Accelerating Mems Design Process Through Machine Learning from Pixelated Binary Images

This paper reports the use of machine learning in accelerating the MEMS design process. Candidate designs are represented by pixelated binary 2D images. Instead of common computational tools like FEA, we use trained neural network for quickly obtaining physical properties of interest for each candidate design. Circular disk resonators are used as an example to demonstrate the capability of our method. After sufficient training with 9000 images, the resulting neural network can serve as a high-speed, high-accuracy analyzer: it can identify four vibrational modes of interest and calculate the corresponding frequencies 4000 times faster than commonly used FEA software, with remarkable accuracy (~98%).