Genetics Optimizes Thin-Film Filters
Daniel S. Burgess
How do you design an optical element when you don't know what it should look like? Simple, say researchers at Fraunhofer Institut für Schicht- und Oberflächentechnik in Brunswick, Germany. Let evolution do the work. Using competing genetic algorithms to produce the design, they have produced a thin-film filter for a laser projection screen that balances optical performance and manufacturing concerns.
Completing genetic algorithms design optimal reflective coatings for laser projection screens (right). The algorithms do not require foreknowledge of the required characteristics of the filter to produce an optimal design, fabricated using a vertical in-line sputtering system (left). Courtesy of Christoph Rickers.
The computer programs now used to design optical thin-film filters minimize the deviation from a specified value, explained Christoph Rickers, a member of the research team. To produce an optimized filter, the designer must be able to quantify the desired spectral distribution of its reflectance at particular wavelengths and to specify the material used, the maximum number of layers and the maximum thickness. Sometimes that is not possible -- but neither is it necessary with genetic algorithms.
"Our implementation is capable of optimizing based on the features or characteristics of a needed reflectance spectrum, rather than a fixed value," Rickers said. "Thus, it represents a new approach, which might, in addition to the available solutions, help to find better solutions for specific applications."
The scientists began their demonstration by considering how the final filter should generally behave. They presumed that it should display improved contrast compared with a white screen, minimize color flop, and employ as few layers as possible to enable deposition on a greater variety of substrates and to reduce production costs. They quantified these features to yield fitness functions, with which it is possible to gauge the relative merits of different designs.
To determine the best combination of design properties, they employed a selective evolutionary algorithm that produced variations in an initial population of designs and that evaluated the resulting ones according to the fitness functions. Crossover and median recombination operators created novel designs based on the properties of the parents, and a mutation step randomly modified a property of the resulting children. Suboptimal designs were reinserted into the chain of operators until they yielded an acceptable set of properties, and the best designs were used to sire another generation.
In the experiment, the software suggested a nine-layer design based on the researchers' choice of SiO2 and Si3N4, for production reasons. They predicted that the design would offer a maximum contrast improvement of 2.75. The actual filter, which they deposited using reactive DC-midfrequency magnetron sputtering and in situ monitoring by spectroscopic ellipsometry, displayed a contrast improvement of 2.5 and should be suitable for deposition on highly flexible plastic substrates.
Although the software shows promise, it is not yet ready for prime time, Rickers said. "Right now, it's just a tool for us, used in this specific project. However, if we find more useful applications for this product, we might rewrite the software for commercial use."
In the meantime, the researchers are working to scale up the process to enable them to deposit the filters on larger substrates. In the initial experiments, they used 5 x 5-cm substrates, but they since have deposited filters on 30 x 30-cm ones, he said, using an in-line sputter-coating plant that can accept 100 x 60-cm substrates.
MORE FROM PHOTONICS MEDIA