Predicting Intake Valve Deposits: A Joint QSAR Project Between LZ and Purdue University Employing Neural Networks and First Principles Modeling

Dan T. Daly (The Lubrizol Corporation, 29400 Lakeland Blvd, Wickliffe, OH 44092, V. Venkatasubramanian, Anantha Sundarum)

Anne. M. Chaka (James Carruthers, Department of Chemical Engineering, Purdue University, Wast Lafayette, IN 47907)



A QSAR study was conducted for the prediction of Intake Valve Deposit
(IVD) formation in a BMW engine test developed for gasoline quality
evaluation.  The analysis performed were: Stepwise Linear Regression,
neural network models using descriptors from Stepwise Linear
Regression analysis as inputs, neural network models using inputs from
Principal Component Analysis and neural network models using inputs
from a Partial Least Square Analysis.  The best neural network based
on the root mean squared error is one using four principal components
as input features and three hidden neurons.  This model was an
improvement over the accuracy of previous linear and neural net
models.  In an attempt to reduce this error even further, workers from
Lubrizol and Purdue calculated a set of 'meta descriptors', which are
based on current theories of IVD formation and yielded a neural net
model derived on first principles.