Deep Learning for Determination of Properties of Semidilute Polymer Solutions

The ability to predict polymer solution viscosity is essential for polymer characterization and processing. Here, we use a synergistic approach combining the scaling theory of polymer solutions and convolutional neural network (CNN) models to obtain system-specific parameters and describe semidilute solution viscosity in the unentangled and entangled solution regimes. The scaling approach relies on the existence of a characteristic microscopic length scale -the solution correlation length (correlation blob size) ?-which uniquely defines macroscopic solution properties. It is based on a relationship between the solution correlation length ? = lg(?)/B and the number of monomers g per correlation volume for polymers with the monomer projection length l. The system-specific set of parameters B-g, B-th , and 1 in the corresponding solution regime with the scaling exponent ? = 0.588, 0.5, and 1, respectively. Applying two CNN models, we obtained the sets B-g and B-th from the solution specific viscosity, ?(sp), as a function of concentration, c, and weight-average degree of polymerization, N-w . The CNN was trained on theoretically generated datasets converted to sparse images representing the normalized specific viscosity ?(sp)/N-w (cl(3))(1/(3?-1)) in the unentangled Rouse regime. The trained CNN was utilized in automated data analysis of the solution viscosity of polystyrene, poly(ethylene oxide), poly(methyl methacrylate), poly(acrylonitrile-co-itaconic acid), cellulose, sodium hyaluronate, hydroxypropyl methyl cellulose, methyl cellulose, hydroxypropyl cellulose, cellulose tris(phenyl carbamate), xanthan gum, galactomannan, and sodium ?-carrageenan in water, organic solvents, and ionic liquids. This approach produced values of the B-parameters with mean absolute percentage differences of less than 6% from the corresponding values determined by the manual data analysis. The B-parameters are then used to obtain the packing number P-e defining the onset of entanglements in polymer solutions and to describe semidilute solution viscosity as a function of concentration and N-w .

Automated summary

By combining the scaling theory of polymer solutions and convolutional neural network (CNN) models, specific parameters were obtained to describe semidilute solution viscosity in different solution regimes. The CNN models were trained on theoretically generated datasets and used to analyze the solution viscosity of various polymers in different solvents. The approach produced accurate results and can be used to predict viscosity based on concentration and polymerization degree.