betaFeedback

WorksPhysics and Astronomy

Quantum-Inspired Tensor-Network Fractional-Step Method for Incompressible Flow in Curvilinear Coordinates

Abstract

We introduce an algorithmic framework based on tensor networks for computing fluid flows around immersed objects in curvilinear coordinates. We show that the tensor network simulations can be carried out solely using highly compressed tensor representations of the flow fields and the differential operators and discuss the numerical implementation of the tensor operations required for computing fluid flows in detail. The applicability of our method is demonstrated by applying it to the paradigm example of steady and transient flows around stationary and rotating cylinders. We find excellent quantitative agreement in comparison to finite difference simulations for Strouhal numbers, forces and velocity fields. The properties of our approach are discussed in terms of reduced order models. We estimate the memory saving and potential runtime advantages in comparison to standard finite difference simulations. We find accurate results with errors of less than 0.3% for flow-field compressions by a factor of up to 20 and differential operators compressed by factors of up to 1000 compared to sparse matrix representations. We provide strong numerical evidence that the runtime scaling advantages of the tensor network approach with system size will provide substantial resource savings when simulating larger systems. Finally, we note that, like other tensor network-based fluid flow simulations, our algorithmic framework is directly portable to a quantum computer leading to further scaling advantages.

Introduction

Engineering computational fluid dynamics (CFD) is subjected to an ever-increasing demand for higher spatial and temporal resolution. An improved dynamic resolution is crucial to understand complex multiphysics and multiscale dynamics whose analysis has so far been impossible due to a lack of computing power. The most prominent example probably relates to the simulation of turbulent flows with the full resolution of all spatiotemporal scales in a direct numerical simulation (DNS). This requirement can pose enormous computational challenges because the range of relevant scales increases nonlinearly with the Reynolds number, which renders DNS infeasible for engineering CFD [1, 2, 3].

Since the ultimate goal of resolving turbulence by DNS remains elusive in CFD, approximate methods based on modeling turbulence physics, coarse-grained or reduced order models (ROMs) have been developed. Turbulence closure models [4, 5, 6] based on empirical parameters substantially reduce the resolution requirements, but at the expense of a lower generality and less detailed results. Alternatively, ROMs [7, 8, 9, 10, 11, 12, 13, 14, 15], utilize scale-resolving strategies that mitigate the curse of dimensionality problem [16] by reducing state space complexities [17, 18, 19, 20, 21, 22], for example, through modal decomposition. However, CPU-efficient ROMs do not provide reliable (long-term) predictions due to the simplification of the dynamics.

Advances in CFD may also arise from the development of novel computing hardware [2]. In this regard, quantum computers (QCs) have the potential to exponentially reduce the amount of memory required for storing highly resolved discretized flow fields [23, 24]. Recent overviews of the prospects of different quantum-based strategies in engineering applications, especially CFD, are provided in [24, 25, 26]. Aiming to apply quantum mechanical concepts, it is important to understand its major bottleneck. This arises from the linear nature of quantum mechanics, and poses a challenge for simulating nonlinear differential equation problems on a QC. Other challenges relate to the measurement of quantum (flow) states as well as hardware-related influences of noise and decoherence.

Proposals for combining the memory advantage with quantum algorithms to solve nonlinear CFD problems are, e.g., based on Carleman linearization [27, 28], efficient quantum algorithms for solving linear(ised) equation systems [29, 30] or quantum realizations of mesoscopic lattice Boltzmann frameworks [31, 32, 33, 34]. Other options refer to transforming the nonlinear differential flow equations into equivalent Schrödinger equations, cf. [35], whose dynamics can be conveniently simulated on a QC using a Hamiltonian simulation algorithm [36, 37, 38, 39, 40].

To balance unfavorable properties of current quantum hardware, the use of hybrid quantum-classical computers can be advantageous. Related approaches often transform the governing transport equations into an optimization problem and parameterize the state. The parameters encode the nonlinearities and are optimized by a corresponding optimization procedure. Such hybrid quantum-classical variational quantum methods [41, 42, 43] benefit from the exponential memory saving and might also achieve a runtime advantage, see, e.g., [24, 44, 45, 46] for CFD applications.

Another avenue is to reformulate classical algorithms using quantum computing strategies to develop quantum-inspired methods that potentially offer higher computational power. Such quantum-inspired methods can be implemented on classical computers but may also use quantum computers as accelerators. Their algorithms often utilize tensor networks [47] as a programming paradigm for the quantum part of the computation [48, 49]. Since sufficiently advanced quantum hardware is not yet available, their performance and accuracy is often analyzed and benchmarked on classical hardware. Fully scale-resolving quantum-inspired tensor network simulations are computationally attractive only if the entanglement – i.e., the internal correlations between different parts of the system – is limited, effectively turning them into ROMs [24, 50]. As conventional intrusive ROMs [51], these methods allow complete CFD simulation within a compressed subspace, akin to studies in quantum many-body systems [52].

Quantics tensor trains [53], also known as a special type of matrix product states (MPS) in quantum physics, and hereafter referred to simply as tensor trains (TTs), have emerged as strong candidates for quantum-inspired tensor network CFD methods. Building on the pioneering work of Dolgov et al. [54], which utilized the TT format to solve semi-discrete parabolic problems, recent studies have successfully extended quantum-inspired ROMs to a broad variety of increasingly advanced CFD problems [54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65]. They mainly utilize conventional finite difference (FD) frameworks and address various challenges faced by conventional CFD, such as incorporating higher-order spatial and temporal approximation schemes. Through this approach, significant speedups have been reported: a 12-fold reduction in computational time for simulating a 2D decaying-turbulence flow on graphic-cards [57], one order of magnitude runtime improvement [58] compared to classical FD for the paradigmatic 2D lid-driven cavity problem and Gourianov et al. [66] reported a reduction of memory and computational cost up to six and three orders of magnitude, respectively, for the (5+1)D joint probability density-function simulation of a chemically reactive turbulent flow.

Applications of TT methods to complex geometries are still in their infancy and simulations around non-rectangular bodies represent a challenge. Nowadays, mesh-based classical CFD methods predominantly refer to body-fitted, unstructured polyhedral meshes [67, 68]. Other approaches employ the immersed boundary method (IBM) to capture complex geometries usually embedded in structured Cartesian grids [69, 70, 71, 72]. In addition, particularly in the area of complex multi-body flow simulations, there are approaches based on overlapping domain-specific grids, which are structured and body-fitted. A key challenge of overset grids is certainly the organization of the dynamic connectivity between grids and the related conservative inter-grid interpolation [73, 74]. To date, only two studies have investigated TT-based CFD methods for complex domains. The first one was published by Peddinti et al. [60]. It employs a direct-forcing IBM approach on a Cartesian mesh to simulate flow around circular and square cylinders, as well as a NACA0042 airfoil. However, the added geometry-representing penalty terms inherent to this method generally introduce extra algorithmic complexity and computational overhead. The second one, reported by Kornev et al. [75], utilizes block-structured Cartesian meshes to simulate flow in a T-shaped channel, yet can not resolve non-rectangular objects.

The present paper reports the implementation of a TT-based CFD procedure for body-fitted structured-grid discretizations using curvilinear coordinates. The transformation of Cartesian differential operators into curvilinear operators within the TT format will be outlined in detail. Analogous to classical approaches, the procedure is based on a fractional-step algorithm [76, 77, 78]. This enables an accurate and efficient simulation of incompressible flows around submerged objects, extending the applicability of TT methods beyond Cartesian grids. An investigation of overset grids is, however, beyond the scope of this article but it paves the way for future TT applications on these type of grids.

This publication is structured as follows: Section II introduces the TT format and discusses the realization and scaling of basic algebraic operations such as addition and multiplication of fields and operators, together with a short introduction into the density matrix renormalization group algorithm used to solve linear systems of equations. The subsequent Section III sketches a generic physical model of an incompressible flow around an obstacle and introduces a discrete curvilinear coordinate transformation. Furthermore, it covers the representation of the flow fields, grid generation, and the construction of curvilinear differential operators, all within the TT format, as well as the implicit-explicit scheme for time integration.

In Sections IV A 1–IV B the proposed framework is applied to two-dimensional, laminar flows around both nonrotating and rotating circular cylinders in steady and transient regimes (20≤Re≤20020 \le Re \le200). We observe excellent prediction and rapid convergence of the TT algorithm toward the reference values characterizing the system dynamics, such as mean lift coefficients and Strouhal numbers. Remarkably, for the steady (non-rotating) cylinder case the proposed TT solver achieves a relative error below 0.3% in the velocity magnitude utilizing only 5.8% degrees of freedom of classical simulations, i.e., a 20-fold compression. The flow dynamics in the transient case are also accurately predicted with an absolute error below 1×10−31 \times10^{-3} in the Strouhal number at compression ratios as low as 18%. Section IV C discusses the computational performance of the TT solver on the transient flow case. We demonstrate that the computational effort required to refine the discretization is significantly reduced when using the TT method compared to classical FD simulations, particularly in cases where the bond dimension does not grow substantially with the system size. Furthermore, we show that the required curvilinear differential operators admit a compression of up to 1000-fold in the TT format, efficiently encoding the orthogonal cylindrical grid. The last section, Section V, is devoted to conclusions and outlook, where we discuss possible implications for future developments of tensor-network-based fluid dynamics.

To accommodate a broad audience, this publication makes extensive use of Appendices to provide additional technical details and background. Appendix A provides additional information regarding the discretization and the curvilinear coordinate transformation. Appendix B introduces the vectorization of a matrix and Appendix C specifies how rows and columns are extracted from the TT representation of a vectorized matrix. The decomposition of a vector into a TT is explained in detail in Appendix D. The grid points used in the simulations and the details on the force computation are specified in Appendix E and F, respectively. Appendix G introduces the algorithmic modifications necessary to compute the transient flow and Appendix H discusses the verification of the in-house FD solver using an OpenFOAM simulation for comparison.

The publication employs lower-case Latin letters to describe field properties and uppercase Latin letters indicate reference quantities used to non-dimensionalize the field quantities. Dimensional field quantities are denoted by a tilde. In a symbolic representation, the number of underlines indicates the tensorial order, e.g., x~\tilde{\boldsymbol{x}} for the position vector and ∇\nabla for the non-dimensional spatial derivative vector. Spatial vectors and tensors are usually defined with reference to Cartesian coordinates. Furthermore, the notation for discretized quantities follows

Vectors (order-1 tensor): v\boldsymbol{v},

Matrices (order-2 tensor): A\boldsymbol{A},

Order-3 or higher tensors: v\boldsymbol{v}, A\boldsymbol{A},

Tensor Trains: V\mathcal{V}, A\mathcal{A}.

Tensor Trains: Principles and Fundamental Operations

This section presents a reduced order modeling framework based on TTs in their binary (dyadic) form – usually called quantics tensor trains or understood as a special case of matrix product states – which we adopt for its compatibility with quantum hardware. Throughout the remainder of the paper, the term TT will always refer to this binary specialization.

We start by defining the notation and describe the encoding of high-dimensional tensors, representing state vectors and operators, into the TT format. Subsequently, the algebraic operations required in this framework are detailed in Section II A. Moreover, the complexity of the operations is discussed in Section II B. Throughout this section, we practice the Einstein summation convention, where repeated indices imply summation. For reordering vectors to matrices and vice versa, the index fusion/extension always assumes a column ordering. In general, it must be chosen consistently, and it mostly depends on the programming language, which is in our case JULIA [79]. For the sake of understandability, the sorting approach is detailed in the Appendix B. For any contraction, we rely on a toolbox determining the most efficient order.

For an introduction to the fundamentals of TTs, a 2n2^n-component vector is reshaped into an order-nn tensor ar1,r2,…,rn\boldsymbol{a}_{r_1,r_2,\ldots,r_n} with binary rk∈{0,1}r_k \in\{0,1\} for k=1,…,nk=1,\ldots,n and subsequently decomposed into a TT representation of nn order-3 tensors A[k]\boldsymbol{A}^{[k]} (TT-cores) [75, 80, 81, 82], as it is schematically depicted in Figure 1. For the representation of tensors in this work, we follow the (graphical) Penrose notation [83], where a vector is an object with one leg (cf. Figure 1), a line sticking out of the blue box/circle, a matrix has two legs, and an order-nn tensor shows nn legs. The number of outgoing legs thereby determines the order of the tensor. A link between any two tensors, indicated with a lowercase Greek letter, is referred to as a bond (or artificial) index implying summation. The outer indices instead are called physical (or real) indices with dimension 2 and indicated by a lowercase Latin letter.

Figure 1. Example for the decomposition of an arbitrary vector a\boldsymbol{a}, reshaped as a 2n2^n-component tensor ar1,r2,…,rn\boldsymbol{a}_{r_1,r_2,\ldots,r_n} into the TT format by performing a series of SVDs, where A[k]\boldsymbol{A}^{[k]} are order-3 tensors with mode indices rk∈{0,1}r_k \in\{0,1\}, and bond indices αk\alpha_k. The corresponding mathematical interpretation is given in Eq. (1).

The decomposition illustrated in Figure 1 is performed by the successive application of singular value decompositions (SVDs). As an alternative, the TT-cross algorithm [84, 85] can be employed, or for certain cases, the TT-vectors can be derived analytically, as is the case of polynomial and trigonometric functions [86]. For further insights, the interested reader is referred to the Refs. [47, 87]. Additionally, an exemplary analytical tensor decomposition is given in Appendix D. In any case, this decomposition is not unique, and a coherent order within a so-called canonical form [47, 87] is recommended. Within the toolbox we maintain a right-canonical form.

