Ionic Liquid Design Using Molecular Simulation and Statistical Methods

Abstract

Solicitous use of time is crucial for material design. In the material domain presented in this work, ionic liquids (ILs), there are theoretically 1014-18 possible pairwise molecular structures—too many to synthesize exhaustively.1,2 For this particular design problem, we turn to computational approaches. But even here an exhaustive approach is intractable. A method of deciding what system to simulate in the first place must be optimized. There are many deterministic and stochastic algorithms available for such search spaces. The Darwinian processes of evolutionary algorithms (EAs), work by mutating a candidate solution until it attains a desired fitness. In this case, the fitness is determined by a quantitative structure property relationship (QSPR) usually in the form of a machine learning (ML) model. Because the ML model is based on learning examples, it pairs well with the search strategy of an EA—starting molecular configurations for the EA are based on the same examples that have informed the ML model. When a particular solution deviates far from the training data (i.e. its molecular structure is different than the structures of the molecules in the training data) the uncertainty estimate in its property prediction is high. When this occurs, the EA solution can be simulated in MD to either: a) explore the structure landscape and inform/update the model or b) exploit the structure landscape because the prediction is close to our target. The holy grail, however, of any material design process is to operate on smooth featural or structural surfaces, and therefore be able to calculate explicit gradients to iterate toward the target property in question. Such a design process is explored through the generative capabilities of a class of stochastic neural networks—variational autoencoders—for the explicit rationalization of desired IL thermodynamic properties.