Hard-constrained neural networks for modelling nonlinear acoustics

Abstract

We model acoustic dynamics in space and time from synthetic sensor data. The tasks are (i) to predict and extrapolate the spatiotemporal dynamics, and (ii) reconstruct the acoustic state from partial observations. To achieve this, we develop acoustic neural networks that learn from sensor data, whilst being constrained by prior knowledge on acoustic and wave physics by both informing the training and constraining parts of the network's architecture as an inductive bias. First, we show that standard feedforward neural networks are unable to extrapolate in time, even in the simplest case of periodic oscillations. Second, we constrain the prior knowledge on acoustics in increasingly effective ways by (i) employing periodic activations (periodically activated neural networks); (ii) informing the training of the networks with a penalty term that favours solutions that fulfil the governing equations (soft-constrained); (iii) constraining the architecture in a physically-motivated solution space (hard-constrained); and (iv) combination of these. Third, we apply the networks on two testcases for two tasks in nonlinear regimes, from periodic to chaotic oscillations. The first testcase is a twin experiment, in which the data is produced by a prototypical time-delayed model. In the second testcase, the data is generated by a higher-fidelity model with mean-flow effects and a kinematic model for the flame source. We find that (i) constraining the physics in the architecture improves interpolation whilst requiring smaller network sizes, (ii) extrapolation in time is achieved by periodic activations, and (iii) velocity can be reconstructed accurately from only pressure measurements with a combination of physics-based hard and soft constraints. In and beyond acoustics, this work opens strategies for constraining the physics in the architecture, rather than the training.