Algorithmic design of wind instrument shape via 3D FDTD and deep learning

Abstract

Traditional design of wind instruments centers around simple shapes such as tubes and cones, whose acoustic properties are well understood and are easily fabricated with traditional manufacturing methods. The advent of additive manufacturing enables the realization of highly complex geometries and new wind instruments with unique sound qualities. While simulation software exists to predict the sound of wind instruments given their shape, the inverse problem of generating a shape that creates a desired sound is challenging given the computational cost of 3D acoustic simulations. In this work we create a fast 3D acoustic wind instrument simulator using GPU acceleration. In addition, we use deep learning to solve the inverse problem of generating a 3D shape that roughly approximates a desired sound when played as a single-reed instrument. Finally we develop an automatic method for determining pitch hole locations for a given shape to generate playable instruments.