Accurate modeling of boundary conditions is crucial in computational physics. The ever increasing use of neural networks as surrogates for physics-related problems calls for an improved understanding of boundary condition treatment, and its influence on the network accuracy. In this paper, several strategies to impose boundary conditions (namely padding, improved spatial context, and explicit encoding of physical boundaries) are investigated in the context of fully convolutional networks applied to recurrent tasks. These strategies are evaluated on two spatio-temporal evolving problems modeled by partial differential equations: the 2D propagation of acoustic waves (hyperbolic PDE) and the heat equation (parabolic PDE). Results reveal a high sensitivity of both accuracy and stability on the boundary implementation in such recurrent tasks. It is then demonstrated that the choice of the optimal padding strategy is directly linked to the data semantics. Furthermore, the inclusion of additional input spatial context or explicit physics-based rules allows a better handling of boundaries in particular for large number of recurrences, resulting in more robust and stable neural networks, while facilitating the design and versatility of such types of networks. 1
New data and review of the spanwise coherence length is provided for flows over cylinders of different cross-sections: circular of diameter d, and rectangular of sectional aspect ratios (breadth (b) to height (d) ratio AR = b/d) of 1, 2 and 3. In the present measurements, the body has both d and spanwise length of 70d fixed, and the Reynolds number (based on d) range 6000–27,000 is covered. Two-point data are obtained from two hot-wire probes, one fixed in the symmetry plane and the other moving on the corresponding spanwise axis. Their position in a cross plane are deduced from preliminary measurement of the mean flow with a single probe, allowing fair comparisons between the different geometries and the introduction of uncertainty bars on coherence length values. At all tested regimes, a very good agreement is noticed between velocity-based and pressure-based coherence experimental data. Coherence length definitions are revisited, and the aeroacoustically consistent, integral length definition is selected, allowing fair synthesis of literature data into a single chart and empirical functions. Definitions for coherence decay models (e.g. Gaussian or Laplacian) are also adapted so that coherence length and coherence integral shall be equivalent. This preliminary work on coherence data and its spanwise integration enables transparent regressions and model selection. Generally, the Gaussian model is relevant for the lift peak, while the coherence exhibits a Laplacian decay at harmonics. On average, at peak Strouhal number, the coherence length for the circular and square cylinders is of 5d while it is of the order of 15d for the rectangular sections. It is concluded that the flow over those latter geometries is still a two-dimensional dynamics at the tone frequency. These values are almost preserved over the tested Reynolds number range. Coherence length value at harmonics is extensively documented. Spanwise coherence length is also discussed as an ingredient of acoustic efficiency.
Localization and quantification of noise sources is an important scientific and industrial problem, the use of phased arrays of microphones being the standard techniques in many applications. Non-physical artifacts appears on the output due to the nature of the method, thus, a supplementary step known as deconvolution is often performed. The use of data-driven machine learning can be a candidate to solve such problem. Neural networks can be extremely advantageous since no hypothesis concerning the environment or the characteristics of the sources are necessary, different from classical deconvolution techniques. Information on the acoustic propagation is implicitly extracted from pairs of source-output maps. On this work, a convolutional neural network is trained to deconvolute the beamforming map obtained from synthetic data simulating the response of an array of microphones. Quality of the estimation and the computational cost are compared to those of classical deconvolution methods (DAMAS, CLEAN-SC). Constraints associated with the size of the dataset used for training the neural network are also investigated and presented.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.