Nonlinear-manifold reduced order models with domain decomposition

Abstract

A nonlinear-manifold reduced order model (NM-ROM) is a great way of incorporating underlying physics principles into a neural network-based data-driven approach. We combine NM-ROMs with domain decomposition (DD) for efficient computation. NM-ROMs offer benefits over linear-subspace ROMs (LS-ROMs) but can be costly to train due to parameter scaling with the full-order model (FOM) size. To address this, we employ DD on the FOM, compute subdomain NM-ROMs, and then merge them into a global NM-ROM. This approach has multiple advantages: parallel training of subdomain NM-ROMs, fewer parameters than global NM-ROMs, and adaptability to subdomain-specific FOM features. Each subdomain NM-ROM uses a shallow, sparse autoencoder, enabling hyper-reduction (HR) for improved computational speed. In this paper, we detail an algebraic DD formulation for the FOM, train HR-equipped NM-ROMs for subdomains, and numerically compare them to DD LS-ROMs with HR. Results show a significant accuracy boost, on the order of magnitude, for the proposed DD NM-ROMs over DD LS-ROMs in solving the 2D steady-state Burgers’ equation.

1 Introduction

In science and engineering, complex tasks often involve repeatedly simulating a large-scale, parameterized, nonlinear system referred to as the full-order model (FOM). Ensuring high fidelity requires a high-dimensional model, leading to significant computational costs and lengthy simulations. As a result, tasks like design optimization become impractical for large-scale problems. Model reduction offers a solution by replacing the FOM with a computationally efficient, low-dimensional model called a reduced-order model (ROM). This ROM approximates the FOM’s behavior with adjustable accuracy, making it suitable for many-query applications. However, construction of accurate and computationally efficient ROMs poses challenges. To address them, we integrate the nonlinear-manifold ROM (NM-ROM) approach with an algebraic domain-decomposition (DD) framework.

Various model reduction methods have been integrated with DD, like reduced basis elements (RBE) Maday and Rønquist [2002, 2004], Iapichino et al. [2012], Antonietti et al. [2016], Eftang et al. [2012], Huynh et al. [2013], Eftang and Patera [2013], Iapichino et al. [2012], and the alternating Schwarz method Buffoni et al. [2009], Barnett et al. [2022], Smetana and Taddei [2022], Iollo et al. [2023]. However, they are often specialized to specific problems, dealing with the physical domain at the PDE level. In contrast, the authors in Hoang et al. [2021] take an algebraic approach by decomposing the FOM at the discrete level and computing linear-subspace ROMs (LS-ROMs) for each subdomain. While LS-ROMs work well in many cases Haasdonk [2017], Quarteroni et al. [2016], Hinze and Volkwein [2005], Gubisch and Volkwein [2017], Cheung et al. [2023], Copeland et al. [2022], Carlberg et al. [2018], Antoulas [2005], Benner and Breiten [2017], Antoulas et al. [2020], Gu [2011], Benner and Breiten [2015], Mayo and Antoulas [2007], Antoulas et al. [2016], Gosea and Antoulas [2018], Choi et al. [2021], Kim et al. [2021], Choi and Carlberg [2019], it is well known that advection-dominated problems and problems with sharp gradients cannot be well-approximated using low-dimensional linear subspaces. These problems are said to have slowly decaying Kolmogorov $n$-width Ohlberger and Rave [2016]. Recent approaches, such as nonlinear-manifold ROMs (NM-ROMs), address these problems by nonlinearly approximating the FOM in a low-dimensional nonlinear manifold. This is typically achieved through training an autoencoder on FOM snapshot data (e.g., Kashima [2016], Hartman and Mestha [2017], Lee and Carlberg [2020], Kim et al. [2022, 2020]). However, training of NM-ROMs is expensive. Indeed, in the monolithic single-domain case, the high-dimensionality of the FOM training data results in a large number of neural network (NN) parameters requiring training. In Barnett et al. [2023] this cost issue was mitigated by first computing a low dimensional proper orthogonal decomposition (POD) model, and then using a NN to train the coefficients in this POD. Instead, we integrate an autoencoder framework with DD. By coupling NM-ROM with DD, one can compute FOM training data on subdomains, thus reducing the dimensionality of subdomain NM-ROM training data, resulting in fewer parameters that need to be trained per subdomain NM-ROM.

