Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Fluorescence microscopy is a key technique for research in biology, medicine etc. Certain molecules, called fluorophores, emit light in a specific colour when they are illuminated. The miniaturisation of a fluorescence microscope onto a chip would make the technique more widely accessible. It would also make it suitable for applications that need fast and parallel analysis, like DNA sequencing. The functioning of a chip for fluorescence microscopy contains several interesting elements. Light from a laser is transported on the chip by waveguides, and these waveguides send light beams from different directions into a central region where an interference pattern forms. The fluorophores within the sample placed on top of the central region feel this light. They start to emit their own light that can be captured with a CMOS sensor, a chip similar to a normal camera. However, the captured light creates blurry pictures, because it is not focused like with normal lenses. Only illuminating specific spots instead of the full sample makes it possible to calculate the origin of the captured light, since we know which part was illuminated. These light spots can then be moved to scan over the entire sample and create a detailed image. The goal is now to design a chip with these abilities. Crossing waves create interference patterns, and interference patterns with bright spots at regular locations are called optical lattices. Not any optical lattice is suitable for structured illumination, it is important that the spots are bright enough and that there is enough distance between them. The integer lattice method, developed by D. Kouznetsov, provides a systematic way of investigating possible lattices, and it also defines how they can be created and translated. This led to the currently existing design of an optical lattice-generating chip. It is 5.6 mm x 5.6 mm, but a lattice is only in the central region of 100 μm x 100 μm generated. This is not efficient and therefore, an alternative way to design a compact chip is desired. A computer can simulate the pattern generated for a device and optimise this for a lattice. This process is called inverse design and produces compact, but irregular devices. Now, the optical lattice can not be translated any more to scan the sample. In this thesis, explainable AI is used to extract design concepts for a compact, regular optical lattice-generating chip. Shapley values, a form of explainable AI, originate from game theory and can indicate the regions in an image most relevant for a neural network. A neural network is a layered structure of mathematical calculations with tunable parameters able to perform very complex tasks. With inverse design, a large database of two classes of optical lattice-generating devices was created, and the neural network was trained on this database to distinguish devices generating different optical lattices. The Shapley values of the dataset with respect to this network were computed and analysed. Specific features similar to the elements of the integer lattice method were recognized. The Shapley values provided guidance towards a new design. This design is defined by mathematics rather than the seemingly random shapes of inverse design, and it is slightly more compact than inverse designs. Unfortunately, it has similar difficulties in fabrication and translating the optical lattice as inverse design, and further research is needed.