The dimension of the bond indices αk\alpha_k connecting individual order-3 tensors can be set by either neglecting singular values below a certain threshold or by manually selecting the first χ\chi diagonal entries in the singular value matrix [47]. In this regard, the complete TT-vector reads as

a:=A[1]α0,α1r1A[2]α1,α2r2⋯A[n]αn−1,αnrn≃ar1,r2,…,rn(1)\mathbf{a} := \mathbf{A}[1]_{\alpha_0,\alpha_1}^{r_1}\mathbf{A}[2]_{\alpha_1,\alpha_2}^{r_2}\cdots\mathbf{A}[n]_{\alpha_{n-1},\alpha_n}^{r_n}\simeq a_{r_1,r_2,\ldots,r_n} \tag*{(1)}

where the dimensions of bond indices αk\alpha_k follow dim⁡(αk)=min⁡(2k,2n−k,χ)\dim(\alpha_k)=\min(2^k,2^{n-k},\chi) for k=0,…,nk=0,\ldots,n. Accordingly, χ\chi is always referred to as the maximal value in the set {dim⁡(α0),dim⁡(α1),…,dim⁡(αn)}\{\dim(\alpha_0),\dim(\alpha_1),\ldots,\dim(\alpha_n)\}, and controls the expressiveness of the representation, possibly affecting its accuracy. In the case where the bond dimension is not under truncation, the TT representation a=ar1,r2,…,rn\mathbf{a}=a_{r_1,r_2,\ldots,r_n} is exact.

Similarly, a matrix MM can be represented as a TT-matrix M\mathcal{M} (TT-operator) by local order-4 tensors M[k]\mathbf{M}[k] with incoming (not primed) and outgoing (primed) modes of the same mode size (square matrices), 2, such that the contraction of incoming mode indices with the mode indices of a TT-vector naturally gives a new TT-vector. The TT-matrix is illustrated in Figure 2.

Figure 2. Graphical depiction of the TT representation M\mathcal{M} of a matrix MM. The red circles with outgoing legs denote TT-cores, where (r1′,…,rn′)(r'_1,\ldots,r'_n) represents the rows of MM and (r1,…,rn)(r_1,\ldots,r_n) enumerates its columns. The corresponding mathematical description is given in Eq. (2).

The matrix MM can be encoded through the successive application of SVDs, defining the maximal bond dimension χ(M)\chi(\mathcal{M}) as the largest value in the set {dim⁡(β0),dim⁡(β1),…,dim⁡(βn)}\{\dim(\beta_0),\dim(\beta_1),\ldots,\dim(\beta_n)\}. The TT cross algorithm [84] may also be employed to further reduce computational effort when obtaining the TT-matrix. However, in most cases, numerical decompositions are not required, as the TTs in these instances can be derived analytically [88].

Each Latin index again holds a mode size of 2. The resulting TT-matrix [75], consisting of order-4 tensors, is given by

M:=M[1]β0,β1r1′,r1M[2]β1,β2r2′,r2⋯M[n]βn−1,βnrn′,rn≃Mr1′,r2′,…,rn′r1,r2,…,rn.(2)\mathcal{M}:=\mathbf{M}[1]_{\beta_0,\beta_1}^{r'_1,r_1}\mathbf{M}[2]_{\beta_1,\beta_2}^{r'_2,r_2}\cdots\mathbf{M}[n]_{\beta_{n-1},\beta_n}^{r'_n,r_n}\simeq M_{r'_1,r'_2,\ldots,r'_n}^{r_1,r_2,\ldots,r_n}. \tag*{(2)}

Again, without truncation, the above representation is exact.

Algebraic Operations

In this TT-representation of matrices and vectors, the common algebraic operations needed for solving PDEs can be performed efficiently. These include vector-vector addition (e.g., a+b\mathbf{a}+\mathbf{b}), matrix-vector product (e.g., Ma\mathcal{M}\mathbf{a}), and elementwise multiplication of two vectors (e.g., a⊙b\mathbf{a}\odot\mathbf{b}), and are introduced in the following Sections II A 1-II A 3. Solving linear systems of equations is done variationally, as explained in Section II A 4. Application of TT operations may result in an increased bond dimension, which is typically controlled through regular truncation of the TT-vector (TT-rounding) [47, 80] to maintain computational efficiency.

Vector-Vector Addition

Given two TT-vectors a\mathbf{a} and b\mathbf{b}, with cores A[k]rk\mathbf{A}[k]^{r_k} and B[k]rk\mathbf{B}[k]^{r_k} and bond dimension χ(a)\chi(\mathbf{a}) and χ(b)\chi(\mathbf{b}), a summation of a\mathbf{a} and b\mathbf{b} results in the new TT-vector c\mathbf{c}. For the summation, each core C[k]\mathbf{C}[k] of c\mathbf{c} is built as

C[k]rk=[A[k]rk00B[k]rk]=A[k]rk⊕B[k]rk,(3)\mathbf{C}[k]^{r_k}= \begin{bmatrix} \mathbf{A}[k]^{r_k} & 0\\ 0 & \mathbf{B}[k]^{r_k} \end{bmatrix} =\mathbf{A}[k]^{r_k}\oplus\mathbf{B}[k]^{r_k}, \tag*{(3)}

where the direct sum ⊕\oplus is the shorthand notation for this block encoding. For k=1k=1 (k=nk=n), the TT-cores of a\mathbf{a} and b\mathbf{b} are contracted in a row (column) vector, respectively. The resulting TT-vector c\mathbf{c} will have a bond dimension χ(c)=χ(a)+χ(b)\chi(\mathbf{c})=\chi(\mathbf{a})+\chi(\mathbf{b}), which may be truncated.

Matrix-Vector Product

The matrix-vector product is realized as a combination of TTs [88]. For the application of a TT-matrix M\mathcal{M} to a TT-vector a\mathbf{a}, multiple contractions along the physical indices are required, as it is depicted in Figure 3. In mathematical terms, the TT-vector b=Ma\mathbf{b}=\mathcal{M}\mathbf{a} has the cores

Figure 3. Graphical representation of TT-matrix acting on a TT-vector, which is achieved by contraction over common physical indices.

B[k]1,(αkβk)rk′=∑rkM[k]1,βkrk′,rkA[1]1,αkrk(4)\mathbf{B}[k]_{1,(\alpha_k\beta_k)}^{r'_k}=\sum_{r_k}\mathbf{M}[k]_{1,\beta_k}^{r'_k,r_k}\mathbf{A}[1]_{1,\alpha_k}^{r_k} \tag*{(4)}

such that the resulting tensor train is

b=B[1]1,(α1β1)r1′B[2](α1β1),(α2β2)r2′⋯B[n](αn−1βn−1),1rn′.(5)\mathbf{b}=\mathbf{B}[1]_{1,(\alpha_1\beta_1)}^{r'_1}\mathbf{B}[2]_{(\alpha_1\beta_1),(\alpha_2\beta_2)}^{r'_2}\cdots\mathbf{B}[n]_{(\alpha_{n-1}\beta_{n-1}),1}^{r'_n}. \tag*{(5)}

Here, (αk βk)(\alpha_k\ \beta_k) denotes the fused bond index formed by combining αk\alpha_k and βk\beta_k, with total dimension $\dim(\alpha_k)\times dim⁡(βk)\dim(\beta_k) (Appendix B). As a result, this increases the bond dimension of the resulting TT-vector, which requires regular truncation (e.g., using SVD) for operations to remain efficient.

Vector-Vector Elementwise Multiplication

As the last fundamental operation, the vector-vector elementwise multiplication, indicated by ⊙\odot, is introduced. For three TT-vectors aa, bb, and cc, the elementwise product c=a⊙bc = a \odot b is computed in two steps. The first step involves contracting the copy tensor δrk,rk′,rk′′rk,rk′,rk′′\delta^{r_k,r_k',r_k''}_{r_k,r_k',r_k''} tensor, defined as [89]

δrk,rk′′rk′={1if rk=rk′=rk′′,0otherwise(6)\delta^{r_k'}_{r_k,r_k''} = \begin{cases} 1 & \text{if } r_k = r_k' = r_k'',\\ 0 & \text{otherwise} \end{cases} \tag*{(6)}

into each TT-core of aa, which effectively doubles its physical dimension (from 2 to 4). This turns the TT-vector aa into a diagonal TT-matrix A\mathcal{A} with the vector entries along its diagonal (cf. Appendix B). Secondly, the resulting TT c=Abc = \mathcal{A}b is computed by the matrix-vector product, as it is described in Section II A 2. Both steps of the vector-vector elementwise operation are illustrated in Figure 4. This increases the bond dimension of the resulting TT-vector according to the matrix–vector product detailed in the previous subsection.

Figure 4. Graphical representation of vector-vector elementwise multiplication between two TT-vectors aa and bb.

Solving Linear Systems of Equations

The discretization of partial differential equations typically leads to linear systems of equations (LSEs). In this section, we introduce a TT-based approach for solving such systems. In classical computing, the unknowns of an LSE can be determined in many ways (e.g., Gauss-elimination, Cholesky decomposition, . . .). Within the TT format, however, one typically employs dedicated solvers; examples are described in Refs. [90, 91, 92, 93]. In the present work we adopt a single-site version of the variational density matrix renormalization group (DMRG) algorithm, also known as the alternating least-squares method, as detailed, for example, in Refs. [91, 92].

Generally, given TT representations of a matrix M\mathcal{M}, a vector aa and a right-hand side (RHS) bb the linear problem reads as

Ma=b.(7)\mathcal{M}a=b . \tag*{(7)}

For application of the DMRG, we define a cost function C(a)C(a), whose global minimum corresponds to the exact solution of Eq. (7) if the matrix M\mathcal{M} is symmetric and positive-semidefinite [54, 94]. If this condition is not satisfied, this cost function has still been shown to perform well [54]; alternatively, a least-squares formulation may be employed [94]. To approximate the solution of the given problem, we solve the optimization problem

min⁡aC(a):=12(aMa)−ab,(8)\min_a C(a) := \frac{1}{2}(a\mathcal{M}a)-ab , \tag*{(8)}

where the unknown TT-vector aa is fully parameterized. Taking the derivative of the cost functional w.r.t. the individual TT-cores of aa reads

∂C∂A[k]=12∂(aMa)∂A[k]−∂(ab)∂A[k]=M(αk−1′rk′αk′),(αk−1rkαk)effA(αk−1rkαk)[k]−b(αk−1′rk′αk′)eff,(9)\frac{\partial C}{\partial A^{[k]}} = \frac{1}{2}\frac{\partial(a\mathcal{M}a)}{\partial A^{[k]}}-\frac{\partial(ab)}{\partial A^{[k]}} = M^{\mathrm{eff}}_{(\alpha_{k-1}'r_k'\alpha_k'),(\alpha_{k-1}r_k\alpha_k)}A^{[k]}_{(\alpha_{k-1}r_k\alpha_k)} -b^{\mathrm{eff}}_{(\alpha_{k-1}'r_k'\alpha_k')} , \tag*{(9)}

where (αk−1rkαk)(\alpha_{k-1}r_k\alpha_k) denotes the fused bond index formed by combining indices αk−1\alpha_{k-1}, rkr_k, and αk\alpha_k (cf. Appendix B). The algorithm is then executed via successive DMRG-style sweeps over all TT-cores [47, 91]. During the update of core A[k]\mathbf{A}^{[k]}, all other cores are held fixed and a small linear system for the optimal A[k]\mathbf{A}^{[k]} is obtained by enforcing the stationarity condition ∂C/∂A[k]=0\partial C/\partial A^{[k]}=0. This step is graphically depicted in Figure 5. The local effective matrix MeffM^{\mathrm{eff}} is of size 2χ2×2χ22\chi^2 \times2\chi^2, so that the matrix-vector product MeffA[k]M^{\mathrm{eff}}A^{[k]} scales as O(χ4)\mathcal{O}(\chi^4).

Figure 5. Graphical representation of the local linear system for the kk-th TT-core A[k]\mathbf{A}[k], obtained by enforcing the stationarity condition ∂C/∂A[k]=0\partial C/\partial\mathbf{A}[k] = 0. Solving this system yields the updated core A[k]\mathbf{A}[k] (orange node).

An alternative approach is to emulate the local matrix-vector product by keeping the constituent tensors of MeffM^{\mathrm{eff}} and applying successive tensor contractions [54]. This is reported to scale as O(χ(M)χ3+χ(M)2χ2)\mathcal{O}(\chi(\mathcal{M})\chi^3+\chi(\mathcal{M})^2\chi^2). However, numerical studies have shown poor convergence in the latter scheme in the case of ill-conditioned systems, rendering preconditioners mandatory. A strategy to find preconditioners directly in the TT format is outlined in [54].

TT Complexity

The operations introduced in the former Sections II A 1-II A 4, come at certain computational expenses. The alternative implementation to the TT operations discussed above is the use of variational methods [95], which avoid the need for intermediate truncation by directly optimizing the solution within the desired bond dimension. A quantitative comparison between both is given in Table I [58, 75], where it is found that variational approaches provide better scaling with increasing bond dimension χ\chi. Throughout our simulations, we employ the variational technique for the most computationally demanding operations, such as elementwise multiplications and the solution of LSEs. For operations with comparatively negligible computational cost, direct methods are preferred, as they avoid the overhead associated with variational approaches and their potential convergence issues in the presence of strong numerical fluctuations. Within the literature, one finds the TT multiplication algorithm [61], or the zip-up [96], which can improve the initially mentioned scalings.

