A great many approaches to problems from diverse branches of chemistry have already been investigated by applying neural networks - from very general tasks such as structure-spectrum relationships, chemical reactivity, secondary and tertiary structures of proteins and process monitoring, to such specific questions as the classification of energy levels in the curium atom and the recognition and classification of aerosol particle distributions, as well as the relationship between the physical structure and the mechanical properties of polyethylene-terephthalate fibers.

This diversity underlines the fact that neural networks represent quite general, broadly applicable problem-solving methods. The final, concrete application is determined only by the nature of the data fed into the neural network.

This article is intended to convey some feeling for the kinds of problems best approached with neural networks, so that the reader will be able to decide whether these methods can be used for his particular problem. The goal is therefore to demonstrate the capabilities of the methods and to give incentive to further applications, rather than to give a comprehensive overview of all the work that has gone before. For this reason only a selection of typical neural network applications in chemistry will be given in the following sections; these will not be organized by application area but rather by problem type, that is, according to whether the task in question is one of classification, modeling, association, or mapping.

There is, of course, a whole series of methods that can be used as alternatives to neural networks - and which have been used successfully for many years. Many tasks for which neural networks are used today could be solved equally well by using statistical and pattern-recognition methods such as regression analysis, clustering methods, and principal component analysis. In many cases these methods should be able to deliver results that are every bit as good. What is unfortunately lacking in most papers is a comparison between the capabilities of neural networks and those of more established methods. Comparisons such as these might bring out the advantages of neural networks more. These advantages include the following:

- |
the mathematical form of the relationship between the input and output data does not need to be provided |

- |
neural networks are also able to represent nonlinear relationships. |

On the other hand, the application of neural networks requires exactly the same care in the formulation of the problem, the representation of information, the selection of data, and the division of data into training and test sets as is necessary for pattern-recognition and statistical methods. It must be stated clearly that the quality of the results obtained from neural networks depends crucially on the work and trouble invested in these subtasks.

Essential aspects must be considered before deploying any kind of neural network.

- |
What is the nature of the task under consideration? Classification, modeling, association, or mapping? |

- |
Which kind of learning process should be chosen? Supervised or unsupervised learning? Is there a set of expected results available for the objects, or must the structure of the information first be found? |

In Table 1 the various methods are correlated with problem types and learning processes to make it easier to decide which neural network method should be used for a specific task.

Table 1. Possibilities for applying neural networks.

Hopfield |
ABAM |
Kohonen |
Back- | |

classification |
x |
x |
x | |

modeling |
x | |||

association |
x |
x |
x | |

mapping |
x |
|||

learning |
unsup. |
unsup. |
unsup. |
sup. |

process [a] |
+ sup. |

[a] unsup. = unsupervised learning, sup. = supervised learning (see Section 4).

In chemistry most problems are of the classification or modeling types. This is one of the reasons why multilayered networks trained by the back-propagation algorithm predominate. The results of a survey of those neural network applications in chemistry that appeared before the end of 1990 show that over 90 % of studies used the back-propagation algorithm [2]. Hopfield networks were used twice, an adaptive bidirectional associative memory only once, and a Kohonen net coupled with the counterpropagation algorithm likewise only once. This distribution need not hold true in future, however, since ABAM and Kohonen networks as well as the counterpropagation algorithm all offer possibilities that have by no means been exhausted.

If we consider only the multilayered model with the back-propagation algorithm, we still have a broad spectrum for the complexities of the networks used: from networks with only 20 weights up to some with 40000, in one case even 500000 weights! The number of data used for training a net should, in the case of the back-propagation algorithm, be at least as large as the number of weights. This rule was by no means always followed. The number of times the training data-set was passed through the net (number of epochs) also varied considerably: from 20 up to 100000 iterations. As may be imagined, the training times increased greatly in proportion with the number of weights and the number of iterations, from a few minutes on a personal computer up to hours on a Cray supercomputer.