We also note that couplings of NNs and DD for solutions of partial differential equations (PDEs) have been considered in previous work (e.g., Li et al. [2020a, b], Sun et al. [2022], Li et al. [2023]). However, these approaches use deep learning to solve a PDE by representing its solution as a NN and minimizing a corresponding physics-informed loss function. In contrast, our work uses autoencoders to reduce the dimensionality of an existing numerical model. The autoencoders are pretrained in an offline stage to find low-dimensional representations of FOM snapshot data, and used in an online stage to significantly reduce the computational cost and runtime of numerical simulations. Our work is the first to couple autoencoders with DD in the reduced-order modeling context.

Here, we extend the work of Hoang et al. [2021] on DD LS-ROM and integrate NM-ROM with hyper-reduction (HR) using shallow, sparse autoencoders discussed in Kim et al. [2022]. We incorporate the NM-ROM approach into this framework because of its success when applied to problems with slowly decaying Kolmogorov $n$-width. DD allows one to compute FOM training snapshots on subdomains, thus reducing the dimensionality of subdomain NM-ROM training data, resulting in fewer parameters that need to be trained per subdomain NM-ROM. We use wide, shallow, and sparse autoencoder architecture, which allows HR to be efficiently applied, thus reducing the complexity caused by nonlinearity and yielding computational speedup. Additionally, we modify the wide, shallow, and sparse architecture used in Kim et al. [2022] to also include a sparsity mask for the encoder input layer as well as the decoder output layer. The proposed DD NM-ROM approach is compared with DD LS-ROM on the 2D Burgers’ equation.

2 DD full order model

First consider the monolithic, single-domain FOM written as a residual equation

$$\boldsymbol{r}(\boldsymbol{x};{\boldsymbol{\mu}})=\boldsymbol{0},$$ (1)

where $\boldsymbol{x}\in\mathbb{R}^{N_{x}}$ is the state, ${\boldsymbol{\mu}}\in{\cal D}\subset\mathbb{R}^{N_{\mu}}$ is a parameter, and $\boldsymbol{r}:\mathbb{R}^{N_{x}}\times\mathbb{R}^{N_{\mu}}\to\mathbb{R}^{N_{x}}$ is the residual function. FOMs of the form (1) typically arise from discretizations of partial differential equations (PDEs). One can reformulate (1) into a DD formulation by partitioning the residual equation into $n_{\Omega}$ systems of equations (so-called algebraic subdomains), coupling them via compatibility constraints, and converting the systems of equations into a least-squares problem, resulting in

$$\min_{(\boldsymbol{x}_{i}^{\Omega},\boldsymbol{x}_{i}^{\Gamma}),i=1,\dots,n_{\Omega}}\quad\frac{1}{2}\sum_{i=1}^{n_{\Omega}}\left\|\boldsymbol{r}_{i}\left(\boldsymbol{x}_{i}^{\Omega},\boldsymbol{x}_{i}^{\Gamma};{\boldsymbol{\mu}}\right)\right\|_{2}^{2},\quad{\rm s.t.}\quad\sum_{i=1}^{n_{\Omega}}\boldsymbol{A}_{i}\boldsymbol{x}_{i}^{\Gamma}=\boldsymbol{0},$$ (2)

