Physics-Informed Neural Networks for Mesh Deformation with Exact Boundary Enforcement

Date

2023-03-03

Author

Aygün, Atakan
Karakuş, Ali

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

196
views

0
downloads

In this work, we have applied physics-informed neural networks (PINN) for solving mesh deformation problems. We used the collocation PINN method to capture the new positions of the vertex nodes while preserving the connectivity information. We use linear elasticity equations for mesh deformation. To prevent vertex collisions or edge overlap, the mesh movement in this work is conducted in steps with relatively small movements. For moving boundary problems, the exact position of the boundary is essential for having an accurate solution. However, PINNs are frequently unable to satisfy Dirichlet boundary conditions exactly. To overcome this issue, we have used hard boundary condition enforcement to automatically satisfy Dirichlet boundary conditions. Specifically, we first trained a PINN with soft boundary conditions to obtain a particular solution. Then, this solution was tuned with exact boundary positions and a proper distance function by using a new PINN considering only the equation residual. To assess the accuracy of our approach, we used the classical translation and rotation tests and compared them with a proper mesh quality metric considering the change in the element area and shape. The results show the accuracy of this approach is comparable with that of finite element solutions. We also solved different moving boundary problems, resembling commonly used fluid-structure interaction problems. This work provides insight into using PINN for mesh-deformation problems without needing a discretization scheme with reasonable accuracy.

URI

https://hdl.handle.net/11511/102637

Conference Name

SIAM Conference on Computational Science and Engineering 2023

Collections

Department of Mechanical Engineering, Conference / Seminar

Suggestions

OpenMETU
Core

Physics informed neural networks for computational fluid dynamics Aygün, Atakan; Karakuş, Ali; Maulik, Romit; Department of Mechanical Engineering (2023-1-17) In this work, we use physics-informed neural networks (PINN) to model the problems generally used in computational fluid dynamics (CFD). Since PINN does not need any discretization scheme or mesh generation, this approach is easier to implement compared to conventional numerical methods. In the context of CFD problems, the effect of network parameters is shown in the convergence and the accuracy of the solution. The solutions to forward problems are chosen with simple geometries since as the nonlinearity in...
PHYSICS INFORMED NEURAL NETWORKS FOR TWO DIMENSIONAL INCOMPRESSIBLE THERMAL CONVECTION PROBLEMS Aygün, Atakan; Karakuş, Ali (2022-10-01) Physics-informed neural networks (PINNs) have drawn attention in recent years in engineering problems due to their effectiveness and ability to tackle problems without generating complex meshes. PINNs use automatic differentiation to evaluate differential operators in conservation laws and hence do not need a discretization scheme. Using this ability, PINNs satisfy governing laws of physics in the loss function without any training data. In this work, we solve various incompressible thermal convection probl...
Binarized Weight Networks for Inverse Problems Ozkan, Savas; Becek, Kadircan; Inci, Alperen; Kutukcu, Basar; Ugurcali, Faruk; Kaya, Mete Can; Akar, Gözde (2020-01-01) In this paper, we present a binarized neural network structure for inverse problems. In this structure, memory requirements and computation time are significantly reduced with a negligible performance drop compared to full-precision models. For this purpose, a unique architecture is proposed based on a residual learning. Precisely, it opts to reconstruct only the error between input and output images, which is eventually centralized the responses around zero. To this end, this provides several advantages fo...
ACOUSTIC SOURCE SEPARATION USING THE SHORT-TIME QUATERNION FOURIER TRANSFORMS OF PARTICLE VELOCITY SIGNALS Hacıhabiboğlu, Hüseyin (2016-03-25) Quaternion Fourier transforms (QFT) provide a powerful tool for the analysis of signals obtained from vector probes. Acoustic particle velocity is one such signal which can be measured with specially designed microphone arrays. This paper presents a time-frequency source separation method based on the short-time quaternion Fourier transform of acoustic particle velocity signals and the k-plane clustering of the vector part of the resulting representation. Two example cases, one with a single and one with tw...
Continuous-time nonlinear estimation filters using UKF-aided gaussian sum representations Gökçe, Murat; Kuzuoğlu, Mustafa; Department of Electrical and Electronics Engineering (2014) A nonlinear filtering method is developed for continuous-time nonlinear systems with observations/measurements carried out in discrete-time by means of UKFaided Gaussian sum representations. The time evolution of the probability density function (pdf) of the state variables (or the a priori pdf) is approximated by solving the Fokker-Planck equation numerically using Euler’s method. At every Euler step, the values of the a priori pdf are evaluated at deterministic sample points. These values are used with Ga...

Citation Formats

A. Aygün and A. Karakuş, “Physics-Informed Neural Networks for Mesh Deformation with Exact Boundary Enforcement,” presented at the SIAM Conference on Computational Science and Engineering 2023, Amsterdam, Hollanda, 2023, Accessed: 00, 2023. [Online]. Available: https://hdl.handle.net/11511/102637.