Machine‐Learning Approaches to Tune Descriptors and Predict the Viscosities of Ionic Liquids and Their Mixtures

Abstract

This work consists on a new chemoinformatic approach based on two complementary artificial intelligence concepts. Random Forest and Kohonen neural network are applied on this context. The former provides a relevance measure of the numerical descriptors encoding either an ionic liquid or its mixtures. The code of a given chemical system is weighted according that relevance measure. The Kohonen neural network is trained with a set of weighted chemical systems. The next step comprises the use of the trained neural network as platform to obtain a tuned profile of numerical descriptors representing a generical chemical system. The tuning mechanism involves the topology of a chemical system-encoding vector in the neural network. The last step comprises the use of the tuned chemical systems to build predictive models of viscosities. The MOLMAP encoding technology is applied to represent ionic liquid systems and its mixtures.

The Melting Point Profile of Organic Molecules: A Chemoinformatic Approach

“…This information is useful when considering databases of reactions comprising exclusively examples where reactions occur. [ 9 ] Different applications highlight the straightforward capacity and adaptability of MOLMAP technology, namely in descriptor's compression, [ 11 ] viscosity prediction, [ 12 ] classification of chemical, [ 13 ] and metabolic reactions, [ 14,15 ] mutagenicity modeling, [ 16 ] and phenolic antioxidant activity prediction. [ 17 ]…”

“…This information is useful when considering databases of reactions comprising exclusively examples where reactions occur. [9] Different applications highlight the straightforward capacity and adaptability of MOLMAP technology, namely in descriptor's compression, [11] viscosity prediction, [12] classification of chemical, [13] and metabolic reactions, [14,15] mutagenicity modeling, [16] and phenolic antioxidant activity prediction. [17] The work here presented uses, as encoding units, binary atomic interactions and the respective pattern of activation, representing a generical chemical system, in a trained Kohonen neural network, in order to obtain its MOLMAP.…”

Prediction of the Phase Composition Profile of Three‐Compound Mixtures in Liquid‐Liquid Equilibrium: A Chemoinformatics Approach

“…This approach enables to compare systems of different nature, including different number of components, and can take into account the molar fractions of each chemical. MOLMAPs have been applied in different studies such as the classification of chemical reactions without assignment of reaction centers, [18] conversion of descriptor's dimensionality in QSAR applications, [19] chemical reactivity evaluation from databases with no negative data, [20] classification of metabolic reactions, [21,22] mutagenicity prediction, [23] QSAR analysis of phenolic antioxidants, [24] viscosity classification [25,26] and gas solubility in ILs. [27] Here the machine learning protocol involved Random Forest models [28,29] that receive information about the mixture (molecules and their relative amount), the temperature and the pressure, and predict the molar fraction of one component of the mixture (one of the two molecules) in a specific phase.…”

Target high‐efficiency ionic liquids to promote H2SO4‐catalyzed C4 alkylation by machine learning

“…Machine learning (ML) is a powerful approach for building quantitative structure-activity relationship (QSAR) models. 26,27 Actually, ML has presented a distinct advantage to develop the QSAR models regarding various properties of ILs or gas solubility in ILs, such as melting points, 25,28 viscosity, 29,30 electrical conductivity, 31 thermal decomposition temperature, 32 critical properties, 33 and toxicity, 34,35 H 2 S solubility, 36 CO 2 solubility, 37,38 and water solubility. 39 Among these reports, the structure features of ILs are usually represented by group contribution, molecular fingerprint, and molecular descriptor.…”