where $\boldsymbol{x}_{i}^{\Omega}\in\mathbb{R}^{N_{i}^{\Omega}}$, $\boldsymbol{x}_{i}^{\Gamma}\in\mathbb{R}^{N_{i}^{\Gamma}}$, $\boldsymbol{r}_{i}:\mathbb{R}^{N_{i}^{\Omega}}\times\mathbb{R}^{N_{i}^{\Gamma}}\times{\cal D}\to\mathbb{R}^{N_{i}^{r}}$, and $\boldsymbol{A}_{i}\in\left\{-1,0,1\right\}^{N_{a}\times N_{i}^{\Gamma}}$ are the $i$-th subdomain interior-state, interface-state, residual function, and compatibility constraint matrix, respectively. The sparsity pattern of the monolithic residual function $\boldsymbol{r}$ determines the structure of the subdomain residual functions $\boldsymbol{r}_{i}$, as well as the decomposition of the state $\boldsymbol{x}$ into subdomain states $(\boldsymbol{x}_{i}^{\Omega},\boldsymbol{x}_{i}^{\Gamma})$. The interior-states $\boldsymbol{x}_{i}^{\Omega}$ are those that are only used to compute the residual $\boldsymbol{r}_{i}$ in the $i$-th subdomain, whereas the interface-states $\boldsymbol{x}_{i}^{\Gamma}$ are also used in the residual computation of neighboring subdomains. The equality constraint determined by $\boldsymbol{A}_{i}$ enforces equality on the overlapping interface states. For further details, see [Diaz et al., 2023, Sec. 2] or [Hoang et al., 2021, Sec. 2].

3 DD nonlinear-manifold reduced order model

For each subdomain $i\in\left\{1,\dots,n_{\Omega}\right\}$, let $\boldsymbol{g}_{i}^{\Omega}:\mathbb{R}^{n_{i}^{\Omega}}\to\mathbb{R}^{N_{i}^{\Omega}}$, $n_{i}^{\Omega}\ll N_{i}^{\Omega}$, and $\boldsymbol{g}_{i}^{\Gamma}:\mathbb{R}^{n_{i}^{\Gamma}}\to\mathbb{R}^{N_{i}^{\Gamma}}$, $n_{i}^{\Gamma}\ll N_{i}^{\Gamma}$, be decoders such that $\boldsymbol{x}_{i}^{\Omega}\approx\boldsymbol{g}_{i}^{\Omega}(\widehat{\boldsymbol{x}}_{i}^{\Omega})$ and $\boldsymbol{x}_{i}^{\Gamma}\approx\boldsymbol{g}_{i}^{\Gamma}(\widehat{\boldsymbol{x}}_{i}^{\Gamma})$. Also let $\boldsymbol{B}_{i}\in\left\{0,1\right\}^{N_{i}^{B}\times N_{i}^{r}}$, $N_{i}^{B}\leq N_{i}^{r}$, denote a row-sampling matrix for collocation HR, and let $\boldsymbol{C}\in\mathbb{R}^{n_{C}\times N_{\overline{A}}}$, $\;n_{C}\ll N_{a}$, be a Gaussian test matrix. The DD NM-ROM is evaluated by solving

$$\min_{(\widehat{\boldsymbol{x}}_{i}^{\Omega},\widehat{\boldsymbol{x}}_{i}^{\Gamma}),i=1,\dots,n_{\Omega}}\quad\frac{1}{2}\sum_{i=1}^{n_{\Omega}}\left\|\boldsymbol{B}_{i}\boldsymbol{r}_{i}\left(\boldsymbol{g}_{i}^{\Omega}\left(\widehat{\boldsymbol{x}}_{i}^{\Omega}\right),\boldsymbol{g}_{i}^{\Gamma}\left(\widehat{\boldsymbol{x}}_{i}^{\Gamma}\right)\right)\right\|_{2}^{2},\quad{\rm s.t.}\quad\sum_{i=1}^{n_{\Omega}}\boldsymbol{C}\boldsymbol{A}_{i}\boldsymbol{g}_{i}^{\Gamma}(\widehat{\boldsymbol{x}}_{i}^{\Gamma})=\boldsymbol{0}.$$ (3)

