AI-Driven Cheminformatics Models of Chemical Mixtures for Sustainable Design of Drop-in Biofuel Blends

Abstract

Abstract Complex chemical mixtures (involving multiple chemical compounds) such as polymers mixtures, ionic liquids, azeotropes, metabolites, and drop-in biofuels, are present in almost all areas of chemical research and industry. Specifically, designing sustainable fuel blends and/or drop-in biofuels by adding eco-friendly pre-mixtures of chemical compounds (compounding) may help to reduce environmental impact. However, experimental testing of all possible pre-mixtures is time and resources consuming. In this context, Cheminformatics approach to complex fuel mixtures is an important challenge of the major relevance. Artificial intelligence/Machine learning (AI/ML) models may help to reduce experimentation cost but there are not publicly available datasets with detailed chemical composition of fuel blends. Consequently, in this work, we assembled a dataset of 1222 fuel blends previously reported with at least 20 compounds each one. After this, Information Fusion and Perturbation Theory Machine Learning (IFPTML) strategy was used to pre-process the data. Next, we seek multiple linear and non-linear AI/ML models able to predict the RON and MOM values of these mixtures. In so doing, Multivariate Linear Regression (MLR), Radial Basis Function (RBF), Multi-Layer Perceptron (MLP), And Deep Neural Network (DNN) algorithms were tested for comparative purposes. The best models found predict the output values with r2 in the range 0.89–0.99 in training and validation series. Last, we run simulations with > 10000 and > 5000 data point of drop-in biofuels and eco-friendly fuel blends both made of a pre-mixture of eco-friendly components and a base blend of reference. We submitted the top scored fuel blends predicted to experimental testing. The experimental results were coincident with computational simulations. In fact, different blends of standard fuel (90%v/v) with different pre-mixtures (10%v/v) comply with specifications. In addition, the C + + code of all the ANN models was released online for public research purposes at the GitHub repository: https://github.com/glezdiazh/biofuels.ptml. The new model may be useful for eco-friendly fuel blends design with lower environmental impact.