OperationDirect approachVariational
c=a+bc = a + bO((χ(a)+χ(b))3)\mathcal{O}\left((\chi(a)+\chi(b))^3\right)O(χ(c)(χ(a)+χ(b))2)\mathcal{O}\left(\chi(c)(\chi(a)+\chi(b))^2\right)
c=a⊙bc = a \odot bO(χ(a)3χ(b)3)\mathcal{O}\left(\chi(a)^3\chi(b)^3\right)O(χ(a)χ(b)2χ(c)+χ(a)2χ(b)χ(c))\mathcal{O}\left(\chi(a)\chi(b)^2\chi(c)+\chi(a)^2\chi(b)\chi(c)\right)
mat-vec⁡ b=Ma\operatorname{mat\text{-}vec}\ b = \mathcal{M}aO(χ(M)3χ(a)3)\mathcal{O}\left(\chi(\mathcal{M})^3\chi(a)^3\right)O(χ(M)χ(a)2χ(b))\mathcal{O}\left(\chi(\mathcal{M})\chi(a)^2\chi(b)\right)
LSE⁡ Ma=b\operatorname{LSE}\ \mathcal{M}a=bN/AO(χ(M)χ(a)+χ(M)2χ(a)2)\mathcal{O}\left(\chi(\mathcal{M})\chi(a)+\chi(\mathcal{M})^2\chi(a)^2\right)

Table 1. Computational complexity of the TT operations introduced in Section II, including subsequent truncation to a target bond dimension (TT-rounding), and variational complexities as discussed in Ref. [95].

Computational Model

This section briefly outlines the governing equations (Section III A) and the classical FD-based discretization strategy based on curvilinear coordinates (Section III B). In the latter, we describe how this discretization is implemented within the TT format, i.e., encoding two-dimensional scalar fields (III B 1), generating a body-fitted grid (III B 2), and transforming the corresponding differential operators (III B 3). The section concludes with detailing the fractional step method employed to solve the incompressible Navier–Stokes equations (INSE) in Section III C.

Governing Equations

The two-dimensional, laminar fluid-dynamics problem is governed by the INSE for a Newtonian fluid with constant density ρ~\tilde{\rho}. Using non-dimensional spatial coordinates x‾=x~‾/L\underline{x}=\underline{\tilde{x}}/L, velocities v‾=v~‾/U∞\underline{v}=\underline{\tilde{v}}/U_{\infty} in addition to a non-dimensional time t=t~/Treft=\tilde{t}/T_{\mathrm{ref}} and pressure p=p~/Prefp=\tilde{p}/P_{\mathrm{ref}}, the system of momentum (10) and continuity (11) equations inside a physical domain Ω\Omega read

∂v‾∂t+(v‾⋅∇)v‾=−∇p+1ReΔv‾+f‾in Ω,(10)\frac{\partial\underline{v}}{\partial t}+(\underline{v}\cdot\nabla)\underline{v}=-\nabla p+\frac{1}{Re}\Delta\underline{v}+\underline{f}\quad\text{in }\Omega, \tag*{(10)}
∇⋅v‾=0in Ω,(11)\nabla\cdot\underline{v}=0\quad\text{in }\Omega, \tag*{(11)}

where ∇\nabla denotes the vector of the non-dimensional spatial derivatives, Re=U∞L/ν~Re=U_{\infty}L/\tilde{\nu} is the Reynolds number, Tref=L/U∞T_{\mathrm{ref}}=L/U_{\infty} is a convective reference time, U∞U_{\infty} is the reference (approach flow) velocity, ν~\tilde{\nu} refers to the (constant) kinematic viscosity of the fluid and LL denotes a reference length, typically the diameter DD of the investigated cylinder studied in this paper, cf. Figure 7. The employed reference pressure refers to Pref=ρ~U∞2P_{\mathrm{ref}}=\tilde{\rho}U_{\infty}^{2} and f=f~L/U∞2f = \tilde{f}L/U_{\infty}^{2} denotes a non-dimensional volume-specific body force.

Figure 7. Dimensionless illustration of the investigated, discretized 2D cylinder case with the grid lines indicated in light-gray and the exterior (interior) boundary of the physical domain Ω\Omega indicated in black (magenta) (cf. Appendix A).

The boundary of the object, denoted Γ\Gamma, is subject to Dirichlet (no-slip) conditions vΓ\boldsymbol{v}_{\Gamma} for the velocity and a natural (Neumann) condition for the pressure. The exterior boundaries of the domain ∂Ω\partial\Omega are assigned to Dirichlet conditions for both velocity and pressure, viz.

v−vΓ=0,∇p⋅n=0,on Γ,\boldsymbol{v}-\boldsymbol{v}_{\Gamma}=0,\qquad\nabla p\cdot\boldsymbol{n}=0,\qquad\text{on }\Gamma,
v−v∞=0,p=0,on ∂Ω.(12)\boldsymbol{v}-\boldsymbol{v}_{\infty}=0,\qquad p=0,\qquad\text{on }\partial\Omega. \tag*{(12)}

Initial conditions are prescribed by a spatially dependent v=vinit.\boldsymbol{v}=\boldsymbol{v}_{\mathrm{init.}} for t=0t=0. Boundary conditions are incorporated using a ghost-point approach. These conditions enter the discrete differential operators as additive correction terms that augment the interior FD stencil. Thereby, each operator application can be expressed as a matrix–vector product plus an explicit boundary-correction vector, while boundary points themselves are excluded from the solution vector [97]. Contributions to the source term ff as volume-specific body forces by the physical problem are neglected in the course of this work.

Discretization and Representation

The spatial discretization of the two-dimensional physical domain Ω\Omega is based on a structured O-type grid G\mathcal{G} [98, 99]. As outlined in Appendix A, the physical coordinates (xij,zij)(x_{ij},z_{ij}) are defined by a mapping to a unit-square computational domain Ω0\Omega_{0} using a simple transfinite interpolation approach [100]. The interior supporting points of the computational domain are located at

ηij=(i+1)δηwith i=0,…,2nη−1,ξij=jδξwith j=0,…,2nξ−1.(13)\begin{aligned} \eta_{ij} &= (i+1)\delta\eta&& \text{with } i=0,\ldots,2^{n_{\eta}}-1,\\ \xi_{ij} &= j\delta\xi&& \text{with } j=0,\ldots,2^{n_{\xi}}-1. \tag*{(13)} \end{aligned}

Hence, employing nηn_{\eta} and nξn_{\xi} TT-cores for the η\eta- and ξ\xi-directions, respectively. We set Nη=2nηN_{\eta}=2^{n_{\eta}} and Nξ=2nξN_{\xi}=2^{n_{\xi}} thereby obtaining an Nη×NξN_{\eta}\times N_{\xi} structured grid with uniform spacings δη=1/(Nη+1)\delta\eta=1/(N_{\eta}+1) and δξ=1/Nξ\delta\xi=1/N_{\xi} (cf. Appendix A). Note that the auxiliary indices i=−1i=-1 and i=Nηi=N_{\eta} coincide with the inner (top) and exterior (bottom) boundaries of Ω\Omega (Ω0\Omega_{0}) domain, while the ξ\xi-direction is treated periodically.

Next, we define a coordinate transformation Φ:Ω0↦Ω\Phi:\Omega_{0}\mapsto\Omega that maps the computational coordinates (ξij,ηij)(\xi_{ij},\eta_{ij}) to the physical coordinates (xij,zij)(x_{ij},z_{ij}) with xij=x(ξij,ηij)x_{ij}=x(\xi_{ij},\eta_{ij}) and zij=z(ξij,ηij)z_{ij}=z(\xi_{ij},\eta_{ij}) [99]. Using this transformation, we approximate the spatial derivatives in the momentum Eqs. (10) and continuity Eq. (11) by their analogues in the computational domain, thereby obtaining curvilinear forms of the differential operators.

2D Field Encoding

Before proceeding with the discretization in the TT format, we first introduce the encoding of two-dimensional scalar fields, here defined over the computational domain Ω0\Omega_{0}, as TT-vectors. Consider, for example, a scalar field a(ξij,ηij)a(\xi_{ij},\eta_{ij}), such as the xx-coordinate x(ξij,ηij)x(\xi_{ij},\eta_{ij}). To represent such a field as a TT-vector, a consistent enumeration of the grid points (ξij,ηij)(\xi_{ij},\eta_{ij}) must first be established. To each point (ξij,ηij)(\xi_{ij},\eta_{ij}), we assign a bit string r=ib∪jb\boldsymbol{r}=i_{b}\cup j_{b} of length n=nη+nξn=n_{\eta}+n_{\xi}, formed by concatenating ibi_{b}, the length nηn_{\eta}, binary representation of row index i∈{0,1,…,Nη−1}i\in\{0,1,\ldots,N_{\eta}-1\}, with jbj_{b}, the length nξn_{\xi} binary representation of column index j∈{0,1,…,Nξ−1}j\in\{0,1,\ldots,N_{\xi}-1\}. Note that ii and jj are the indices defined in Eq. (13). Each bit rkr_{k} is then interpreted as the two modes {0,1}\{0,1\} of the indices of an order-nn tensor ar1,r2,…,rn\boldsymbol{a}_{r_{1},r_{2},\ldots,r_{n}} that encodes N=2nN=2^{n} discretized field values as presented in Section II, where rkr_{k} is the kthk^{\mathrm{th}} bit of r\boldsymbol{r}. After decomposing ar1,r2,…,rn\boldsymbol{a}_{r_{1},r_{2},\ldots,r_{n}} for n−1n-1 times, as illustrated in Figure 6, one obtains the TT-vector representation of the field

Figure 6. Graphical representation of a TT decomposition of a 2D field. Here, iki_k (jkj_k) represents the value of the kthk^{\mathrm{th}} bit of bit string ibi_b (jbj_b).

a=A[1]r1A[2]r2…A[n]rn≃ar1,r2,…,rn=f(ξij,ηij).(14)\boldsymbol{a}=\boldsymbol{A}^{[1]r_{1}}\boldsymbol{A}^{[2]r_{2}}\ldots\boldsymbol{A}^{[n]r_{n}}\simeq\boldsymbol{a}_{r_{1},r_{2},\ldots,r_{n}}=f(\xi_{ij},\eta_{ij}). \tag*{(14)}

Vector fields, such as the velocity v\boldsymbol{v}, are treated by encoding xx- and zz-components of a field separately into two different TT-vectors.

While the aforementioned method yields a flexible encoding of various fields, the computational cost is proportional to the number of grid points N=NηNξN=N_{\eta}N_{\xi}. Practically, a field with recognized features (e.g., uniform, harmonic, exponential, ...) has a low-rank representation that allows a direct construction of its TT-cores that only costs O(log⁡(N)2)\mathcal{O}(\log(N)^{2}) [86], or it can be found through tensor cross interpolation [85]. For the simulations conducted in this paper, the TT-vectors of the initial flow fields are directly constructed within the TT format, thus avoiding the decomposition of order-nn tensors. By keeping the bond dimension fixed to a constant value throughout each step of the algorithm described in Section III C, the time evolution and subsequent analysis are performed without requiring the full numerical representation of the fields, thereby preserving the overall logarithmic scaling in NN.

Mesh Generation in TT Format

The discretization (xij,zij)(x_{ij}, z_{ij}) is classically defined by a grid G\mathcal{G}, which can either be read in or generated directly in the TT format. In either case, the number of interior grid points must satisfy the constraint 2nη×2nξ2^{n_\eta} \times2^{n_\xi} to be compatible with the TT format. To preserve the logarithmic scaling in the number of points NN, the latter approach is preferable, as it avoids the explicit construction and decomposition of full NN-component tensors, and is hence discussed in the following.

For a convex shape Γ\Gamma with supporting points on the surface given by (xj(Γ),zj(Γ))(x^{(\Gamma)}_j, z^{(\Gamma)}_j), the interior grid points can be constructed either uniformly along the surface normal direction, or using a non-uniform distribution with step sizes that gradually increase with the distance from the surface Γ\Gamma, e.g.,

si(lin)=Q(i+1)δη(linear)(15)s_i^{(\mathrm{lin})} = Q(i+1)\delta\eta\qquad\text{(linear)} \tag*{(15)}

or

si(exp)=Qκ(i+1)δη−1κ−1(exponential),(16)s_i^{(\mathrm{exp})} = Q\frac{\kappa^{(i+1)}\delta\eta- 1}{\kappa- 1} \qquad\text{(exponential)}, \tag*{(16)}

where QQ denotes the maximal normal distance to the surface Γ\Gamma and κ\kappa is a stretching parameter. For κ→0\kappa\to0, the exponential si(exp)s_i^{(\mathrm{exp})} spacing reduces to linear si(lin)s_i^{(\mathrm{lin})}. The grid coordinates are constructed as a tensor product of 1D functions defined along the η\eta- and ξ\xi-directions, respectively:

xij=1ixj(Γ)+sinjxand(17)x_{ij} = \mathbf{1}_i x_j^{(\Gamma)} + s_i n_j^x \qquad\text{and} \tag*{(17)}
zij=1izj(Γ)+sinjz,(18)z_{ij} = \mathbf{1}_i z_j^{(\Gamma)} + s_i n_j^z, \tag*{(18)}

where sis_i is either si(lin)s_i^{(\mathrm{lin})} or si(exp)s_i^{(\mathrm{exp})}, 1i\mathbf{1}_i denotes a vector of ones, and njxn_j^x, njzn_j^z are the xx- and zz-components of the unit normal vector at surface point jj. For the orthogonal cylinder mesh illustrated in Figure 7, Eqs. (17), (18) agree with the results of a slightly more involved classical transfinite interpolation method [101].