If HR is not applied (i.e., $\boldsymbol{B}_{i}=\boldsymbol{I}$ in (3)), the ROM’s computational savings are limited because evaluation of residuals $\big{(}\widehat{\boldsymbol{x}}_{i}^{\Omega},\widehat{\boldsymbol{x}}_{i}^{\Gamma}\big{)}\to$ $\big{(}\boldsymbol{g}_{i}^{\Omega}(\widehat{\boldsymbol{x}}_{i}^{\Omega}),\boldsymbol{g}_{i}^{\Gamma}(\widehat{\boldsymbol{x}}_{i}^{\Gamma})\big{)}\to$ $\boldsymbol{r}_{i}\big{(}\boldsymbol{g}_{i}^{\Omega}\big{(}\widehat{\boldsymbol{x}}_{i}^{\Omega}\big{)},\boldsymbol{g}_{i}^{\Gamma}\big{(}\widehat{\boldsymbol{x}}_{i}^{\Gamma}\big{)}\big{)}$ scales with the size $N_{i}^{\Omega}$ and $N_{i}^{\Gamma}$ of the FOM. Thus, HR is applied to decrease the computational complexity caused by the nonlinearity of $\boldsymbol{r}_{i}$, and increase the computational speedup. We use [Carlberg et al., 2013, Algo. 3] to greedily compute a row sampling matrix $\boldsymbol{B}_{i}$ for collocation HR. The application of HR to the decoders $\boldsymbol{g}_{i}^{\Omega}$ and $\boldsymbol{g}_{i}^{\Gamma}$ is discussed further in Sec. 3.1. Following Hoang et al. [2021], we apply a Gaussian test matrix $\boldsymbol{C}\in\mathbb{R}^{n_{C}\times N_{a}}$, $\;n_{C}\ll N_{a}$, to convert the compatibility constraints into a so-called “weak compatibility constraint", which decreases the number of constraints to avoid making the DD ROM over-determined.

The DD FOM (2) and DD NM-ROM (3) are solved using an inexact Lagrange-Newton sequential quadratic programming (SQP) solver, where the Hessian of the Lagrangian is replaced with a Gauss-Newton approximation. This avoids computation of second order derivatives of residuals and constraints in (3), but still achieves good convergence for (2) and (3). For further details, see Diaz et al. [2023].

The DD NM-ROM (3) formulation has several benefits. Training, i.e., computation of the $\boldsymbol{g}_{i}^{\Omega}$ and $\boldsymbol{g}_{i}^{\Gamma}$ is local, involves few parameters, and can be done in parallel. The ROMs can be adjusted to localized features of the problem, which may result in smaller ROMs. Parallelization can be used to speed up ROM computation/training and ROM execution.

3.1 NM-ROM architecture and training

We use single-layer, wide, and sparse decoders with smooth activation functions to represent the maps $\boldsymbol{g}_{i}^{\Omega}$ and $\boldsymbol{g}_{i}^{\Gamma}$. The corresponding encoders, denoted $\boldsymbol{h}_{i}^{\Omega}$ and $\boldsymbol{h}_{i}^{\Gamma}$, are also single-layer, wide, and sparse. Shallow networks are used for computational efficiency; fewer layers correspond to fewer repeated matrix-vector multiplications when evaluating the decoders. The shallow depth necessitates a wide network to maintain enough expressiveness for use in NM-ROM. Smooth activations (i.e., swish) are used to ensure that $\boldsymbol{g}_{i}^{\Omega}$ and $\boldsymbol{g}_{i}^{\Gamma}$ are continuously differentiable. Normalization and de-normalization layers are also applied at the encoder input and decoder output layers, respectively.

Sparsity is applied at the decoder output layer so that HR can be applied. The sparsity allows one to compute a subnet, which only keeps track of the hidden nodes required to compute the output nodes that remain after HR. Further details can be found in [Kim et al., 2022, Sec. 3.2], [Diaz et al., 2023, Sec. 5.3]. We also apply a sparsity mask to the encoder input layer so that the autoencoders are symmetric across the latent layer. The sparsity pattern has a tri-banded structure inspired by 2D finite difference stencils, where the number of nonzeros per band and the separation between bands are hyper-parameters.

