydrates of non-ionic molecules and the hydrates and chloride salts of several amino acids. The ability to handle flexible pharmaceutical molecules is demonstrated by a validation study on aripiprazole with one and two molecules per asymmetric unit. Salient features of the energy landscapes of other pharmaceutical molecules are discussed. Statistics are presented for the accuracy of tailor-made force fields, and the energy ranking performance of several DFT-d flavors is compared.

scholarly journals Accelerating crystal structure prediction by machine-learning interatomic potentials with active learning

2019 ◽  
Vol 99 (6) ◽  
Author(s):  
Evgeny V. Podryabinkin ◽  
Evgeny V. Tikhonov ◽  
Alexander V. Shapeev ◽  
Artem R. Oganov
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Active Learning ◽  
Structure Prediction ◽  
Interatomic Potentials ◽  
Crystal Structure Prediction

scholarly journals CrySPY: a crystal structure prediction tool accelerated by machine learning

2021 ◽  
Vol 1 (1) ◽  
pp. 87-97
Author(s):  
Tomoki Yamashita ◽  
Shinichi Kanehira ◽  
Nobuya Sato ◽  
Hiori Kino ◽  
Kei Terayama ◽  
...  
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Structure Prediction ◽  
Prediction Tool ◽  
Crystal Structure Prediction

scholarly journals Ternary mixed-anion semiconductors with tunable band gaps from machine-learning and crystal structure prediction

2019 ◽  
Vol 3 (3) ◽  
Author(s):  
Maximilian Amsler ◽  
Logan Ward ◽  
Vinay I. Hegde ◽  
Maarten G. Goesten ◽  
Xia Yi ◽  
...  
Keyword(s):  
Crystal Structure ◽  
Machine Learning ◽  
Structure Prediction ◽  
Band Gaps ◽  
Crystal Structure Prediction
Sign in / Sign up

Export Citation Format

Share Document