The Chilean physicist and luthier Sebastian Gonzalez (post-doc researcher) and the professional mandolin player Davide Salvi (PhD student), researchers at Dipartimento di Elettronica, Informazione e Bioingegneria’s Musical Acoustic Lab, have published on Nature Scientific Reports along with Prof. Augusto Sarti and Prof. Fabio Antonacci the study “A Data-Driven Approach to Violinmaking”, which shows how a simple and effective neural network is able to predict the vibrational behavior of violin plates. This prediction is obtained from a limited set of geometric and mechanical parameters of the plate.
The research shows how Artificial Intelligence, physical simulation and craftsmanship can all join forces to shed light on the art of violin making. The ability to predict the sound of a violin design can truly be a game changer for violin makers, as not only will it help them do better than the “grand masters”, but it will also help them explore the potential of new designs and materials.
Violins are extremely complex objects, and their geometry is defined by their outline, arching on the horizontal and vertical sections. The inspiration of this study came from a historical drawing on display at the “Museo del Violino” in Cremona. Politecnico di Milano researchers developed a model that describes the violin’s outline as the conjunction of arcs of nine circles. Thanks to this representation and an efficient model of the curvature of the plate, based on the renowned “Messiah” violin by Stradivarius, researchers were able to draw a violin plate as a function of 35 parameters.
By randomly changing such parameters, such as radii and center position of the circles, arching, thickness, mechanical characteristics of the wood, etc., they built a dataset of violins, which includes shapes that are very similar to those used in violin making, but also designs that had never been seen before. Such shapes constituted the input for the neural network. Advanced tools for the modeling of vibrations were used for characterizing the acoustic behavior of each violin in the dataset.
The next step was to understand if a simple neural network would be able to predict the acoustic behavior of a violin plate, starting from its parameters. The answer turned out to be positive, with an accuracy that came close to 98%, exceeding any expectations.
To read the full-text version of the article: https://www.nature.com/articles/s41598-021-88931-z