Use Of ANNs Research Articles

A combination of dynamic measurement protocol and advanced data treatment to resolve the mixtures of chemically similar analytes with potentiometric multisensor system

Data processing techniques and measuring protocol are very important parts of the multisensor systems methodology. Complex analytical tasks like resolving the mixtures of two components with very similar chemical properties require special attention. We report on the application of non-linear (artificial neural networks, ANNs) and linear (projections on latent structures, PLS) regression techniques to the data obtained from the flow cell with potentiometric multisensor detection of neighouring lanthanides in the Periodic System of the elements (samarium, europium and gadolinium). Quantification of individual components in mixtures is possible with reasonable precision if dynamic components of the response are incorporated thanks to the use of an automated sequential injection analysis system. The average absolute error in prediction of lanthanides with PLS was around 1 × 10(-4)mol/L, while the use of ANNs allows the lowering of prediction errors down to 2 × 10(-5)mol/L in certain cases. The suggested protocol seems to be useful for other analytical applications where simultaneous determination of chemically similar analytes in mixtures is required.

A New Algorithm for Spatial Learning of Artificial Neural Networks Based on Lattice Models of Chemical Structures for QSAR Analysis

analysis of experimental data do not require any preset functional form. Moreover, this approach establishes relationships capable of revealing resemblances even between groups of data belonging to different types. Once established, these relationships allow predicting the properties of objects not entering into the initial group of data. In the field of chemical investigations, most widely used are networks employing a learning process based on forward signal propagation, called feed forward neural networks. Best known among the learning algorithms is that based on error back-propagation [9]. However, the ANN applications are mostly restricted to the analysis of qualitative signs [10, 11]. This is explained by the relatively low learning rate of the neural networks with a large number of neurons in the input layer, whereby direct use of ANNs for the analysis of substances possessing a large number of signs involves time-consuming computations. Since each sign in this approach is treated as independent, the analysis may encounter various correlation effects between signs; this additionally complicates the network learning process. In this study, we attempted at eliminating these problems by using a special algorithm for the analysis of data on the spatial molecular structures of compounds. This algorithm is essentially a combination of ANNs with error back-propagation learning and Kohonen self-organizing maps [12]. The principle of our method consists in initially determining certain cluster zones in the space around molecules, followed by using the corresponding averaged values for the ANN learning and predicting the activity of new compounds. In order to assess the quality of models found by the new algorithm, we employed the PLS method [5] used in the CoMFA data analysis. The computer code of PLS was realized in C++ as described in [13] and adapted using the Kubinyi recommendations concerning the CoMFA data analysis [14]. Verification of the proposed method was based on a comparison of the results obtained in this study to the PLS data reported by Kubinyi [15].