Application of Deep Neural Network in the Design of Microresonators Subject to Microfabrication Constraints

Deep Neural Networks (DNNs) have transformed various fields. However, their applications in the design of micro-electromechanical systems (MEMS) have been limited. Here, we demonstrate a data driven approach, for the application of current state of art DNNs in design of passive microresonators targeting for operation in MHz frequencies. The dimensions of the microresonators are constrained based on a photolithography process on SOI wafers followed by dry-etching the Si device and the oxide layers. Accordingly, the design simplifies to single 2D layout representing the photolithography of the photoresist etch mask. We aim to find MEMS layouts for designing microresonators with minimal human interaction in the design process. For this purpose, we integrated an end-to-end machine learning model in TensorFlow with a finite element analysis model in ANSYS. The methodology integrated generation of training/validation/test data using finite element analysis, classification of the data, training DNNs specialized for predicting specific design criteria, and training generative networks in order to produce new designs. The fabrication constraints were incorporated in the finite element analysis, as well as the algorithm for the classification of the training/validation/test data. Upon achieving the datasets, DNNs with various architectures were trained for the prediction and classification of the resonant frequencies and errors associated with the given designs. The fabrication errors included unconstrainted parts and any feature size below the photolithography limits. Upon the completion of an end-to-end machine learning model, we assessed the influence of various parameters, such as the DNNs’ architectures, activation functions, and the training algorithm on critical metrics such as accuracy, loss, and the overfitting of the models. In order to validate the models, we compared the training, validation, and test results; additionally, the model was tested for the classification of microresonators that were specially designed by human. Upon training the DNNs specialized for the classification of a microresonator design, they are employed for training generative neural networks in a multi-output classification network. The generative networks were trained such that they could find a design that featured desirable values for metrics such as resonant frequencies or areas. The performance of the microresonators, designed by the generative neural networks, were then verified by the finite element model, and their properties were discussed based on the common design of mechanical microresonators according to the existing literature. This work suggests a multidisciplinary approach for leveraging the current advancements in the field of machine learning in research on microresonators, and eventually the route towards a data driven design of multi-layer multifunctional microsystems.