A deep neural network (DNN) and Gaussian process regression (GPR) model are trained on the Polystyrene Cloud Points Dataset.1 All model details, data preprocessing, and training protocols can be found in Ref. 1. Both models output polystyrene cloud point temperatures as a function of volume fraction for select solvents.
Many solvents have very little training data and may not extrapolate to other Mw well. For instance, 1-dodecanol has 1 reported Mw. Predictions around this Mw are expected to be accurate, but will deteriorate as the Mw input deviates significantly from the experimental Mw. Note that curve shape is highly dependent on data quality and amount of available data in the low concentration regime. Hence, some curves may appear to be concave up (instead of concave down) when predicting upper critical solubility (UCS), and vice versa for lower critical solubility (LCS), depending on the amount of data available for the solvent at low concentrations. The models will be updated as new cloud point data is processed.
The code generates a binodal curve for a binary mixture of polystyrene in a select solvent. Required inputs are: solvent, polymer Mw, volume fraction range, polydispersity, and pressure. Number of volume fractions is not required (default is 100) and determines the number of predicted values between φmin and φmax (larger numbers result in smoother curves). Predictions can be made using the neural network (NN) or Gaussian process regression (GPR) model.
LCS - include the upper-bound lower critical solubility (one phase region found with decreasing temperature) in addition to the upper critical solubility (one phase region found with increasing temperature) cloud points.
Find Nearest Experimental Data (by Mw) - search Polystyrene Cloud Points Dataset for experimental data closest to input Mw value.