Room Temperature Ionic Liquids (RTILs), which are broadly defined as salts that are liquid below 100 °C, offer an alternative to typical organic solvents. Favourable properties, such as high thermal stability, large melting range, low vapour pressure (and consequently low flammability) and miscibility with both polar and nonpolar compounds, provide the motivation for these systems to become next-generation solvents for synthesis, catalysis and separation technologies. Since all ILs consist of both a cation and anion, interchange between existing and novel ions can deliver a vast set of new solvent systems, much quicker than derivatisation of regular, molecular systems. Consequently, the millions of potential systems for exploration make property prediction of novel ILs an exciting research opportunity. Coupled with the global demand to reduce waste and develop processes within the principles of sustainable and green chemistry, there is increasing interest in moving away from conventional solvent systems.
To speed up rational design methodologies and avoid costly synthetic routes, cost-effective alternatives to experimental screening of potential new ILs for favourable physical and mechanical properties are needed. The estimation of thermodynamic properties is an essential component of this task and the application of high-performance computing is central to achieving this. Estimation methods can come in various forms but can be divided into three broad categories: (1) molecular/atomistic simulation and computation, (2) empirical modelling (e.g. equations of state) or (3) quantitative structure property relationships, driven by machine learning approaches. The first requires a classical or quantum mechanical description of molecules and a theoretical framework (such as statistical mechanics) to derive macroscopic properties from these; the second is immersed in classical thermodynamics and exploits the relationships between state variables; the last finds patterns in data and can build linear or nonlinear functions of user-defined features to express physical properties without the need for direct, prior knowledge of these. Each comes with its own strengths and weaknesses in terms of resources needed (empirical information and computational cost), domain of applicability, accuracy and ease-of-use.
In this presentation, the prediction power of molecular simulation and machine learning methods as applied to ILs is put to the test. The constant pressure heat capacity is chosen as the target—this property expresses the response of a system’s energy to a temperature change and can be both quantified theoretically and measured experimentally with reasonable ease. A test set of five structurally diverse ionic liquids have been picked and their temperature dependent heat capacities calculated using classical molecular dynamics (MD) simulations with various force field implementations as well as a selection of machine learning algorithms. In addition to discussing the accuracy, the strengths and weaknesses of the difference approaches are also compared.
MD simulations, of systems consisting of 10k to 20k atoms, were run using the AMBER code, which has been implemented on both GPU and CPU architectures. The former is highly efficient and speedups in excess of 20x can be obtained (> 500 ns/day). The CPU code runs in parallel using an MPI implementation that scales well up to 64 processors (system size dependent). The Scikit-learn Python framework was used for machine learning model development with the Keras API and TensorFlow as backend for the artificial neural network models. In addition to built-in support for multithreading, where applicable, embarrassingly parallel steps in model training and validation were optimized using the mpi4py module.