A review on machine learning algorithms for the ionic liquid chemical space

Abstract

In this review article, the authors discuss the use of machine learning algorithms as tools for the prediction of physical and chemical properties of ionic liquids.

Computer Simulations of Deep Eutectic Solvents: Challenges, Solutions, and Perspectives

“…ML is based on statistical processing of large datasets and detecting correlations between input and output data (for example, between structure and property) and using them to predict properties of new compounds [ 1 , 182 , 183 ]. One of the most common applications of ML is prediction of the quantitative structure-property relationship (QSPR) [ 1 , 184 ]. For this purpose, the most frequently used mathematical models are Artificial Neural Networks (ANNs) and methods of regression analysis [ 1 , 182 , 184 ].…”

“…One of the most common applications of ML is prediction of the quantitative structure-property relationship (QSPR) [ 1 , 184 ]. For this purpose, the most frequently used mathematical models are Artificial Neural Networks (ANNs) and methods of regression analysis [ 1 , 182 , 184 ].…”

“…ANN models consist of a set of nodes connected to each other and distributed over several layers, similar to the neuron cells in a human brain [ 182 ]. The data analysis typically begins with translating the structure from a chemical to mathematical language by coding it with molecular descriptors, or features [ 184 ]. There are a number of different types of descriptors; they can originate, for example, from properties with a clear physical meaning (e.g., molecular mass, the number of carbons in the structure, HOMO or LUMO energies and other characteristics obtained through the quantum chemical calculations) or from any type of topological indices [ 185 ].…”

“…ML algorithms can be used not only for the construction of FFs, but also in post-processing of simulation data or/and their interpretation [ 201 , 202 , 203 ]. Moreover, the results of the MD simulation can be used as an input for ML models [ 184 ]. Although some works exist where MD and ML methods have been applied synergistically, to the best of our knowledge, there are currently no examples of the implementation of these two approaches in combination with DES research.…”

“…Moreover, in order to avoid the common problem of overfitting, some advanced ML models use the molar structure (e.g., graph-convolutional neural networks [ 204 ]) as their direct input. As a result, the number of parameters may be so large that they require a substantial amount of data to properly estimate the weights [ 184 ]. Thus, gathering a sufficient amount of experimental data is a major obstacle for developing advanced ML models, since training of these models may require thousands and even hundreds of thousands of entries on molecular properties that may not be available.…”