(763b) A COSMO-Based Inverse Machine Learning Application for Mixture Product Design

Modern society pays unprecedented attention to chemical products for both environmental impact and human-centered developments. Generally, chemical-based products can be categorized into single species product, multiple-species, and devices. Multiple-species products such as formulated mixtures and blends are one of the most widely used products due to their tailor-made, controlled and adjustable properties for satisfying product requirements. The common design methods for such kinds of products are based on trial-and-error experimentations. However, when the available options are many or when sufficient data are not available, methods based on property models are suggested [1]. For example, optimization-based mathematical programming methods are developed to integrate target property models into a MINLP/MILP model to obtain candidate ingredients [2]. However, there are a spate of physicochemical properties for which models of acceptable accuracy or application range, such as odor or color, are still not available. For the purpose of saving time, money and human resources, use of machine learning (ML) for modelling such properties, is an option worth considering. On the other hand, even though ML is regarded as a promising method for formulating property models, it cannot be incorporated into the product design methodology directly because of its highly non-linear characteristics. However, ML models can be utilized for product design through a decomposition-based approach, in which ML is responsible for verifying feasible molecules obtained from the established MINLP model [3].

Inspired by the well-known optimization-based Computer-Aided Molecular Design (CAMD) approach [4], an inverse application framework of machine learning for mixture product design is proposed in this study. First, a ML model is established to verify the correlation between structural descriptors of the product and its target properties. Specifically, the Artificial Neural Network is utilized to formulate the ML model, and σ-profile spectra descriptors generated from COnductor-like Screening MOdel (COSMO) [5] are employed as the structural descriptors. Such descriptors of a product are represented by ten S_σ-profile descriptors, which are the integrations of ten areas from one σ-profile spectrum. Next, the Inverse Neural Network is established where the inputs are target properties while the outputs are structural descriptors. Notably, since the number of inputs for the design phase is often less than the outputs, multiple ML models are established, and each model is responsible for predicting one S_σ-profile descriptor. Afterwards, all the established ML models are employed to design the ten S_σ-profile descriptors for potential products based on the target properties. Since the COSMO spectrum of a mixture is equal to the linear combination of COSMO spectra of its ingredients, potential combination of ingredients is screened out within an assigned database [6] and further selected through a statistical tool based on Euclidean distance. Finally, this framework is demonstrated using a fragrance design case study, and the results are verified by using experimental data from the literature [6].

ACKNOWLEDGEMENTS:

The authors gratefully acknowledge the financial support from “Natural Science Foundation of China” (No. 21808025) and “the Fundamental Research Funds for the Central Universities DUT17RC(3)008”.

REFERENCES:

[1] Zhang, L., Fung, K. Y., Wibowo, C., Gani, R. (2017). Advances in Chemical Product design. Reviews in Chemical Engineering, 2017, 34(3).

[2] Zhang, L., Kalakul, S., Liu, L., Elbashir, N. O., Du, J., Gani, R. (2018). A Computer-Aided Methodology for Mixture-Blend Design. Applications to Tailor-Made Design of Surrogate Fuels, Industrial & Engineering Chemistry Research, 57(20), 7008-7020.

[3] Zhang, L., Mao, H., Liu, L., Du, J., Gani, R. (2018). A machine learning based computer-aided molecular design/screening methodology for fragrance molecules. Computers & Chemical Engineering, 115, 295-308.

[4] Gani, R., Nielsen B., Fredenslund, A. (1991). A group contribution approach to computer-aided molecular design. AIChE Journal, 37, 1318-1332.

[5] Zhou, T., Mcbride, K., Zhang, X., Qi, Z., Sundmacher, K. (2015). Integrated Solvent and Process Design Exemplified for a Diels-Alder Reaction. AIChE Jounal, 61(1), 147-158.

[6] Keller, A., Vosshall, L. B. (2016). Olfactory perception of chemically diverse molecules. BMC Neuroscience, 17(1), 55.