To train the autoencoders, we first generate FOM snapshots in an offline stage by solving (2) at parameters $\left\{{\boldsymbol{\mu}}_{\ell}\right\}_{\ell=1}^{M}$, and collect interior- and interface-state snapshot datasets $\boldsymbol{X}_{i}^{\Omega}\in\mathbb{R}^{N_{i}^{\Omega}\times M}$ and $\boldsymbol{X}_{i}^{\Gamma}\in\mathbb{R}^{N_{i}^{\Gamma}\times M}.$ Alternatively, one can solve the monolithic FOM (1) at each ${\boldsymbol{\mu}}_{\ell}$ and restrict the corresponding states $\boldsymbol{x}({\boldsymbol{\mu}}_{\ell})$ to interior-states $\boldsymbol{x}_{i}^{\Omega}({\boldsymbol{\mu}}_{\ell})$ and interface-states $\boldsymbol{x}_{i}^{\Gamma}({\boldsymbol{\mu}}_{\ell})$ for each subdomain. We use the latter approach. The autoencoders $(\boldsymbol{h}_{i}^{\Omega},\boldsymbol{g}_{i}^{\Omega})$ and $(\boldsymbol{h}_{i}^{\Gamma},\boldsymbol{g}_{i}^{\Gamma})$ are then trained in parallel by minimizing the respective MSE losses

$${\cal L}_{i}^{\Omega}=\frac{1}{M}\sum_{\ell=1}^{M}\left\|\boldsymbol{x}_{i}^{\Omega}({\boldsymbol{\mu}}_{\ell})-\boldsymbol{g}_{i}^{\Omega}(\boldsymbol{h}_{i}^{\Omega}(\boldsymbol{x}_{i}^{\Omega}({\boldsymbol{\mu}}_{\ell})))\right\|_{2}^{2},\;{\cal L}_{i}^{\Gamma}=\frac{1}{M}\sum_{\ell=1}^{M}\left\|\boldsymbol{x}_{i}^{\Gamma}({\boldsymbol{\mu}}_{\ell})-\boldsymbol{g}_{i}^{\Gamma}(\boldsymbol{h}_{i}^{\Gamma}(\boldsymbol{x}_{i}^{\Gamma}({\boldsymbol{\mu}}_{\ell})))\right\|_{2}^{2}$$ (4)

for each subdomain $i=1,\ldots,n_{\Omega}$. The snapshots undergo a random 90-10 split for training and validation, and the MSE loss is minimized using the Adam optimizer over $2000$ epochs with a batch size of $32$. We also apply early stopping Prechelt [1998] with a patience of $300$ and reduce the learning rate on plateau with an initial learning rate of $10^{-3}$. The implementation was done in PyTorch and used the PyTorch Sparse and SparseLinear packages.

4 Numerical experiment: 2D Burgers’ equation

We compare the DD LS-ROM of Hoang et al. [2021] and the proposed DD NM-ROM with HR for the 2D steady-state Burgers equation. The DD LS-ROM can be regarded as a specific instance within the realm of DD NM-ROMs, where the encoders and decoders defined in Equation (4) are exchanged for linear operators derived through singular value decomposition. We compute the relative error as

$$e=\left(\frac{1}{n_{\Omega}}\sum_{i=1}^{n_{\Omega}}\Big{(}\left\|\boldsymbol{x}_{i}^{\Omega}-\boldsymbol{g}_{i}^{\Omega}(\widehat{\boldsymbol{x}}_{i}^{\Omega})\right\|_{2}^{2}+\left\|\boldsymbol{x}_{i}^{\Gamma}-\boldsymbol{g}_{i}^{\Gamma}(\widehat{\boldsymbol{x}}_{i}^{\Gamma})\right\|_{2}^{2}\Big{)}/\Big{(}\left\|\boldsymbol{x}_{i}^{\Omega}\right\|_{2}^{2}+\left\|\boldsymbol{x}_{i}^{\Gamma}\right\|_{2}^{2}\Big{)}\right)^{1/2}.$$ (5)

All training and computations were performed on the Lassen machine at Lawrence Livermore National Laboratory, which consists of an IBM Power9 processor with NVIDIA V100 (Volta) GPUs, clock speed between 2.3-3.8 GHz, and 256 GB DDR4 memory. The code can be found at https://anonymous.4open.science/r/DDNMROM_NeurIPS-4160/.

