A temperature-constrained training algorithm has been devised for back
propagation neural networks. The use of a temperature constraint softe
ns the neural network modeling. This algorithm furnishes more stable m
odels and may decrease training time substantially. These networks are
resistant to overfitting and avoid the problem of overtraining. Tempe
rature-constrained and conventional backpropagation networks are evalu
ated with a synthetic data set and mass spectra from compounds whose m
olecular formulas are C8H18O. Confusion matrices are used for evaluati
ng classification performance.