To translate Eqs. (17) and (18) into TTs, the NξN_\xi-component vectors containing to the xx- and zz-coordinates of the object’s surface Γ\Gamma, xj(Γ)x_j^{(\Gamma)} and zj(Γ)z_j^{(\Gamma)}, as well as the components of the unit normal vector, njxn_j^x and njzn_j^z, with j=0,…,Nξ−1j = 0, \ldots, N_\xi- 1 must each be encoded as TT-vectors of length nξn_\xi. These are denoted as x(Γ)x^{(\Gamma)}, z(Γ)z^{(\Gamma)}, nxn^x, and nzn^z, respectively. Furthermore, the TT-vectors of length nηn_\eta corresponding to 1i\mathbf{1}_i and sis_i, with i=0,…,Nη−1i = 0, \ldots, N_\eta- 1, are analytically prescribed and low-rank. The TT-cores of 1i\mathbf{1}_i and linear spacing si(lin)s_i^{(\mathrm{lin})} can be constructed following Ref. [86], while the strategy for the exponential spacing si(exp)s_i^{(\mathrm{exp})} is described in Ref. [49], and denoting them by 1\mathbf{1} and ss. The final TT-vectors of the interior coordinates (xij→x(x_{ij} \to x and zij→z)z_{ij} \to z) are then realized by concatenating the above TT-vectors as

x=1∪x(Γ)+s∪nxand(19)x = \mathbf{1} \mathbin{\cup} x^{(\Gamma)} + s \mathbin{\cup} n^x \qquad\text{and} \tag*{(19)}
z=1∪z(Γ)+s∪nz,(20)z = \mathbf{1} \mathbin{\cup} z^{(\Gamma)} + s \mathbin{\cup} n^z, \tag*{(20)}

where ∪\mathbin{\cup} here denotes a∪b=A[1]…A[nη]B[1]…B[nξ]a \mathbin{\cup} b = A[1] \ldots A[n_\eta]B[1] \ldots B[n_\xi]. The resulting TT-vector is then of length nη+nξn_\eta+ n_\xi.

Constructing Curvilinear Operators in TT Format

The construction of curvilinear operators as TT-matrices exemplarily presented for the physical-space derivative ∂x\partial_x follows (cf. Appendix A)

∂x=1J((∂ηz)∂ξ−(∂ξz)∂η),(21)\partial_x = \frac{1}{J}\left((\partial_\eta z)\partial_\xi- (\partial_\xi z)\partial_\eta\right), \tag*{(21)}

where JJ is the Jacobian determinant J=(∂ξx)(∂ηz)−(∂ηx)(∂ξz)J = (\partial_\xi x)(\partial_\eta z) - (\partial_\eta x)(\partial_\xi z). The same construction strategy applies analogously to the physical-space derivative ∂z\partial_z as well as to the Laplacian operator Δ\Delta. We note that all computational-space derivatives are discretized using a second-order central FD scheme (cf. Appendix A).

Given the TT-vector representations of the internal grid points, χ\chi and zz, as detailed in Section III B2, the derivative operators ∂ξ\partial_\xi and ∂η\partial_\eta can be directly constructed as low-rank TT-matrices [88], Dξ\mathcal{D}_\xi and Dη\mathcal{D}_\eta, respectively. Accordingly, the TT-vector jj of the Jacobian determinant JJ is computed as

j=(Dξχ)(Dηz)−(Dηχ)(Dξz).(22)j = (\mathcal{D}_\xi\chi)(\mathcal{D}_\eta z) - (\mathcal{D}_\eta\chi)(\mathcal{D}_\xi z). \tag*{(22)}

Next, the computation of c1=(Dηz)⊙j−1c_1 = (\mathcal{D}_\eta z) \odot j^{-1} and c2=(Dξz)⊙j−1c_2 = (\mathcal{D}_\xi z) \odot j^{-1} is executed where the inverse j−1j^{-1} is found as the solution of the LSE Ja=1Ja = 1, where JJ denotes the diagonal TT-matrix representation of jj and 11 a TT-vector of ones. Thus c1c_1 and c2c_2 are found by a combination of known operations, cf. Section II A 3-II A 4. As a final step, the prefactors c1c_1, c2c_2 represented as TT-vectors, need to be multiplied into the TT-matrices Dξ\mathcal{D}_\xi and Dη\mathcal{D}_\eta to construct the partial derivative in xx-direction in TT format, viz. Dx=c1Dξ−c2Dη\mathcal{D}_x = c_1\mathcal{D}_\xi- c_2\mathcal{D}_\eta. This operation is realized as

Dx=C1Dξ−C2Dη,(23)\mathcal{D}_x = C_1\mathcal{D}_\xi- C_2\mathcal{D}_\eta, \tag*{(23)}

where C1C_1 (C2C_2) are diagonal TT-matrices that have c1c_1 (c2c_2) on its diagonal using the δrk,rk′\delta_{r_k,r_k^{\prime}} tensor as introduced in Section II A 3. The matrix-matrix product C1DξC_1\mathcal{D}_\xi is performed similarly to a TT-matrix with TT-vector contraction with changes w.r.t. the physical dimension, and, in the worst case, scaling as O(χ(Dξ)χ(C1)3)\mathcal{O}(\chi(\mathcal{D}_\xi)\chi(C_1)^3). Note that the bond dimension χ(C1)\chi(C_1) generally depends on the grid coordinates, while χ(Dξ)\chi(\mathcal{D}_\xi) is low-rank and independent of the grid [88]. Moreover, they are independent of the flow regime, as it requires the construction of the curvilinear operators only once for a given FD scheme and discretization. Lastly, in constructing the curvilinear operators as TT-operators, a truncation is performed after each TT operation by discarding singular values below 10−1410^{-14} (cf. Section II).

Fractional Step Method

A fractional step method [97] is used to advance the flow field described by the momentum Eqs. (10) and the continuity Eq. (11) from t(m)t^{(m)} to t(m+1)=t(m)+δtt^{(m+1)} = t^{(m)}+\delta t. In this approach, the pressure gradient-free momentum equations are used to determine an intermediate velocity v∗\boldsymbol{v}^{*} [102]

v∗−v(m)δt+(v(m)⋅∇)v(m)=1ReΔv∗⟺(1−δtReΔ)v∗=v(m)−δt(v(m)⋅∇)v(m).(24)\frac{\boldsymbol{v}^{*}-\boldsymbol{v}^{(m)}}{\delta t}+\left(\boldsymbol{v}^{(m)}\cdot\nabla\right)\boldsymbol{v}^{(m)}=\frac{1}{Re}\Delta\boldsymbol{v}^{*} \Longleftrightarrow \left(1-\frac{\delta t}{Re}\Delta\right)\boldsymbol{v}^{*}=\boldsymbol{v}^{(m)}-\delta t\left(\boldsymbol{v}^{(m)}\cdot\nabla\right)\boldsymbol{v}^{(m)}. \tag*{(24)}

The intermediate velocity v∗\boldsymbol{v}^{*} is subsequently corrected by solving a pressure Poisson equation [97], which is controlled by the divergence of the intermediate velocity field, i.e.,

Δp(m+1)=1δt∇⋅v∗.(25)\Delta p^{(m+1)}=\frac{1}{\delta t}\nabla\cdot\boldsymbol{v}^{*}. \tag*{(25)}

The gradient of the computed pressure p(m+1)p^{(m+1)} is used to enforce the continuity constraint for the velocity at the next time step t(m+1)t^{(m+1)} with

v(m+1)=v∗−δt∇p(m+1).(26)\boldsymbol{v}^{(m+1)}=\boldsymbol{v}^{*}-\delta t\nabla p^{(m+1)}. \tag*{(26)}

Instead of an adaptive, Courant-based time step, δt\delta t is kept constant to maintain the efficiency in the TT format. The resulting method is equivalent to the fractional step method originally proposed by Chorin [76]. The Poisson equation (25) is solved in TT format using the variational optimization outlined in Section II A 4. This procedure is particularly well suited to elliptic equations, and we include the Jacobian determinant JJ to account for the varying cell sizes in the curvilinear discretization [94]. Consequently, the modified optimization problem for the Poisson Eq. (25) reads

min⁡aC(a):=12(a(jM(Δ)a)−a(j⊙b)).(27)\min_{\boldsymbol{a}} C(\boldsymbol{a}) := \frac{1}{2}\left(\boldsymbol{a}\left(j\mathcal{M}^{(\Delta)}\boldsymbol{a}\right)-\boldsymbol{a}(j\odot\boldsymbol{b})\right). \tag*{(27)}

To facilitate the direct comparison of the TT solver with its classical counterpart, the fractional step method described above is implemented twofold. The first implementation follows the TT methodology, while the second refers to a classical FD implementation, which is verified against literature and finite-volume results generated with OpenFOAM. The interested reader finds more information in Appendix H.

Applications

In this section, we present a detailed comparison between our TT solver and a reference classical FD solver for the paradigmatic 2D cylinder flow, considering both a non-rotating and a rotating cylinder. We start by introducing the specific problem setup of the cylinder and its discretization. Further we introduce flow characteristics and diagnostic quantities used to evaluate solution accuracy and computational efficiency. Results of the first Subsection IV A are dedicated to the non-rotating cylinder in steady (Re=20Re = 20) and transient (Re≥50Re \ge50) regimes. The subsequent Subsection IV B discusses the transient flow around the rotating cylinder (Re=100Re = 100). Section IV C complements the numerical results with insights into computational performance. All runs were carried out on 8 cores of an AMD EPYC 9654 CPU and 32 GB of RAM, except stated otherwise.

Figure 7 depicts the test case for a cylinder that features a unit diameter DD boundary Γ\Gamma which is centered in a circular domain Ω\Omega with the boundary ∂Ω\partial\Omega placed with diameter 32D32D. The spatial discretization refers to a block-structured O-type grid, as shown by the light-gray grid lines in Figure 7. Three successively refined grids, each equidistant in the circumferential direction, are considered:

G1:N=256×256, nη=8∧nξ=8,\mathcal{G}_1 : N = 256 \times256,\ n_{\eta}=8\land n_{\xi}=8,
G2:N=256×512, nη=8∧nξ=9,\mathcal{G}_2 : N = 256 \times512,\ n_{\eta}=8\land n_{\xi}=9,
G3:N=512×512, nη=9∧nξ=9,(28)\mathcal{G}_3 : N = 512 \times512,\ n_{\eta}=9\land n_{\xi}=9, \tag*{(28)}

with minimal grid spacings at the cylinder surface, each taken as the smaller of the normal spacing or the circumferential arc length, that are approximately hmin⁡≈6.78×10−3h_{\min}\approx6.78\times10^{-3}, hmin⁡≈6.14×10−3h_{\min}\approx6.14\times10^{-3} and hmin⁡≈3.39×10−3h_{\min}\approx3.39\times10^{-3}, respectively. Here, hmin⁡=hmin⁡/Dh_{\min}=h_{\min}/D denotes the smallest step size in the physical domain which usually occurs at the interior boundary in the radial direction. The grid points are specified in Appendix E. While the grids can be constructed directly within the TT format as outlined in Section III B 2, in this study we construct the grid points (xij,zij)(x_{ij},z_{ij}) and the associated metric terms in a classical way and subsequently encode them into TT-vectors. This ensures comparability between FD and TT-based solvers, which therefore operate with identical coordinate transformations. The computational overhead due to the transformation of the classical grid into the TT format is neglected here. Yet, its advantageous efficiency should be exploited for any other application.

The assessment of transient flow phenomena employs the Strouhal number

St=ω~D2πU∞,(29)St = \frac{\widetilde{\omega}D}{2\pi U_{\infty},} \tag*{(29)}

where ω~\widetilde{\omega} [rad/s] is the circular frequency of the oscillating lift force F~L\widetilde{F}_{L}. The lift force F~L\widetilde{F}_{L} is defined as the force per unit length acting perpendicular to the approach flow, whereas the drag force F~D\widetilde{F}_{D} acts parallel to it. The corresponding non-dimensional coefficients read

CL=2F~Lρ~U∞2DandCD=2F~Dρ~U∞2D.(30)C_{L} = \frac{2\widetilde{F}_{L}}{\widetilde{\rho}U_{\infty}^{2}D} \quad\text{and}\quad C_{D} = \frac{2\widetilde{F}_{D}}{\widetilde{\rho}U_{\infty}^{2}D}. \tag*{(30)}

The predictive accuracy is assessed by the L2L_{2}-norm of a scalar field s(ξ,η)s(\xi,\eta), which accounts for the variable grid spacing across the domain, i.e.,

∥s∥L2=∫Ω0∣s(ξ,η)∣2J dξdη≈δξδη∣sij∣2Jij.(31)\lVert s\rVert_{L_{2}} = \sqrt{\int_{\Omega_0} \lvert s(\xi,\eta)\rvert^{2}J\,\mathrm{d}\xi\mathrm{d}\eta} \approx \sqrt{\delta\xi\delta\eta\lvert s_{ij}\rvert^{2}J_{ij}}. \tag*{(31)}

Consequently, a relative difference between a TT solution ss and a classical solution sˉ\bar{s} reads

ϵL2=∥s−sˉ∥L2∥sˉ∥L2,(32)\epsilon_{L_{2}} = \frac{\lVert s-\bar{s}\rVert_{L_{2}}}{\lVert\bar{s}\rVert_{L_{2}}}, \tag*{(32)}

and is used to evaluate the error of the TT format. We also consider the maximal difference error

ϵmax⁡=∥s−sˉ∥∞.(33)\epsilon_{\max} = \lVert s-\bar{s}\rVert_{\infty}. \tag*{(33)}

Moreover, time averages are denoted by an overline, e.g., the error ϵL2‾\overline{\epsilon_{L_{2}}}. To quantify the compression achievable through the TT representation, we introduce the number of variables parameterizing the solution (NVPS), which is the total number of values needed to be stored to represent a state [58]. The NVPS of a TT-vector depends on the dimensions of its indices and is computed by

NVPSTT=2∑k=1nαk−1αk,(34)\mathrm{NVPS}_{\mathrm{TT}} = 2\sum_{k=1}^{n}\alpha_{k-1}\alpha_{k}, \tag*{(34)}

where αk\alpha_k denotes again the bond index between the kk-th and (k+1)(k+1)-th TT-core, and n=log⁡(N)n=\log(N) the number of TT-cores. Therefore, the NVPS for a TT-vector grows only as O(nχ2)O(n\chi^2) [58]. In contrast, in a classical FD description, the solution vector length grows exponentially as O(2n)O(2^n).

Non-Rotating Cylinder

Within this section, the non-rotating cylinder is modeled by zero velocity conditions such that vΓ=0\boldsymbol{v}_{\Gamma}=0 in Eq. (12). The initial Section IV A 1 compares the INSE solutions (10-12) obtained from the classical FD method with TT-based solutions for the steady state flow at Re=20Re=20, and scrutinizes the accuracy of the TT’s steady-state solution for several bond dimensions. The subsequent Section IV A 2 investigates the transient regime for increasing Reynolds numbers, where emphasis is given to the ability of TT to accurately predict vortex shedding and the related oscillating forces.

