(16d) Development of Accurate Surrogate Models for Crystal Structure Prediction: Bridging the Gap between Force Fields and DFT

Many organic molecules of industrial interest exhibit crystalline polymorphism, forming multiple solid-state structures, each displaying different physico-chemical properties. Therefore, understanding the crystal energy landscape and identifying all likely polymorphs of a commercial drug or chemical can be of great importance, especially if the compound is delivered in solid form, as is the case with suspended agrochemicals or oral dosage forms of pharmaceuticals. The discovery of new polymorphs is desirable from an R&D perspective and can help to avoid potentially disastrous situations from an operational standpoint. These observations have motivated the development of methods for polymorph structure prediction that can now be applied to molecules of industrial relevance, provided limits conformational flexibility [1].

While the field of Crystal Structure Prediction (CSP) has seen growing success over the last few decades, current approaches are limited by a conflict between accuracy and computational efficiency. On the one hand, force field-based approaches have been a staple of CSP for many years, yet due to limitations in accuracy, findings may not always align with experimental observations. Alternatively, Density Functional Theory (DFT) has emerged as a highly reliable method for the prediction of relative polymorph stabilities [2], but the computational cost of such approaches may be far beyond the resources typically available to many organisations or research groups for routine applications. The disparity between force fields and DFT is exemplified by the difference in computational cost, which often spans many orders of magnitude. Due to an increasing drive to study pharmaceutical compounds in the early stages of product development, it becomes imperative that crystal energy landscapes can be mapped quickly and reliably [3]. This motivates the development of new methods that can approach the accuracy of DFT, while remaining computationally tractable. This is especially important as the complexity of target systems continues to grow with increasing molecular size and flexibility, or with the number of components in the system (such as hydrates with unknown or variable stoichiometry).

In this talk, we will demonstrate recent advancements in force field parameter estimation capabilities for a well-established class of hybrid ab initio/empirical models used in CSP [4]. Using carefully curated training sets that include crystal energies, and optimal lattice geometries obtained from DFT calculations, the empirical component of this hybrid model is re-parameterised to create an effective surrogate model of DFT. Such data, while expensive to generate, are only calculated once and can be re-used to adjust the empirical component subject to changes in the ab initio term. This parameter estimation process can now be performed rapidly to obtain globally optimal parameter sets, and as the form of the force-field potential is unchanged, the computational cost remains low. The effectiveness of this approach has been demonstrated in a recent publication on the development of system specific parameters [5]. Further to this, we investigate the development of transferable parameters from large sets of DFT training data which include atoms common to organic molecules. We investigate the accuracy and computational efficiency of the resulting potentials relative to DFT models.

[1] Nyman, Jonas, and Susan M. Reutzel-Edens. "Crystal structure prediction is changing from basic science to applied technology." Faraday discussions 211 (2018): 459-476.

[2] Hoja, Johannes, et al. "Reliable and practical computational description of molecular crystal polymorphs." Science advances 5 (2019): eaau3338.

[3] Price, Sarah L., and Susan M. Reutzel-Edens. "The potential of computed crystal energy landscapes to aid solid-form development." Drug Discovery Today 21 (2016): 912-923.

[4] Pantelides, Constantinos C., Claire S. Adjiman, and Andrei V. Kazantsev. "General computational algorithms for ab initio crystal structure prediction for organic molecules." Prediction and Calculation of Crystal Structures. Springer, Cham, 2014. 25-58.

[5] Bowskill, David H., Sugden, Isaac J., George, Neil, Keates, Adam, Webb, Jennifer, Pantelides, Constantinos C., and Adjiman, Claire S. “Efficient parameterization of a surrogate model of molecular interactions in crystals.”, In Computer Aided Process Engineering. Elsevier, 2020 (in press)