Neural networks in mining molecular properties from tabulated data

  • Authors:

  • W. Bradford Davis;Ray Hefferlin

  • Affiliations:

  • Davis Research Group, Rio Linda, California and Physics Department, Southern Adventist University, Collegedale, Tennessee;Davis Research Group, Rio Linda, California and Physics Department, Southern Adventist University, Collegedale, Tennessee

  • Venue:
  • ICCOMP'05 Proceedings of the 9th WSEAS International Conference on Computers
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

We produce global estimates of data for vibration frequencies of main-group diatomic molecules for the first time, using a very sophisticated-neural network program designed for forecasting financial markets. Input vector selection, frequency compensation, genetic-algorithm variable selection, hidden nodes, multiple transfer functions, Kalman filtering, and a root-mean power evaluation function are used to produce 56 models. Of these, an optimal number of 23 models is selected. Mean error measures over the 23 networks are determined for all molecules and 181 outliers are deleted. The number of predictions per molecule varies from 18 to 23; and the average of the 99% confidence limits of these predictions is 9.04%. To test these estimates, data for 116 molecules were gleaned from the literature. The average 1/2-spread in these literature data for molecules is approximately 4.4%. 90 of our predictions agreed with the literature values to within the literature 1/2-spread plus the predictions' 99% confidence limits. 18 more, a total of 108, of them were outside of the sum of the two error measures by less than 50% of the prediction magnitude.