Concept of a neural system for real-time evaluation of spectroscopic measurements

A hardware implementation of a Backpropagation feedforward neural network has been designed. The tool was proposed for reflectometric measurements integrated together with photosensor arrays. The intelligent reflectometric sensor is being implemented in a multi-chip-module approach. A logarithmic input transformation is applied for easing the misalignment and parameter scatter correction. It also allows for easy ratio calculation by subtraction for normalization with the reference value. The neural network was designed for complexities up to 100 inputs, 30 hidden neurons and 5 outputs. The digital building blocks (neurons) utilize a logic approximation of the sigmoid nonlinearity and the possibility of weight scaling. These hardware solutions result in a simultaneous area reduction and speed gain, at the cost of slightly decreased performance. Simulations of the proposed neural system prove applicability for evaluation of optical measurements were performed for reflectometric and ellipsometric data thin porous layers. Hardware simulations showed good correspondence to the optimum-case neural software simulations.