Steady Flow (Re=20)

To study the effects introduced by the compression inherent to the TT format, the differences between both strategies are analyzed for increasing bond dimensions χ\chi. We note that the bond dimension is imposed uniformly on the TT-vectors of the two cartesian velocities uu, vv, and the pressure pp, such that each field has the same value of χ\chi. For the sake of completeness, the accuracy of the classical FD code is verified against a frequently used finite volume procedure in Appendix § 1.13. For the stability criterion of the convective term, a maximum velocity as Umax⁡=γU∞U_{\max}=\gamma U_{\infty} is estimated following the usual practice in the literature (cf. [97]); here we adopt γ=1.5\gamma=1.5. Accordingly, the simulation on the medium-fine grid G2G2 is advanced with a time step δt=δt~D/U∞≈3.8×10−3\delta t=\tilde{\delta t}D/U_{\infty}\approx3.8\times10^{-3}, so that the resulting convective Courant number remains below unity.

Starting from a uniform initial condition with vinit=(U∞,0)T\boldsymbol{v}_{\mathrm{init}}=(U_{\infty},0)^{\mathrm{T}}, solutions are extracted after the system evolves over 8100 steps in pseudo time, which corresponds to more than 30 passage times D/U∞D/U_{\infty}. Figure 8(a) displays the steady-state streamlines and horizontal-velocity (uu) contours at Re=20Re=20 predicted by the classical FD method, while Figure 8(b) shows the corresponding TT solution with χ=20\chi=20. In both panels the u=0u=0 isoline is depicted, solid red for the FD and dashed green for the TT. Moreover, Figure 8(b) includes an inset that magnifies these isolines for direct comparison. Qualitatively, the predicted fields show excellent agreement, and the classical Föppl vortex pair behind the cylinder is correctly captured.

Figure 8. Streamlines and horizontal velocity (uu) contours for the steady-state flow at Re=20Re = 20, predicted by (a) the classical FD method and (b) the TT method with χ=20\chi= 20. The u=0u = 0 isoline is highlighted in solid red for FD and dashed green for TT, with an inset in panel (b) showing a close-up comparison of both isolines. Predictive difference ϵL2(32)\epsilon_{\mathrm{L2}}^{(32)} between the velocity magnitude field (c) and the individual Cartesian velocity components uu, vv (d) obtained from the TT and the classical FD simulations for the flow around the cylinder at Re=20Re = 20 using the medium-fine grid G2\mathcal{G}2. Panel (c) shows the temporal evolution for different bond dimensions χ\chi. Panel (d) depicts the steady state results for uu (circles) and vv (squares), supplemented by the NVPS fraction for each χ\chi (crosses) to indicate the corresponding compression rate.

Figure 8(c) depicts the evolution of the relative difference (error) ϵL2\epsilon_{L_{2}} between the velocity magnitude obtained from TT-based simulations and FD predictions, evaluated across a range of fixed bond dimensions χ\chi. The error increases during the initial transient phase before stabilizing after around 25 passage times at distinct levels for each χ\chi, remaining below 2×10−32\times10^{-3}. The maximum difference in uu is bounded by ϵmax⁡≤9.9×10−3\epsilon_{\max}\leq9.9\times10^{-3} (χ=20\chi=20), decreasing to 2.2×10−72.2\times10^{-7} (χ=50\chi=50). Similarly, the vv-field satisfies ϵmax⁡≤1.3×10−2\epsilon_{\max}\leq1.3\times10^{-2} (χ=20\chi=20), dropping to ϵmax⁡≤7.9×10−8\epsilon_{\max}\leq7.9\times10^{-8} (χ=50\chi=50). The negligible lift coefficient of ∣CL∣=1×10−4\lvert C_{L}\rvert=1\times10^{-4} for the FD solver exhibits its largest deviation w.r.t. the TT solution at χ=25\chi=25 with ∣CL∣=3×10−3\lvert C_{L}\rvert=3\times10^{-3}. For all the bond dimension χ>30\chi>30 the lift coefficient is upper bounded by ∣CL∣≤1.7×10−4\lvert C_{L}\rvert\leq1.7\times10^{-4}, closely resembling the FD solution. Moreover, in the energy norm commonly used to assess ROMs (cf. [103]), the steady-state relative error in total kinetic energy Ekin=12∥v∥L22E_{\mathrm{kin}}=\frac{1}{2}\lVert\boldsymbol{v}\rVert_{L_{2}}^{2} compared to FD decays rapidly with bond dimension, from 7.65×10−57.65\times10^{-5} at χ=20\chi=20, to 7.16×10−87.16\times10^{-8} at χ=35\chi=35, and down to 3.62×10−113.62\times10^{-11} at χ=50\chi=50. Figure 8(d) shows the final differences in uu and vv after 8100 time steps.

In general, a remarkable agreement between the TT and FD solution is observed. It is evident that with approximately 5.8% of the NVPS, the TT-based approach achieves a relative error of 3×10−33 \times10^{-3} (or 0.3%0.3\%) in the velocity magnitude, achieving an outstanding 20-fold compression. Increasing the NVPS fraction to approximately 15%15\% further reduces the relative error by three magnitudes to 5.8×10−65.8 \times10^{-6}. In the most demanding case with χ=50\chi= 50, the method still requires only around 1/41/4 of the NVPS while achieving highly accurate solutions with relative errors on the order of 10−810^{-8}. An extrapolation of the errors suggests that the accuracy can in principle be reduced to machine precision by further increasing the bond dimension. More details will be found in the analysis of the time-dependent flow in the following section.

Transient Flow

We now proceed to analyze unsteady flow regimes that emerge for Re≥50Re \ge50. Our goal is to demonstrate that TT can not only efficiently capture simple steady-state flows but also faithfully reproduce transient dynamics. In particular, we investigate the trade-off between parametrization (NVPS) complexity and pointwise solution errors, noting that exact local accuracy may not be required to capture the essential dynamics, aligning with the core idea of ROMs. The simulations apply the two coarser grids G1\mathcal{G}1 and G2\mathcal{G}2 with a time step of δt=δt~U∞/D≈3.4×10−3\delta t = \widetilde{\delta t}U_{\infty}/D \approx3.4 \times10^{-3} and δt=δt~U∞/D≈3.1×10−3\delta t = \widetilde{\delta t}U_{\infty}/D \approx3.1 \times10^{-3}, respectively. Here, the maximum velocity is estimated with γ=2\gamma= 2, i.e., Umax⁡=γU∞U_{\max} = \gamma U_{\infty}, so that the resulting Courant number remains below unity for all ReRe considered. Expecting a Strouhal number of St≈0.2St \approx0.2 corresponds to a temporal resolution of approximately 60 time steps per period.

Two algorithmic modifications are employed to compute the transient flows: (1) we trigger transient effects by employing asymmetries within initial conditions. To this end, we add an initial perturbation upert.u_{\mathrm{pert.}} to the initial velocity field, which consists of a 5%5\% velocity modulation (as given in Appendix G) and (2) we apply a pre-conditioning of the Poisson equation within the TT format, allowing to efficiently simulate higher bond dimensions [54].

Figure 9 displays snapshots of the TT for χ=40\chi= 40 and the FD-predicted velocity magnitudes ∣v∣|\mathbf{v}| for Re=150Re = 150. Results refer to the coarse grid G1\mathcal{G}1 at around 84.85 passage times, i.e., approximately 2.5×1042.5 \times10^{4} time steps when the flow has reached a periodic state. Qualitatively, both velocity fields are almost indistinguishable from each other, demonstrating that the TT simulation at χ=40\chi= 40 is capable of capturing the Kármán vortex street similar to the classical approach. Figure 10 illustrates the corresponding pointwise differences of the velocity magnitude ∣v∣|\mathbf{v}|, defined as ∣∣vTT∣−∣vFD∣∣||\mathbf{v}^{\mathrm{TT}}| - |\mathbf{v}^{\mathrm{FD}}||, in the region x~/D∈[−2,12]\tilde{x}/D \in[-2,12], z~/D∈[−4,4]\tilde{z}/D \in[-4,4]. The bond dimension χ\chi increases from top to bottom across the panels, with χ=40\chi= 40, 6060, and 8080. Results are again evaluated after around 2.5×1042.5 \times10^{4} time steps, when the flow reflects a periodic behavior in time. Figure 10 demonstrates that each increase of χ\chi by 20 bond units reduces the local deviation by roughly two orders of magnitude. For instance, at the location (x~/D,z~/D)=(8,2)(\tilde{x}/D,\tilde{z}/D)=(8,2) (marked by ‘×’ in Figure 10), the deviation reads 5.2×10−35.2\times10^{-3}, 6.3×10−56.3\times10^{-5}, and 9.7×10−89.7\times10^{-8} for increasing values of χ\chi. This aligns with the convergence of TT towards FD in the limit of increasing χ\chi. For the given system, the maximum possible bond dimension is χ=256\chi=256. With a reduction to χ=80\chi=80 and a corresponding 87% of NVPS compared to FD, a maximum error of 2.54×10−62.54\times10^{-6} in the velocity magnitude ∣v∣|v| is recovered, cf. Figure 10(c). Nevertheless, lower χ\chi values are sufficient to capture the essential physics for this ReRe, as it was concluded before. This is confirmed by the error patterns in the cylinder’s wake. As expected, the error patterns are aligned to the wake and appear similar for different values of χ\chi. This indicates a consistent resolution of the dynamics with only differences in the resolution accuracy, and underlines the suitability of the TT-based approach as a ROM.

Figure 10. Pointwise difference of the velocity magnitude ∣v∣|v| predicted by the TT and classical FD method for the flow around a non-rotating cylinder at Re=150Re=150 on the coarse grid G1\mathcal{G}1. White crosses indicate the probe location (x~/D,z~/D)=(8,2)(\tilde{x}/D,\tilde{z}/D)=(8,2), for the comparison of velocity components.

The capability of TTs to predict the relevant integral force quantities is assessed next. To this end, the Strouhal number is determined from the oscillations of the lift force coefficient CLC_L (30). Accordingly, the computation of CLC_L can be performed entirely in the TT format, without requiring an expensive contraction to the full classical representation. Specifically, it involves extracting one-dimensional TT-vectors of η\eta-constant lines adjacent to the surface Γ\Gamma (cf. Ref. [58]), combining TT operations outlined in Section II, and applying a summation (average-like) operation as reported in [60]. Furthermore, the mapping from the computational domain

Ω0\Omega_0 to the physical domain Ω\Omega must again be taken into account, as detailed in Appendix F.

Figure 11 presents the temporal evolution of CLC_L over 50 passage times for Re=100Re=100, 150, and 200. The figure compares the results obtained from the TT approach on the medium grid G2\mathcal{G}2 for two different bond dimensions, i.e., χ=40\chi=40 (orange) and χ=60\chi=60 (magenta), and the FD method (gray). The TT results clearly capture distinct oscillations. However, the agreement between χ=40\chi=40 and FD slightly deteriorates as ReRe increases. Notably, the phase seems to be shifting during the simulation of the higher ReRe number in combination with χ=40\chi=40. This is attributed to an expected increase of the required bond dimension with ReRe at the same accuracy, as also seen in Ref. [58].

Temporal evolution of the lift coefficient $C_{\mathrm{L}}$ for the flow around a non-rotating cylinder

Figure 11. Temporal evolution of the lift coefficient CLC_{\mathrm{L}} for the flow around a non-rotating cylinder on the medium grid G2G_2. Results predicted by the TT method are shown for bond dimensions χ=40\chi= 40 (orange dashed) and χ=60\chi= 60 (magenta dashed), with the FD reference solution (gray solid).

To study how the observed deterioration in Figure 11 affects the accuracy of the predicted Strouhal numbers, the angular frequencies of the lift forces were evaluated for a range of bond dimensions for the two coarser grids G1G_1 and G2G_2, respectively. Results were obtained from successive (interpolated) zero crossings of CLC_{\mathrm{L}} over 50 passage times (t∈[50,100]t \in[50,100]); these crossings were then used to compute the averaged shedding period. The difference between Strouhal numbers obtained from the TT simulations and the corresponding FD are presented in Figure 12 for bond dimensions χ=40\chi= 40, 60 and χ=80\chi= 80. As a point of reference, we obtain St≈0.1916St \approx0.1916 for Re=200Re = 200 from the FD results using G2G_2 which presents a 10−310^{-3} absolute error w.r.t. the value of St≈0.1977St \approx0.1977 reported in Ref. [97]. This allows us to conclude that the prediction by TT is indeed generally very accurate, as confirmed by Figure 12. Results demonstrate that as χ\chi increases, the difference in the predicted Strouhal number decreases monotonically, consistent with the observed decrease in pointwise field differences shown in Figure 10. Moreover, we observe the expected trend where, for a fixed system size and bond dimension, the error in StSt increases with growing ReRe. Despite this, the TT-based model is particularly promising, as increases in the bond dimension χ\chi can provide increasingly accurate Strouhal number estimates, which may be used in a Richardson extrapolation scheme to approximate the exact value.

Figure 12. Absolute difference of the Strouhal number StSt predicted by the TT and the classical FD method for the flow around a non-rotating cylinder on the two coarser grids G1G_1 (straight) and G2G_2 (dot-dashed).

