Options
Applications of backpropagation neural networks in chemistry
Author
Hu, Fang
Supervisor
Teo, Khay Chuan
Abstract
Artificial neural networks (ANN), especially the backpropagation neural networks (BPNN) have recently attracted the interests of many researchers in the field of chemistry. The advantages of BPNN over classical methods have been proven, but BPNN methods still have many aspects worth to be studied such as to improve the performance of the algorithm in order to overcome some of its limitation.
The objectives of this work are to improve the neural network methods, to apply the BPNN in some selected chemistry areas, and to compare the neural network method with multivariate linear regression (MLR) method.
In-house developed programs were part of this study. The programs were developed using C language. Moreover, all of the programs were written and run on a Pentium III 450 PC with 64 Mbytes memory under a Window 98 operating system. They were proven to be successful.
BPNN have been used to predict six kinds of physical properties, namely heat capacity, boiling point, density, refractive index, Gibbs free energy and enthalpy of alkanes. The molecular distance-edge vector λ was used as input for the BPNN. It was noted that the BPNN prediction results for 25 alkanes are indeed better than the MLR results.
The BPNN method was also used to predict the 13C NMR chemical shifts for alkanes. From the experimental results obtained by the BPNN and MLR methods, it was noted that the BPNN calculation provides better prediction results than the MLR. The results indicated that there is a nonlinear dependence between the descriptors and the intended property, 13C chemical shift. Undoubtedly the use of BPNN is justified.
The performance of the BP algorithm is affected by the presence of outliers in the experimental data. In order to solve this problem a robust BPNN algorithm has also been developed. The developed robust BPNN demonstrated better predicted results even when the training data set contains 25% of outliers.
The objectives of this work are to improve the neural network methods, to apply the BPNN in some selected chemistry areas, and to compare the neural network method with multivariate linear regression (MLR) method.
In-house developed programs were part of this study. The programs were developed using C language. Moreover, all of the programs were written and run on a Pentium III 450 PC with 64 Mbytes memory under a Window 98 operating system. They were proven to be successful.
BPNN have been used to predict six kinds of physical properties, namely heat capacity, boiling point, density, refractive index, Gibbs free energy and enthalpy of alkanes. The molecular distance-edge vector λ was used as input for the BPNN. It was noted that the BPNN prediction results for 25 alkanes are indeed better than the MLR results.
The BPNN method was also used to predict the 13C NMR chemical shifts for alkanes. From the experimental results obtained by the BPNN and MLR methods, it was noted that the BPNN calculation provides better prediction results than the MLR. The results indicated that there is a nonlinear dependence between the descriptors and the intended property, 13C chemical shift. Undoubtedly the use of BPNN is justified.
The performance of the BP algorithm is affected by the presence of outliers in the experimental data. In order to solve this problem a robust BPNN algorithm has also been developed. The developed robust BPNN demonstrated better predicted results even when the training data set contains 25% of outliers.
Date Issued
2000
Call Number
QD39.3.M3 Hu
Date Submitted
2000