In ellipsometry, the thickness and optical properties of thin films may be determined by light reflection, both in situ and ex situ. However, these useful film parameters are not measured directly but must be computed from the measured parameters (angles psi and Delta) using an appropriate system of equations derived from reflecting surface morphology. The most popular solving method, variably damped least squares (VDLS), is slow and troubled by local minima in the solution surface. Artificial neural networks (ANN) avoid these problems but have not been trained to be better than about 5% accurate. The work presented here demonstrates a cascade of three ANN levels exhibiting a typical overall accuracy better than 0.4% with speed orders of magnitude faster than that of VDLS methods. This ANN cascade for a material system contains 1800 three-layer perceptrons with ten hidden neurons, each requiring 70 weights plus 20 range variables and other statistics which comes to 16.2 MB for 100 wavelengths. Assuming the ranges of film material optical properties can be fitted into 30 such 'subranges' yields 486 MB uncompressed. This data can easily fit onto a CD-ROM in the form of a semantic database for efficient storage and selective retrieval.