The implementation was done sequentially, but to highlight potential advantages of a parallel implementation, the reported wall clock time for computing subdomain-specific quantities for the SQP solver is taken to be the largest wall clock time incurred among all subdomains. The wall clock time for the remaining steps of the SQP solver is set to the overall wall clock time.

We consider the 2D steady-state Burgers’ equation

$$u\frac{\partial u}{\partial x}+v\frac{\partial u}{\partial y} =\nu\left(\frac{\partial^{2}u}{\partial x^{2}}+\frac{\partial^{2}u}{\partial y^{2}}\right), u\frac{\partial v}{\partial x}+v\frac{\partial v}{\partial y} =\nu\left(\frac{\partial^{2}v}{\partial x^{2}}+\frac{\partial^{2}v}{\partial y^{2}}\right)$$ (6)

for $(x,y)\in[-1,1]\times[0,0.05]$ with viscosity $\nu=0.1$. As in Hoang et al. [2021], we use the exact solution $u_{ex}=-2\nu\frac{\partial}{\partial x}\psi\,/\psi$,  $v_{ex}=-2\nu\frac{\partial}{\partial y}\psi\,/\psi$, where $\psi(x,y;a,\lambda)=a(1+x)+\left(e^{\lambda(x-1)}+e^{-\lambda(x-1)}\right)\cos(\lambda y)$ and $(a,\lambda)$ are parameters, and its restriction to the boundary as Dirichlet boundary conditions. The PDE is discretized using centered finite differences with with $482$ uniformly spaced grid points in the $x$-direction and $26$ uniformly spaced grid points in the $y$-direction. For ROM training, we collected $6400$ FOM snapshots corresponding to varying $(a,\lambda)\in[1,10^{4}]\times[5,25]$ (see Fig. 1) in a uniform $80\times 80$ grid. We use ROMs to predict the out-of-sample case $(a,\lambda)=(7692.5384,21.9230)$.

(a) $(a,\lambda)=(1,25)$.

(b) $(a,\lambda)=(10^{4},5)$.

Figure 1: FOM $u$ and $v$ components for different $(a,\lambda)$. The distance of the shock from the left boundary and its steepness are determined by $a$ and $\lambda$, respectively.

First we use DD problem with $4$ uniformly sized subdomains in a $2\times 2$ configuration and vary the ROM sizes $n_{i}^{\Omega}$ and $n_{i}^{\Gamma}.$ Table 1 shows that NM-ROM has an order of magnitude lower error than LS-ROM with and without HR when comparing ROMs of the same size. In the non-HR case, LS-ROM only achieves order $10^{-3}$ error for a ROM with $96$ total DoF (error = $2.66\times 10^{-3}$), while NM-ROM can achieve a similar error with only $36$ DoF (error = $2.42\times 10^{-3}$) and a higher speedup (speedup = $26.2$) compared to LS-ROM with similar accuracy (speedup = $18.3$). LS-ROM achieves a much higher speedup in the HR cases while retaining similar errors from the non-HR cases. NM-ROM also retains high accuracy after HR, and gains an extra $15$-$20$ times speedup after applying HR.

	$n_{i}^{\Omega}$	$n_{i}^{\Gamma}$	DoF	Error	Speedup	Error (HR)	Speedup (HR)
LS-ROM	$6$	$3$	$36$	$2.06\times 10^{-2}$	$48.7$	$1.78\times 10^{-2}$	$340.0$
	$8$	$4$	$48$	$1.98\times 10^{-2}$	$30.0$	$1.44\times 10^{-2}$	$347.6$
	$10$	$5$	$60$	$1.50\times 10^{-2}$	$16.3$	$1.16\times 10^{-2}$	$329.6$
	$16$	$8$	$96$	$2.66\times 10^{-3}$	$18.3$	$3.23\times 10^{-3}$	$280.4$
NM-ROM	$6$	$3$	$36$	$2.42\times 10^{-3}$	$26.2$	$2.60\times 10^{-3}$	$44.7$
	$8$	$4$	$48$	$1.28\times 10^{-3}$	$21.7$	$1.64\times 10^{-3}$	$43.9$
	$10$	$5$	$60$	$1.09\times 10^{-3}$	$15.0$	$1.19\times 10^{-3}$	$43.6$
	$16$	$8$	$96$	$7.87\times 10^{-4}$	$13.9$	$9.80\times 10^{-4}$	$37.5$