Rotating Cylinder

As a further demonstration of the framework’s versatility, we examine the case of a cylinder rotating with the non-dimensional rotation rate θ˙=Dθ˙~/(2U∞)=1\dot{\theta} = D\tilde{\dot{\theta}}/(2U_{\infty}) = 1, which gives rise to the widely studied phenomenon known as the Magnus effect. To this end, the boundary condition of the cylinder is changed to a no-slip condition with a non-vanishing prescribed tangential velocity

UΓ=θ˙U∞.(35)U_{\Gamma} = \dot{\theta}U_{\infty}. \tag*{(35)}

The components of a normalized velocity vΓ\mathbf{v}_{\Gamma} at the discrete surface points are given by

uΓj=θ˙⋅sin⁡(2πj δξ)andvΓj=−θ˙⋅cos⁡(2πj δξ),(36)\begin{aligned} u_{\Gamma j} &= \dot{\theta}\cdot\sin(2\pi j\,\delta\xi) \quad\text{and}\\ v_{\Gamma j} &= -\dot{\theta}\cdot\cos(2\pi j\,\delta\xi), \tag*{(36)} \end{aligned}

where j=0,…,Nξ−1j = 0,\ldots,N_{\xi}-1. The initial conditions are set to a uniform flow field uinit=(U∞,0)\mathbf{u}_{\mathrm{init}}=(U_{\infty},0) with no added perturbation, as employed in Section IV A 1, since the asymmetry of the boundary conditions is sufficient to trigger the vortex shedding at θ˙=1\dot{\theta}=1 [104].

The application of Eq. (36) realizes a counter-clockwise rotation. As a consequence of the rotation, the cylinder is subject to time-averaged circulation, which mainly results in a shift of the oscillating lift force. Figure 13 shows the temporal evolution of the lift-force coefficient CLC_{\mathrm{L}} (30) for θ˙=1\dot{\theta}=1 (magenta) in comparison to the non-rotating case (θ˙=0\dot{\theta}=0, orange) predicted by the TT method for Re=100Re=100 on the coarse grid G1G_1. The maximum velocity is estimated with γ=3\gamma=3, i.e., Umax⁡=γU∞U_{\max}=\gamma U_{\infty}, so that the resulting Courant number is below unity. Accordingly, the employed time step refers to δt=δt~U∞/D=2×10−3\delta t=\delta\tilde{t}U_{\infty}/D=2\times10^{-3}. The TT-predicted time-averaged lift force C‾L\overline{C}_{\mathrm{L}} and the lift Strouhal number are outlined in Table II for three different bond dimensions χ=40\chi=40, 50, and 60. These predictions compare favorably with our FD-predicted and literature-reported values extracted from [104, 105] which were obtained for a similar resolution.

Discretization (interior)χ\chiC‾L\overline{C}_{L} absC‾L\overline{C}_{L} %StSt absStSt %
241 ×\times 241–2.4881–0.1655–
256 ×\times 256402.49190.150.16620.42
256 ×\times 256502.49080.110.16630.49
256 ×\times 256602.49080.110.16630.49
256 ×\times 256FD2.49080.110.16630.49

Table II. Comparison of C‾L\overline{C}_{L} and lift StSt values for the flow around a rotating cylinder at Re=100Re = 100, θ˙=1\dot{\theta} = 1. Values reported by Kang et al. [105] (bold) are compared with values obtained using the TT method with different bond dimensions χ\chi and the classical FD solver.

Image not extracted

Figure 13. Temporal evolution of the TT-predicted (χ=40\chi= 40) lift-force coefficient CLC_{L} for the flow around a rotating cylinder (UΓ=U∞θ˙U_{\Gamma} = U_{\infty}\dot{\theta}) at Reynolds Re=100Re = 100 on the coarse grid G1\mathcal{G}_{1}.

It is noteworthy, that the TT approximation used for the low relative error of 0.15% for C‾L\overline{C}_{L} and 0.42% for the Strouhal prediction just uses χ=40∼30%\chi= 40 \sim30\% of NVPS on the coarse grid G1\mathcal{G}_{1}. A convergence towards the corresponding FD solver is observed as χ\chi increases, whereas convergence toward literature-reported values is not observed, which might be attributed to differences in the simulation methods and grids. Nonetheless, we conclude that the TT-predicted values exhibit excellent agreement with the literature while reducing the computational effort at least by a third.

Figure 14 shows the corresponding velocity magnitude fields predicted by the FD solver (a,b) and the TT solver with χ=40\chi= 40 (c,d), and the respective pointwise error distribution (e,f) analogously to Figure 10. The depicted snapshots clearly describe the formation of the wake and include the initial transient state (t≈4t \approx4) and the final periodic state (t≈15t \approx15). The deflection of the flow is visible and differs from the structure in Figure 9, which is due to the rotation of the cylinder. We observe well-defined vortices are convected downstream in the wake region. An increase in errors is detected in time, which we attribute to the accumulation of numerical noise and truncation errors in the TT format.

Figure 9. Snapshots of velocity magnitude ∣v∣|v| extracted at tD/U∞=84.85tD/U_{\infty}=84.85 from TT in (a) (χ=40\chi=40) and a classical FD (b) simulation at Re=150Re=150 when using the coarse grid G1\mathcal{G}1.

Image not extracted

Figure 14. Velocity magnitude snapshots predicted by the FD method (a,b) and the TT method (χ=40\chi= 40, c,d) on the coarse grid G1\mathcal{G}_{1} for the flow around a rotating cylinder (θ˙=1\dot{\theta} = 1) at Re=100Re = 100. Panels (e) and (f) show the corresponding pointwise errors between both methods. All error values below 10−710^{-7} are rendered as black in the visualization.

Computational Characteristics & Performance

In this section, we examine the computational effort of the algorithm. We confine the analysis to the assessment of the transient, non-rotating cylinder case at Re=150Re = 150 (cf. Section IV A 2) and assume that changes of the physical model are not affecting the parameterization of the numerics.

Following the fractional step approach introduced in Section III C, the three most computationally expensive operations are identified. These are (i) computing the RHS of the rearranged momentum Eqs. (24), involving elementwise multiplications of TT-vectors to evaluate the convection term, which scales as O(χ4)\mathcal{O}(\chi^{4}), (ii) solving an LSE (cf. Eq. (24)) for both components of v∗\mathbf{v}^{*}, scaling as O(χ3)\mathcal{O}(\chi^{3}), and (iii) solving the Poisson Eq. (25) with O(χ3)\mathcal{O}(\chi^{3}). The given scalings are taken from Table I. The breakdown of the absolute computational effort required by these three most expensive operations is evaluated over 2.4×1042.4 \times10^{4} time steps and displayed in Figure 15.

Figure 15. Illustration of the computing time of the three most expensive algorithmic tasks per time step for the flow around a cylinder at Re=150Re = 150 on the coarse grid G1G1 for three different bond dimensions.

As evident in Figure 15, each computational effort (i)–(iii) grows with rising χ\chi while the ratios between the operations remain nearly constant over time. The results of Figure 15 indicate that the wall-clock time per time step remains nearly constant throughout the simulation. We attribute this behavior to the Poisson preconditioner, which enables a rapid—if not fully optimal—pressure projection even when the Poisson system is ill-conditioned and the RHS is noisy.

In contrast, an unpreconditioned DMRG solver performs less effectively with the truncation and rounding errors that accumulate over time, leading to an increase in the number of DMRG sweeps and, consequently, a higher cost per step.

With a focus on the individual operations, solving the linear system (ii), for v∗v^* requires the least computational effort, significantly less than solving the Poisson problem (iii), despite both being LSEs (cf. Figure 15). This discrepancy can be attributed to the better conditioning of the left-hand side matrix (1+δtRe∇2)(1 + \frac{\delta t}{Re}\nabla^2) in the momentum update, resulting in faster convergence in the variational optimization. Moreover, we observe that computing the RHS (i) via elementwise multiplications of TT-vectors has roughly the same computational effort as solving the Poisson problem (iii).

All these operations scale with the bond dimension χ\chi of the fluid field representations. However, it is important to emphasize that the overall complexity also depends on the bond dimensions of the TT-matrices representing the differential operators (cf. Table I). Accordingly, Figure 16a illustrates the bond dimension of the curvilinear operators as TT (dashed line) w.r.t. the corresponding operators defined on a Cartesian mesh (solid line). As it is evident in this figure, the generalized operators require a comparable bond dimension, increasing only slightly for larger grid sizes. Furthermore, the number of NVPS of the TT (NVPSTT\mathrm{NVPS}^{\mathrm{TT}}) exhibits a remarkable reduction – up to three orders of magnitude – compared to the conventional sparse matrix representation (cf. Figure 16b). For this comparison, we consider a sparse matrix size of NVPSFD=5N\mathrm{NVPS}^{\mathrm{FD}} = 5N that represents the central point and its four neighboring points in the second-order central FD approximation of the Laplacian operator. Diagonal neighbors are not considered, due to the orthogonality of the grid.

Figure 16. Compressibility analysis of the generalized TT-operators. Panel (a) depicts the required bond dimension of the curvilinear and Cartesian operators. Panel (b) shows the ratio of NVPS needed to represent the curvilinear operators in TT and sparse format. Please note that the lines connecting the data points are added solely for better visualization of the emerging trends.

The wall-clock time of our TT-based approach is compared to the classical FD in Figure 17. As it is depicted in this figure, both approaches reach similar computational times for the given system sizes. The slopes of the TT simulations remain constant and the growth in bond dimension χ\chi for increasing system size is only reflected in a yy-axis offset. On the other hand, the computational complexity of the classic FD scales exponentially with the number of tensors (TT-cores) nn. Despite the different LSE solving strategies applied to TT and FD, such as the non-variational approach or the absence of preconditioning in the FD, the extracted conclusions from Figure 17 remain valid. An improvement in the FD computation times would shift the break-even point along the xx-axis toward higher tensor counts. Given that industrial CFD cases routinely exceed 20 million NVPS (≈24\approx24 tensors), we anticipate that tensor-network methods like the one proposed here could deliver substantial computational gains.

Figure 17. Scaling of the average wall-clock time per iteration (sampled every 200th iteration) for the Re=150Re = 150 simulations discussed in Subsection IV A 2.

Conclusions & Outlook

In this work, we demonstrated that the TT format can serve as an efficient ROM framework for simulating INSE in curvilinear coordinates. To the best of our knowledge, this study presents the first application of tensor network methods to the encoding and simulation of fluid fields on non-uniform, body-fitted meshes, extending their use beyond the uniform discretizations commonly considered in previous works [58, 60, 75]. This development marks a significant advancement toward the realization of an industrial-grade Quantum Computational Fluid Dynamics toolbox, as body-fitted meshes are integral to many established CFD programs [97, 106, 107].

The methodology presented here ranges from outlining the mesh generation process, formulated within the TT format, to describing the curvilinear transformation of differential operators using known TT operations, and includes the efficient readout of relevant quantities of interest, such as the lift coefficient and the Strouhal number, directly in the TT format [60]. Furthermore, the approach is reinforced through the ghost point strategy for boundary treatment that eliminates the need for penalty terms in the system of equations. This enables the versatile application of common engineering boundaries within the TT format. Similarly, the fractional-step method avoids determining additional degrees of freedom required to enforce the pressure-velocity coupling as realized in Refs. [57, 60, 108].

This allows the implementation of various flow scenarios around an immersed cylinder, ranging from Stokes flows to vortex streets with instabilities at large Reynolds numbers, and even the Magnus effect induced by cylinder rotation. In all cases, the TT-based solver demonstrates outstanding agreement with a significant reduction up to 20-fold of the NVPS required to represent fields and up to 1000-fold for the operators.

This work lays the foundation for future investigations into how obstacle geometry, mesh generation, upwind-biased schemes [46, 62, 109], parallelization [110], and grid properties (e.g., orthogonality) influence the achievable compression of operator representations and further increase the TT’s efficiency. In particular, certain mesh types–such as elliptic grids–may naturally yield low-rank, geometry-aware operators, offering potential for shape optimization and other industrial applications. A natural next step is to investigate the laminar and turbulent flow around an immersed ellipsoid, which breaks the rotational symmetry seen in the cylinder, as well as the flow around NACA airfoils [111], providing a compelling test case. Particular focus will be placed on high-Reynolds number regimes, where the curvilinear approach is expected to be advantageous compared to IBM.

Acknowledgments

This publication and the current work have received funding from the European Union’s Horizon Europe research and innovation program (HORIZON-CL4-2021-DIGITAL-EMERGING-02-10) under grant agreement No. 101080085 QCFD. PS acknowledges financial support by the DLR-Quantum-Fellowship Program. DJ and TH are partly funded by the Cluster of Excellence “Advanced Imaging of Matter” of the Deutsche Forschungsgemeinschaft (DFG)—EXC 2056– project ID390715994. DJ acknowledges the support by DFG project “Quantencomputing mit neutralen Atomen” (JA 1793/1-1, Japan-JST-DFG-ASPIRE 2024) and the Hamburg Quantum Computing Initiative (HQIC) project EFRE. The EFRE project is co-financed by ERDF of the European Union and by “Fonds of the Hamburg Ministry of Science, Research, Equalities and Districts (BWFGB)”. The numerical data is acquired using the computational cluster Hummel-2 of the University of Hamburg, which is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 498394658. This work was inspired in part by the engaging discussions and expert input from Airbus, BMW Group, and AWS during the Quantum Mobility Quest challenge. The authors thank Philipp Thoma for fruitful discussions.

Author Declarations

Declaration of Interest

The authors have no interest to disclose.

Author Contributions

Nis-Luca van Hülst: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review and editing; Pia Siegl: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing – original draft, Writing – review and editing; Paul Over: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review and editing; Sergio Bengoechea: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review and editing; Tomohiro Hashizume: Conceptualization, Investigation, Methodology, Supervision, Visualization, Writing – original draft, Writing – review and editing; Mario Guillaume Cecile: Conceptualization, Investigation, Methodology, Supervision, Visualization, Writing – original draft, Writing – review and editing; Thomas Rung: Project administration, Funding acquisition, Supervision, Conceptualization, Methodology, Resources, Writing - review and editing; Dieter Jaksch: Project administration, Funding acquisition, Supervision, Conceptualization, Methodology, Resources, Writing - review and editing.

