Novel Predictive Model for Determining Filtration Volume vs. Time for Nano Compounds with Multi-modal Particle Size Distribution.

摘要

A novel predictive model for accurately determining filtrate mass correlation with respect with the input variables. The solution is based on training a neural network consisting of ten hidden layers using the Levenberg- Marcquard back propagation algorithm to recognize the correlation between the input variables, including the filtration time, and the filtrate mass. The main hurdle found was how to exactly organize the data in order to get the best Mean Square Error (MSE) and correlation (R) values to ensure a good prediction and a strong relationship between the input variables and the output variable which in this case is the Filtrate Mass. This Model is proven to work with an estimated error of about 4.02% and 4.9% with a total sample size of 4974. Each sample consisted of a row containing 114 input variables, including the time for reaching the target Filtrate Mass, and one output variable which is the Filtrate Mass. Many Models were tried but the model with the best MSE and R values was the model in which the complete PSD was inserted as variables. I decided to use the PSD particle size as a variable name and the volume in percent as the value for the variable and include it in the input array to the neural network. The output array of the neural network consisted of the empirical data of the filtrate mass over time. This model will work for n-modal compounds since all the information from the PSD is already taken into consideration with the model.