|Structure Input||3D-Structure Generation||Calculation of Properties||Structure Coding||Network Training||IR Spectrum Simulation|
The neural network learns the relationship between structure and spectrum by analyzing a set of pairs of molecules and spectra in a training phase. For each data point, a pair of structure and spectrum, two steps are performed:
Determining the most similar neuron
For each data point, the most similar neuron is determined by calculating the rms error between the structure code (red) of the training structure and the input (upper) block of the neural network.
Adjustment of the neuron weights
The weights of the winning neuron are adjusted to become more similar to the training data point. The weights of the neighbor neurons are also adjusted with the rate of adjustment decreasing with increasing distance from the winning neuron. The adjustment is applied to the structure (green) block and the spectrum (red) block of the network. This training step establishes the correlation between structures and spectra and stores it in the network.
A trained counterpropagation network acts as an interpolator and is able to predict a spectrum for a molecule it has never seen before.