Data availability

A software example for simulating the 2D heat equation on the employed cylindrical grid is available via Ref. [112]. The data that support the findings of this study are available upon reasonable request from the authors and at Ref. https://doi.org/10.25592/uhhfdm.17687.

References

Discretization and Coordinate Transformation

In Figure 18 the two domains of interest are illustrated, i.e., the computational (a) and the physical (b) domain. The former corresponds to a uniformly discretized unit square Ω0=[0,1]×[0,1)\Omega_0 = [0,1] \times[0,1), while the latter is discretized using an O-type grid. The coloring highlights the corresponding boundaries of the two domains; for instance, the outer boundary ∂Ω\partial\Omega of the physical domain (shown in black) maps to the top boundary of the unit square (also black).

Figure 18. Computational domain Ω0\Omega_0 with uniform discretization (left) and physical domain Ω\Omega around a cylinder discretized using an O-type grid (right). In both, light-gray points indicate grid nodes, and colored lines highlight corresponding boundaries.

The computational domain Ω0\Omega_0 is discretized using (2nη+2)×2nξ(2^{n_\eta} + 2) \times2^{n_\xi} grid points in the η(i)\eta(i)- and ξ(j)\xi(j)-directions, respectively, given by

ηij=(i+1)δηξij=jδξi=−1,…,2nη,j=0,…,2nξ−1.(37)\begin{aligned} \eta_{ij} &= (i+1)\delta\eta\\ \xi_{ij} &= j\delta\xi \end{aligned} \qquad \begin{aligned} i &= -1,\ldots,2^{n_\eta},\\ j &= 0,\ldots,2^{n_\xi}-1. \tag*{(37)} \end{aligned}

Here, i=−1i=-1 and i=2nηi=2^{n_\eta} correspond to ghost points representing the top (black) and bottom (magenta) boundaries, respectively. These grid points are not part of the solution domain but are incorporated through the ghost-point approach. The associated grid spacings are δη=1/(2nη+1)\delta\eta= 1/(2^{n_\eta}+1) and δξ=1/2nξ\delta\xi= 1/2^{n_\xi}. The difference in spacing reflects the boundary conditions: periodicity in the ξ\xi-direction versus non-periodicity in η\eta.

The Cartesian spatial derivatives required by the momentum (10) and continuity (11) equations (INSE) are expressed in the reference coordinate system (ξ,η)(\xi,\eta) by applying the chain rule, e.g., ∂x=∂xξ∂ξ+∂xη∂η\partial_x = \partial_x^\xi\partial_\xi+ \partial_x^\eta\partial_\eta, and using a coordinate transformation, i.e., ∂xξ=∂ηz/J\partial_x^\xi= \partial_\eta z/J, ∂xη=−∂ξz/J\partial_x^\eta= -\partial_\xi z/J, ∂yξ=−∂ηx/J\partial_y^\xi= -\partial_\eta x/J and ∂yη=∂ξx/J\partial_y^\eta= \partial_\xi x/J, viz.

(∂x,∂z)=1J(∂ηz−∂ξz−∂ηx∂ξx)(∂ξ,∂η)T,(38)(\partial_x,\partial_z)=\frac{1}{J} \begin{pmatrix} \partial_\eta z & -\partial_\xi z\\ -\partial_\eta x & \partial_\xi x \end{pmatrix} (\partial_\xi,\partial_\eta)^{\mathsf T}, \tag*{(38)}

for first spatial derivatives and

Δ=1J(∂ξ,∂η)1J(g22−g12−g21g11)(∂ξ,∂η)T,(39)\Delta=\frac{1}{J}(\partial_\xi,\partial_\eta)\frac{1}{J} \begin{pmatrix} g_{22} & -g_{12}\\ -g_{21} & g_{11} \end{pmatrix} (\partial_\xi,\partial_\eta)^{\mathsf T}, \tag*{(39)}

for the Laplace operator. Here, JJ denotes the Jacobian determinant and gg the metric tensor. They are given by the following expressions

J=(∂ξx)(∂ηz)−(∂ηx)(∂ξz)=det⁡(∂ξx∂ηx∂ξz∂ηz),J=(\partial_\xi x)(\partial_\eta z)-(\partial_\eta x)(\partial_\xi z)=\det \begin{pmatrix} \partial_\xi x & \partial_\eta x\\ \partial_\xi z & \partial_\eta z \end{pmatrix},
g11=(∂ξx)2+(∂ξz)2,g22=(∂ηx)2+(∂ηz)2,g12=(∂ξx)(∂ηx)+(∂ξz)(∂ηz).(40)g_{11}=(\partial_\xi x)^2+(\partial_\xi z)^2,\qquad g_{22}=(\partial_\eta x)^2+(\partial_\eta z)^2,\qquad g_{12}=(\partial_\xi x)(\partial_\eta x)+(\partial_\xi z)(\partial_\eta z). \tag*{(40)}

Writing out the Laplacian operator one obtains a sum of seven operator terms

Δ=1J{∂ξ(g22J)∂ξ+g22J∂ξξ−∂ξ(g12J)∂η−∂η(g12J)∂ξ−2g12J∂ξη+∂η(g11J)∂η+g11J∂ηη}.(41)\Delta=\frac{1}{J}\left\{ \partial_\xi\left(\frac{g_{22}}{J}\right)\partial_\xi +\frac{g_{22}}{J}\partial_{\xi\xi} -\partial_\xi\left(\frac{g_{12}}{J}\right)\partial_\eta -\partial_\eta\left(\frac{g_{12}}{J}\right)\partial_\xi -2\frac{g_{12}}{J}\partial_{\xi\eta} +\partial_\eta\left(\frac{g_{11}}{J}\right)\partial_\eta +\frac{g_{11}}{J}\partial_{\eta\eta} \right\}. \tag*{(41)}

The derivatives w.r.t. ξ\xi and η\eta are approximated with second-order central difference methods. The first derivatives of a generic scalar field ϕ\phi read

∂ξϕ∣i,j=ϕi,j+1−ϕi,j−12δξ+O(δξ2),∂ηϕ∣i,j=ϕi+1,j−ϕi−1,j2δη+O(δη2).(42)\left.\partial_{\xi}\phi\right|_{i,j} = \frac{\phi_{i,j+1}-\phi_{i,j-1}}{2\delta_{\xi}} + O(\delta_{\xi}^{2}), \qquad\left.\partial_{\eta}\phi\right|_{i,j} = \frac{\phi_{i+1,j}-\phi_{i-1,j}}{2\delta_{\eta}} + O(\delta_{\eta}^{2}). \tag*{(42)}

Similarly, the corresponding second derivatives read

∂ξξϕ∣i,j=ϕi,j+1−2ϕi,j+ϕi,j−1δξ2+O(δξ2),∂ηηϕ∣i,j=ϕi+1,j−2ϕi,j+ϕi−1,jδη2+O(δη2).(43)\left.\partial_{\xi\xi}\phi\right|_{i,j} = \frac{\phi_{i,j+1}-2\phi_{i,j}+\phi_{i,j-1}}{\delta_{\xi}^{2}} + O(\delta_{\xi}^{2}), \qquad\left.\partial_{\eta\eta}\phi\right|_{i,j} = \frac{\phi_{i+1,j}-2\phi_{i,j}+\phi_{i-1,j}}{\delta_{\eta}^{2}} + O(\delta_{\eta}^{2}). \tag*{(43)}

and finally mixed derivatives are approximated by

∂ξηϕ∣i,j=ϕi+1,j+1−ϕi+1,j−1−ϕi−1,j+1+ϕi−1,j−14δξδη+O(δξ2+δη2).(44)\left.\partial_{\xi\eta}\phi\right|_{i,j} = \frac{\phi_{i+1,j+1}-\phi_{i+1,j-1}-\phi_{i-1,j+1}+\phi_{i-1,j-1}}{4\delta_{\xi}\delta_{\eta}} + O(\delta_{\xi}^{2}+\delta_{\eta}^{2}). \tag*{(44)}

Sorting Vector to Matrix and Reverse

The column-major ordered vectorization of matrix A∈R3×3A \in\mathbb{R}^{3\times3}, denoted vec⁡(A)\operatorname{vec}(A), is

A=[a11a12a13a21a22a23a31a32a33] → vec⁡(A)=[a11a21a31a12a22a32a13a23a33]T.(45)A = \begin{bmatrix} a_{11} & a_{12} & a_{13} \\ a_{21} & a_{22} & a_{23} \\ a_{31} & a_{32} & a_{33} \end{bmatrix} \ \to\ \operatorname{vec}(A) = \begin{bmatrix} a_{11} & a_{21} & a_{31} & a_{12} & a_{22} & a_{32} & a_{13} & a_{23} & a_{33} \end{bmatrix}^{\mathsf{T}}. \tag*{(45)}

Read-out of Column or Row

Assuming the two-dimensional scalar field fij:=f(ξij,ηij)∈R2nη×2nξf_{ij} := f(\xi_{ij},\eta_{ij}) \in\mathbb{R}^{2^{n_{\eta}}\times2^{n_{\xi}}} is represented as a TT, we write

fij≃A[1]i1A[2]i2⋯A[nη]inηA[nη+1]j1⋯A[n]jnξ,(46)f_{ij} \simeq\mathbf{A}^{[1]i_1}\mathbf{A}^{[2]i_2}\cdots\mathbf{A}^{[n_{\eta}]i_{n_{\eta}}}\mathbf{A}^{[n_{\eta}+1]j_1}\cdots\mathbf{A}^{[n]j_{n_{\xi}}}, \tag*{(46)}

where total number of TT-cores is n=nη+nξn=n_{\eta}+n_{\xi}. Each core A[ℓ]∈Rαℓ−1×2×αℓ\mathbf{A}^{[\ell]} \in\mathbb{R}^{\alpha_{\ell-1}\times2\times\alpha_{\ell}} is equipped with a corresponding binary digit ik,jk∈{0,1}i_k,j_k \in\{0,1\} of the row and column indices, respectively. To extract the II-th row, i.e., the slice fIjf_{Ij}, we fix the first nηn_{\eta} physical indices using the binary representation (I1,…,Inη)∈{0,1}nη(I_1,\ldots,I_{n_{\eta}})\in\{0,1\}^{n_{\eta}} of II, resulting in the contracted form

fIj≃A[1]i1=I1A[2]i2=I2⋯A[nη]inη=InηA[nη+1]j1⋯A[n]jnξ.(47)f_{Ij} \simeq\mathbf{A}^{[1]i_1=I_1}\mathbf{A}^{[2]i_2=I_2}\cdots\mathbf{A}^{[n_{\eta}]i_{n_{\eta}}=I_{n_{\eta}}}\mathbf{A}^{[n_{\eta}+1]j_1}\cdots\mathbf{A}^{[n]j_{n_{\xi}}}. \tag*{(47)}

This procedure effectively contracts the first nηn_{\eta} tensor cores to form a matrix MI∈R1×αnηM_I \in\mathbb{R}^{1\times\alpha_{n_{\eta}}}, which is then left-multiplied into the remaining TT-cores

fIj≃MI⋅(A[nη+1]j1⋯A[n]jnξ).(48)f_{Ij} \simeq M_I\cdot\left(\mathbf{A}^{[n_{\eta}+1]j_1}\cdots\mathbf{A}^{[n]j_{n_{\xi}}}\right). \tag*{(48)}

This yields a TT representation of the row fIj∈R2nξf_{Ij}\in\mathbb{R}^{2^{n_{\xi}}}. Similarly, to extract the JJ-th column, i.e., the slice fiJf_{iJ}, we fix the last nξn_{\xi} physical indices using the binary representation (J1,…,Jnξ)∈{0,1}nξ(J_1,\ldots,J_{n_{\xi}})\in\{0,1\}^{n_{\xi}} of JJ, resulting in the contracted form

fiJ≃A[1]i1A[2]i2⋯A[nη]inηA[nη+1]j1=J1⋯A[n]jnξ=Jnξ.(49)f_{iJ} \simeq\mathbf{A}^{[1]i_1}\mathbf{A}^{[2]i_2}\cdots\mathbf{A}^{[n_{\eta}]i_{n_{\eta}}}\mathbf{A}^{[n_{\eta}+1]j_1=J_1}\cdots\mathbf{A}^{[n]j_{n_{\xi}}=J_{n_{\xi}}}. \tag*{(49)}

This procedure effectively contracts the last nξn_{\xi} tensor cores to form a matrix MJ∈Rαnη×1M_J \in\mathbb{R}^{\alpha_{n_{\eta}}\times1}, which is then right-multiplied into the remaining TT-cores:

fiJ≃(A[1]i1A[2]i2⋯A[nη]inη)⋅MJ.(50)f_{iJ} \simeq\left(\mathbf{A}^{[1]i_1}\mathbf{A}^{[2]i_2}\cdots\mathbf{A}^{[n_{\eta}]i_{n_{\eta}}}\right)\cdot M_J. \tag*{(50)}

This yields a TT representation of the column fiJ∈R2nηf_{iJ}\in\mathbb{R}^{2^{n_{\eta}}}.

