Abstract:Despite great improvements in computing hardware and developments of new methodologies for solving partial differential equations (PDEs), solving PDEs numerically can still be computationally prohibitive for certain applications. This computational difficulty is especially true when the solution of PDEs is tied to characterization, control, design, or other inverse problems in general. Most of the traditional PDE solution strategies, such as the Finite Element Method or Finite Volume Method can require hundreds of thousands of degrees of freedom to accurately capture the behavior of even relatively simple physical system. The computational cost is several orders of magnitude higher for solving optimization problems (a common approach to solve inverse problems), which require obtaining several solution fields. Hence, model order reduction is necessary to enable the solution of such optimization problems with sufficient efficiency to allow practical applicability. Several approaches have been developed to create accurate reduced-order models (ROMs) of physical systems with dramatically reduced computational expense. Yet, several questions remain as to the optimal approach to create ROMs for a given physical system to ensure suitable accuracy, and even more so in relation to inverse problem applications. The objective of the present work is to address issues relating to the creation of suitably accurate ROMs that can be utilized for the solution of a variety of computational inverse mechanics problems. This work focuses on ROMs that utilize proper orthogonal decomposition (POD) to create a reduced-order basis for the PDE solution from a set of previously obtained potential solution fields (i.e., snapshots) of the system of interest. First, a generally applicable algorithm is presented to efficiently create accurate ROMs for use in solving inverse problems in material characterization (i.e., nondestructive evaluation). This algorithm is based upon a novel concept for maximizing the diversity of the system snapshots used to create the ROM. Results show that by maximizing the snapshot diversity, the accurate generalization of the resulting ROM is substantially improved, which then improves inverse problem solution capabilities. Then, a comprehensive study is presented of the capability of a set of different approaches for reduced-order modeling (all still using POD) to represent systems involving flow past bluff bodies. One of the relatively unexplored issues when creating ROMs for fluid flows is the accuracy with respect to changes in Reynolds (Re) number. The present work uses the generalized POD technique to create ROMs that are capable of predicting flow field not only at different time levels, but also at different Re numbers. Finally, the ROM strategies explored for fluid flow were extended to investigate the applicability to optimal flow control problems. The ROM of flow was used along with an optimization technique and adjoint method for gradient calculation to solve a control problem to reduce the drag force in flow past one or more cylinders. Results of the solution of this optimal control problem show that the developed ROMs are capable of solving complex optimization problems with significantly reduced computational expense.

Compressed Neural Networks for Reduced Order Modeling

Deep neural networks for nonlinear model order reduction of unsteady flows

A Non-Intrusive Reduced Order Model with Transformer Neural Network and Its Application

Data-driven reduced-order modeling for nonlinear aerodynamics using an autoencoder neural network

Data-driven reduced order modeling of environmental hydrodynamics using deep autoencoders and neural ODEs

Reduced-Order Modeling of Steady and Unsteady Flows with Deep Neural Networks

An artificial neural network framework for reduced order modeling of transient flows

A fast and accurate physics-informed neural network reduced order model with shallow masked autoencoder

Reduced order modelling of intracranial aneurysm flow using proper orthogonal decomposition and neural networks

Explicable hyper-reduced order models on nonlinearly approximated solution manifolds of compressible and incompressible Navier-Stokes equations

On Reduced-Order Modeling of Gas–solid Flows Using Deep Learning

Data-Driven Nonintrusive Model-Order Reduction for Aerodynamic Design Optimization

Polytopic autoencoders with smooth clustering for reduced-order modeling of flows

Polytopic Autoencoders with Smooth Clustering for Reduced-order Modelling of Flows

Artificial neural network-based reduced-order modeling for turbulent wake of a finite wall-mounted square cylinder

Non-intrusive reduced order modeling of natural convection in porous media using convolutional autoencoders: comparison with linear subspace techniques

Reduced-order modeling of advection-dominated systems with recurrent neural networks and convolutional autoencoders

Standardized Non-Intrusive Reduced Order Modeling Using Different Regression Models With Application to Complex Flow Problems

Reduced order modeling for flow and transport problems with Barlow Twins self-supervised learning

Reduced-Order Modeling Toward Solving Inverse Problems in Solid Mechanics and Fluid Dynamics

Reduced-order modeling for stochastic large-scale and time-dependent flow problems using deep spatial and temporal convolutional autoencoders