Fragrance Product Design With A Hybrid Machine Learning And Mechanistic Modeling Approach

Fragrance is a kind of products that are closely related to human social life,and it is capable of improving products' quality and further their competitiveness.Currently,fragrance design mainly relies on human experience and experimental methods,which costs a plenty of time,manpower and money.In recent years,product engineering methods,such as Computer-Aided Molecular Design(CAMD),are applied among thousands of chemical-based products.With mechanistic physicochemical property models,CAMD is able to design candidate products according to requirements of customers,and it therefore shortens the R&D period of products like fragrance.On the other hand,Machine Learning(ML),represented by Artificial Neural Network(ANN),is attracting increasing attentions because it could model Structure-Odor Relationship(SOR)and further predict odors.As a result,a hybrid approach of ML-based odor prediction model and mechanistic physicochemical property models is proposed in this thesis to design fragrance,including fragrance molecule(aroma)design and mixed fragrance design.Main contents of this thesis are concluded as following:(1)A COSMO-based descriptor of molecular surface charge distribution is introduced to represent the molecular structure and correlated with odor via ML.Then a framework of MLbased SOR is proposed to build ML model for odor prediction.Within the framework,a greedy parameter tuning strategy is employed towards the tricky difficulty during the establishment of ANN.The strategy is able to improve the efficacy of model while avoiding model's over-fitting.Lastly,the performances of built ML models are verified,in order to guarantee that models could be used in fragrance design.(2)A Computer-Aided Aroma Design(CAAD)framework,which integrates ML,is established in this chapter.Firstly,requirements of aroma are determined,which are further converted into physicochemical properties.Then aroma design is abstracted as a MINLP problem consisting of ML and mechanistic models.Finally,a decomposition-based algorithm is used to solve the problem.In the first step,the mathematical programming consists of structural constraints and linear property constraints are solved;in the second step,non-linear property constraints are solved based on solutions of last step;in the last step,ML is utilized to predict odors of solutions in step two,and candidate aromas are selected if their predicted odors are satisfied.Here,candle aroma is regarded as the case study to testing the CAAD framework.After molecular design and verification,five molecules are found as aromas appearing in the normal life,that demonstrates the practical value of the framework.(3)A ML-based Computer-Aided Mixed Fragrance Design(CAMFD)framework is proposed in this chapter.Firstly,requirements of products and their corresponding physicochemical properties are decided.ML is then employed to predict descriptors of mixed fragrance based on target properties.Finally,Generate-And-Test method is applied.In the first step,according to predicted results,candidate mixed fragrances are screened and generated within the database.And in the second step,mechanistic models are used to test properties of mixtures' components.The framework is employed to design substituted mixed fragrance of cis-3-hexenyl propionate.With the verification of database and experiments,two mixtures are obtained finally which are similar to the target.The result highlights the effectiveness of employed framework.