Design to Acoustics Through Deep Learning

Numerical methods are the standard approach for acoustic simulation (e.g. finite element method). However, they are computationally expensive. Neural network-based surrogate models can increase the speed and provide gradients for design optimization.

• The FQO method can make accurate prediction while accelerating computation several magnitudes (after training) [1].
• We can control the design process using Guided Flow Matching and show that this method compares favourably to existing design optimization methods [2].

[1] van Delden et al.: Learning to predict structural vibrations, NeurIPS 2024

[2] van Delden et al.: Minimizing Structural Vibrations via Guided Flow Matching Design Optimization, under review 2025

In project phase 1 we investigated beading plate structures, that can be modelled as two dimensional projections. Many real-world use cases require 3D structures.

• How do neural network-based surrogate models evolve as the complexity increases?
• Which neural network architecture are necessary to handle more complex 3D structures?
• Several architectures were discussed in computer vision and graphics (like voxel grids, triplanes or point clouds). An important aspect is that the representation should be non-parametric, i.e. use variable number of parameters depending on the complexity of the structure. This will allow to represent a broad spectrum of structures.

Having developed flexible models for complex structures, we could benefit from transfer learning. Can we improve performance by initializing with existing network weights obtained from a similar problem?

In practice, a designer has to find a tradeoff between multiple criteria. While our design optimization methods used only a single criterion in the past, we will investigate the combination of gradients from multiple surrogate models. While this is straightforward in the guided flow matching framework, we expect properly weighting and scaling  to be challenging.