Study On Machine Learning And Its Applications In Liquid Viscosity Prediction

With the rapid development of machine learning in recent years,it has been widely studied in natural science and engineering fields.In the field of molecular property prediction,molecular representation learning is one of the current promising research directions,which can achieve endto-end molecular property prediction based on molecular structure.The viscosity of organic liquids is closely related to the transport phenomena in chemical engineering processes,and it is crucial for the screening of solvents,simulation of chemical processes,and scaling up of equipment.In this paper,we developed a series of machine learning models and molecular representation learning models for viscosity prediction and molecular representation learning.The main work of this paper is as follows:(1)A transition state theory-inspired neural network(TSTiNet)is constructed for viscosity prediction of deep eutectic solvents(DESs).The model utilizes theoretical equations to describe the viscosity-temperature and viscosity-molar ratio relationships and neural networks to calculate equation parameters.The proposed TSTiNet was able to achieve an average absolute relative deviation(AARD)of 6.84%and an R2 of 0.9805,which is significantly better than other existing viscosity prediction models for DESs.(2)Considering the importance of molecular structure information with different granular sizes for molecular representation,a plug-and-play graph transformation layer(LineEvo)in graph neural networks(GNNs)was constructed,which transforms fine-grained molecular structure information into coarse-grained one based on the idea of line graphs.On the nine molecular property benchmark datasets,GNNs containing the LineEvo layer were able to significantly outperform the original GNNs.In addition,the relationship between the LineEvo layer and the kWL test was analyzed to theoretically explain the reason for the ability of LineEvo to enhance the expressive power of the original GNNs.(3)To include all intramolecular interactions in GNNs,a molecular force field-inspired neural network(FFiNet)is proposed.FFiNet describes interatomic interactions based on the functional form of molecular force field and updates node representations according to the interactions.FFiNet finally achieves the best performance on eight benchmark datasets of molecular properties.In addition,FFiNet was able to accurately predict protein-ligand affinities and significantly outperformed other existing models.The final FFiNet visualization results show the interpretability of FFiNet.(4)To enable FFiNet to generate temperature-dependent molecular representations,a Boltzmann layer based on statistical thermodynamics was constructed,which describes molecular representations as a function of temperature using the Boltzmann distribution.Combining FFiNet with Boltzmann layers,an OrVNet model was constructed,which was able to achieve an AARD of 8.08%and R2 of 0.9500 on a large organic liquid viscosity dataset,and significantly outperformed other existing viscosity prediction models.Subsequently,OrVNet was applied to the viscosities of biofuels and binary organic mixtures as well as 17 other transport and thermodynamic properties,and the results showed that the model continued to achieve the best predictions on these datasets.