Information Technology Reference
In-Depth Information
Fig. 1.41.
Molecules that have chemical idiosyncrasies, whose properties may be
poorly predicted by neural networks
of their parsimony, neural networks of very small size (5 to 7 hidden neurons)
provide better results, on the same databases, than multilinear regression
techniques that are used traditionally in the field [Duprat 1998].
Interestingly, the values of logP of some molecules were systematically
either poorly learnt (when those molecules were in the training set) or poorly
predicted (when present in the test set). In such a situation, one should first
be suspicious of a measurement or typing-in error. If such is not the case, then
one should conclude that the molecules have idiosyncrasies that are not shared
by the other examples; in the present vase, it turns out that the molecules
of interest are either strongly charged (tetracycline and caffeine, shown on
Fig. 1.41), or, by contrast, interact very weakly with the solvent (perylene, 1-4
pentadiene, see Fig. 1.41). Thus, neural networks are able to detect anomalies;
anomaly detection is actually one of the main areas of applications of neural
networks.
1.4.7 An Application in Formulation: The Prediction of the
Liquidus Temperatures of Industrial Glasses
In the same spirit as the previous application, albeit in a completely different
field, thermodynamic parameters of materials can be predicted as a function
of their formulation. Of specific interest is the prediction of the liquidus tem-
peratures of oxide glasses. That temperature is the maximal temperature at
which crystals are in thermodynamic equilibrium with the liquid; the predic-
tion of the liquidus temperature is important for the glass industry, because
the value of the viscosity at the liquidus temperature has a strong impact
Search WWH ::
Custom Search