Table 1: Relative error and speedup for LS-ROM and NM-ROM with and without HR for varying ROM size. We use $N_{i}^{B}=100$ HR nodes per subdomain in the HR case.

Next we examine the per-subdomain reduction in the required number of autoencoder parameters for different subdomain configurations compared to the monolothic single-domain NM-ROM. We use the notation $2\times 1$ subdomains to indicate $2$ subdomains in the $x$-direction and $1$ subdomain in the $y$-direction. As expected, from Table 2, we see that the maximum number of NN parameters per subdomain decreases significantly as more subdomains are used. Furthermore, the total number of NN parameters in the DD cases also decreases relative to the single-domain case. We also note that the error increases as more subdomains are used. We kept the ROM size $(n_{i}^{\Omega},n_{i}^{\Gamma})=(6,3)$ constant for each subdomain configuration to isolate the effect of DD on the number of NN parameters, but this may cause overfitting in the $16$ subdomain case. More careful hyper-parameter tuning is necessary to mitigate increases in error as the number of subdomains is increased.

Subdomains	Max # subdomain params.	Reduction	Total # params.	Error
$1\times 1$	$2.995\times 10^{6}$	$0.0$ %	$2.995\times 10^{6}$	$1.08\times 10^{-3}$
$2\times 1$	$1.147\times 10^{6}$	$61.7$ %	$2.307\times 10^{6}$	$1.27\times 10^{-3}$
$2\times 2$	$5.257\times 10^{5}$	$82.4$ %	$2.384\times 10^{6}$	$2.42\times 10^{-3}$
$4\times 2$	$2.617\times 10^{5}$	$91.3$ %	$2.391\times 10^{6}$	$4.26\times 10^{-3}$
$8\times 2$	$1.297\times 10^{5}$	$95.7$ %	$2.406\times 10^{6}$	$4.58\times 10^{-2}$

Table 2: Max number of NN parameters per subdomain, the per-subdomain reduction in number of NN parameters, the total number of parameters, and the corresponding error for different subdomain configurations. For the single-domain case, an NM-ROM of dimension $n=9$ is used. For the DD cases, $(n_{i}^{\Omega},n_{i}^{\Gamma})=(6,3)$, resulting in $9$ DoF per subdomain. HR was not used to evaluate the NM-ROMs in these examples.

5 Conclusion

We extended the DD framework of Hoang et al. [2021] and compute ROMs using NM-ROM with HR as presented in Kim et al. [2022]. Our experiments on the 2D Burgers’ equation show that NM-ROM achieves an order of magnitude lower relative error than LS-ROM in nearly all cases tested. While LS-ROM with HR achieves much higher speedup than NM-ROM with HR, NM-ROM is still the clear winner in terms of ROM accuracy for a given ROM size. Moreover, HR allows NM-ROM to gain an extra $15$-$20$ time speedup compared to the non-HR cases. While the speedup is not as drastic as for LS-ROM, these speedup gains for NM-ROM are the highest that have been achieved for NM-ROM to our knowledge. We also showed that using the DD approach significantly decreases the number of required NN parameters per subdomain compared to the monolithic single-domain NM-ROM. In future work, we plan to apply DD NM-ROM to more challenging problems, including those with slowly decaying Kolmogorov $n$-width and to time-dependent problems. Other directions for future research include a greedy sampling strategy when choosing which FOM snapshots to compute for NM-ROM training and applying the DD NM-ROM framework to decomposable or component-based systems.