Vector to Tensor Train Decomposition

The decomposition of a vector into a TT is exemplified here analytically for a vector of size 23=82^3 = 8 (n=3n = 3). As a preliminary requirement, mention the rank core product, denoted by ⋈\bowtie, which acts as an array multiplication given vector entities as

[(ab)(cd)]⋈[αβγδ]=(ab)⊗(αβ)+(cd)⊗(γδ)=(aα+cγaβ+cδbα+dγbβ+dδ)(51)\left[\begin{pmatrix}a\\b\end{pmatrix}\begin{pmatrix}c\\d\end{pmatrix}\right]\bowtie\begin{bmatrix}\alpha\\\beta\\\gamma\\\delta\end{bmatrix}=\begin{pmatrix}a\\b\end{pmatrix}\otimes\begin{pmatrix}\alpha\\\beta\end{pmatrix}+\begin{pmatrix}c\\d\end{pmatrix}\otimes\begin{pmatrix}\gamma\\\delta\end{pmatrix}=\begin{pmatrix}a\alpha+c\gamma\\a\beta+c\delta\\b\alpha+d\gamma\\b\beta+d\delta\end{pmatrix} \tag*{(51)}
[(abcd)(efgh)]⋈[αβγδ]=(abcd)⊗(αβ)+(efgh)⊗(γδ)=(aα+eγaβ+eδbα+fγbβ+fδcα+gγcβ+gδdα+hγdβ+hδ)(52)\left[\begin{pmatrix}a\\b\\c\\d\end{pmatrix}\begin{pmatrix}e\\f\\g\\h\end{pmatrix}\right]\bowtie\begin{bmatrix}\alpha\\\beta\\\gamma\\\delta\end{bmatrix}=\begin{pmatrix}a\\b\\c\\d\end{pmatrix}\otimes\begin{pmatrix}\alpha\\\beta\end{pmatrix}+\begin{pmatrix}e\\f\\g\\h\end{pmatrix}\otimes\begin{pmatrix}\gamma\\\delta\end{pmatrix}=\begin{pmatrix}a\alpha+e\gamma\\a\beta+e\delta\\b\alpha+f\gamma\\b\beta+f\delta\\c\alpha+g\gamma\\c\beta+g\delta\\d\alpha+h\gamma\\d\beta+h\delta\end{pmatrix} \tag*{(52)}

If one now defines a vector t=(1,2,3,4,5,6,7,8)T\boldsymbol{t}=(1,2,3,4,5,6,7,8)^{\mathrm{T}}, its tensor decomposition is

[(11)(01)]⋈[(1100)(0122)]⋈(1222)=(12345678)(53)\left[\begin{pmatrix}1\\1\end{pmatrix}\begin{pmatrix}0\\1\end{pmatrix}\right]\bowtie\left[\begin{pmatrix}1\\1\\0\\0\end{pmatrix}\begin{pmatrix}0\\1\\2\\2\end{pmatrix}\right]\bowtie\begin{pmatrix}1\\2\\2\\2\end{pmatrix}=\begin{pmatrix}1\\2\\3\\4\\5\\6\\7\\8\end{pmatrix} \tag*{(53)}

Each of the square brackets represents one tensor core A[k]\mathbf{A}[k], where k∈{0,1,2}k \in\{0,1,2\}. Furthermore, the bond dimension and the physical dimension are 2 in all cases.

Grid points in Application

The classically generated grid consists of N=Nη×Nξ:=2nη×2nξN=N_{\eta}\times N_{\xi}:=2^{n_{\eta}}\times2^{n_{\xi}} interior support points, with indices i=0,…,Nη−1i=0,\ldots,N_{\eta}-1 and j=0,…,Nξ−1j=0,\ldots,N_{\xi}-1, provided as a tensor (outer) product of 1D functions as

xij=D2(D∂ΩD)i+1Nη+1cos⁡(2πjδξ),zij=D2(D∂ΩD)i+1Nη+1sin⁡(2πjδξ).(54)x_{ij}=\frac{D}{2}\left(\frac{D_{\partial\Omega}}{D}\right)^{\frac{i+1}{N_{\eta}+1}}\cos(2\pi j\delta\xi),\qquad z_{ij}=\frac{D}{2}\left(\frac{D_{\partial\Omega}}{D}\right)^{\frac{i+1}{N_{\eta}+1}}\sin(2\pi j\delta\xi). \tag*{(54)}

Force Computation

The Strouhal number can be computed using different methods, depending on the available data and flow characteristics. In this study, we determine the Strouhal number from the oscillations in the dimensionless lift force FLF_L, following the approach of [97]. The lift force corresponds to the zz-projection of the total force as

FL=ezT∫Γτ(v,p)n ds≈2πδξ12∑j=0Nξ−1(cos⁡2πjδξsin⁡−2πjδξ)T(τ21τ22)(55)F_L=\mathbf{e}_z^{\mathrm{T}}\int_{\Gamma}\boldsymbol{\tau}(\mathbf{v},p)\mathbf{n}\,ds\approx2\pi\delta\xi\frac{1}{2}\sum_{j=0}^{N_{\xi}-1}\begin{pmatrix}\cos2\pi j\delta\xi\\\sin-2\pi j\delta\xi\end{pmatrix}^{\mathrm{T}}\begin{pmatrix}\tau_{21}\\\tau_{22}\end{pmatrix} \tag*{(55)}

Here, the stress tensor components need to be determined on the cylinder surface via

τ21=2Re(12(∂v∂x∣Γ+∂u∂z∣Γ)),τ22=−(p∣Γ+23Re(∇v)∣Γ)+2Re∂v∂z∣Γ(56)\tau_{21}=\frac{2}{Re}\left(\frac{1}{2}\left(\left.\frac{\partial v}{\partial x}\right|_{\Gamma}+\left.\frac{\partial u}{\partial z}\right|_{\Gamma}\right)\right),\qquad\tau_{22}=-\left(\left.p\right|_{\Gamma}+\frac{2}{3Re}\left.(\nabla\mathbf{v})\right|_{\Gamma}\right)+\left.\frac{2}{Re}\frac{\partial v}{\partial z}\right|_{\Gamma} \tag*{(56)}

where derivatives are computed in the computational domain Ω0\Omega_0 and then transformed. A similar explicit formula can be derived for the dimensionless drag force FDF_D.

Algorithm Modifications for Transient

First, we implement an initial modulation onto the velocity field to allow the vortex shedding to emerge. To do so, we employed the following modulation

uij(t=0)=A0e(iδη)6−1e−1sin⁡(2πjδξ),(57)u_{ij}(t=0)=A_0\frac{e^{(i\delta\eta)^6}-1}{e-1}\sin(2\pi j\delta\xi), \tag*{(57)}

with perturbation strength A0=0.05A_0=0.05. This modulation introduces an asymmetry w.r.t. the branch cut, while it vanishes radially towards the far-field boundary ∂Ω\partial\Omega. This was motivated to avoid strong perturbations on the comparably large cells in the far field, potentially making the system stiffer. Second, we employ a preconditioner for solving the Poisson equation within the TT approach to enable faster convergence in terms of the number of sweeps in the variational scheme [47], and to avoid dealing with highly ill-conditioned LSE in the local variational updates. This allows us to make use of the fast matvec routine outlined in [54], i.e., emulating the local matrix-vector multiplication by successive tensor contractions scaling with O(χ3)O(\chi^3) instead of assembling the local matrix and naively applying it to a vector with scaling of O(χ4)O(\chi^4).

Verification of the Finite Difference Reference

To validate the feasibility of the presented results, the in-house FD code is benchmarked against reference values from OpenFOAM [113] and the literature [97, 114, 115] for the case of flow around a cylinder. For a quantitative comparison, the drag coefficient CDC_D as a function of the Reynolds number over the range Re∈[1,200]Re\in[1,200] is given in Figure 19.

Figure 19. Drag coefficient as a function of the Reynolds number ReRe for the flow around a cylinder from Stokes to subcritical regimes. The benchmark study compares results from literature [97, 114, 115], OpenFOAM [113] and the developed FD solver.

The verification uses ◯\bigcirc marks for literature references, □\square marks for the numerical results from OpenFOAM [113], and △\triangle for results of the FD solver under validation. Overall, the results confirm good agreement, though minor discrepancies are apparent, especially at lower ReRe. These deviations can be attributed mainly to two factors: 1) discretization differences between the numerical models (finite approximation, grid resolution) and 2) differences in the computational setup (boundary layers, pressure correction approach, boundary conditions).

  1. The application of different meshes between the models in Figure 19 is a source of discrepancy in the results. Furthermore, the references either approximate the vorticity-streamfunction formulation in [114, 115], or apply finite

volume discretizations [97, 113], which inherently conserve fluxes, whereas the in-house FD solver uses central stencils introducing dispersion errors.

  1. Another source of error lies in the computational model itself. As conventional CFD solvers can handle more complex flow phenomena, different types of boundaries and pressure correction approaches are implemented. In contrast, the developed FD code relies on Dirichlet-Neumann pressure boundaries to guarantee numerical stability. The differences in the boundary conditions are important as the wall treatment directly influences the computation of forces and reflects on the drag coefficient CDC_{\mathrm{D}}. As a consequence, the corresponding literature symbols do not correctly align with the numerical results that include inertial contributions.

In conclusion, the FD solver delivers good accuracies and shows a constant offset to the given references, indicating a well-functioning of the model and motivating a reliable basis for benchmarking the TT solver under discussion.

References

  1. [1]R. D. Moser, Chapter 1 - numerical challenges in turbulence simulation, in Numerical Methods in Turbulence Simulation, Numerical Methods in Turbulence, edited by R. D. Moser (Academic Press, 2023) pp. 1–43.
  2. [2]J. P. Slotnick, A. Khodadoust, J. Alonso, D. Darmofal, W. Gropp, E. Lurie, and D. J. Mavriplis, CFD vision 2030 study: a path to revolutionary computational aerosciences, Tech. Rep. (NASA, 2014).
  3. [3]A. Monin, A. Yaglom, and J. Lumley, Statistical Fluid Mechanics: Mechanics of Turbulence, Dover books on physics No. Vol. 1 (Dover Publications, 2007).
  4. [4]D. Wilcox, Turbulence Modeling for CFD (Third Edition) (Hardcover) (2006).
  5. [5]B. Launder and D. Spalding, The numerical computation of turbulent flows, in Numerical Prediction of Flow, Heat Transfer, Turbulence and Combustion, edited by S. V. Patankar, A. Pollard, A. K. Singhal, and S. P. Vanka (Pergamon, 1983) pp. 96–116.
  6. [6]S. B. Pope, Turbulent Flows (Cambridge University Press, 2000).DOI
  7. [7]N. Aubry, On the hidden beauty of the proper orthogonal decomposition, Theoretical and Computational Fluid Dynamics 2, 339 (1991).
  8. [8]L. Sirovich, Turbulence and the dynamics of coherent structures part i: Coherent structures, Quarterly of Applied Mathematics 45, 561 (1987).
  9. [9]K. Taira, S. L. Brunton, S. T. M. Dawson, C. W. Rowley, T. Colonius, B. J. McKeon, O. T. Schmidt, S. Gordeyev, V. Theofilis, and L. S. Ukeiley, Model analysis of fluid flows: An overview, AIAA Journal 55, 4013 (2017).
  10. [10]P. J. Schmid, Dynamic mode decomposition of numerical and experimental data, Journal of Fluid Mechanics 656, 5–28 (2010).
  11. [11]J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems (Society for Industrial and Applied Mathematics, 2016).
  12. [12]S. L. Brunton, B. R. Noack, and P. Koumoutsakos, Machine learning for fluid mechanics, Annual Review of Fluid Mechanics 52, 477 (2020).
  13. [13]S. L. Brunton and J. N. Kutz, Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control (Cambridge University Press, 2022).
  14. [14]P. Pfeffer, F. Heyder, and J. Schumacher, Reduced-order modeling of two-dimensional turbulent rayleigh-bénard flow by hybrid quantum-classical reservoir computing, Phys. Rev. Res. 5, 043242 (2023).DOI
  15. [15]D. Ramezanian, A. G. Nouri, and H. Babaee, On-the-fly reduced order modeling of passive and reactive species via time-dependent manifolds, Computer Methods in Applied Mechanics and Engineering 382, 113882 (2021).
  16. [16]R. Bellman and S. Dreyfus, Dynamic Programming, Vol. 33 (Princeton University Press, 2010).
  17. [17]G. Berkooz, P. Holmes, and J. L. Lumley, The proper orthogonal decomposition in the analysis of turbulent flows, Annual Review of Fluid Mechanics 25, 539 (1993).
  18. [18]P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley, Turbulence, Coherent Structures, Dynamical Systems and Symmetry, 2nd ed., Cambridge Monographs on Mechanics (Cambridge University Press, 2012).
  19. [19]P. J. Holmes, J. L. Lumley, G. Berkooz, J. C. Mattingly, and R. W. Wittenberg, Low-dimensional models of coherent structures in turbulence, Physics Reports 287, 337 (1997).DOI
  20. [20]K. Kunisch and S. Volkwein, Control of the Burgers equation by a reduced-order approach using proper orthogonal decomposition, Journal of optimization theory and applications 102, 345 (1999).DOI
  21. [21]C. W. Rowley, Model reduction for fluids, using balanced proper orthogonal decomposition, International Journal of Bifurcation and Chaos 15, 997 (2005).
  22. [22]D. Venturi, On proper orthogonal decomposition of randomly perturbed fields with applications to flow past a cylinder and natural convection over a horizontal plate, Journal of Fluid Mechanics 559, 215–254 (2006).DOI
  23. [23]P. Givi, A. J. Daley, D. Mavriplis, and M. Malik, Quantum speedup for aerospace and engineering, AIAA Journal 58, 3715 (2020).DOI
  24. [24]D. Jaksch, P. Givi, A. J. Daley, and T. Rung, Variational quantum algorithms for computational fluid dynamics, AIAA Journal 61, 1885 (2023).
  25. [25]S. S. Bharadwaj and K. R. Sreenivasan, Quantum computation of fluid dynamics, in Indian Academy of Sciences Conference Series, Vol. 3 (2020) pp. 77–96.DOI
  26. [26]C. A. Riofrío, J. Klepsch, J. R. Fînžgar, F. Kiwit, L. Hölscher, M. Erdmann, L. Müller, C. Kumar, Y. A. Berrada, and A. Luckow, Quantum computing for automotive applications (2024), arXiv:2409.14183 [quant-ph].arxiv.org/abs/2409.14183
  27. [27]C. Sanavio, R. Scatamacchia, C. de Falco, and S. Succi, Three carleman routes to the quantum simulation of classical fluids, Physics of Fluids 36, 057143 (2024).DOI
  28. [28]C. Sanavio, E. Mauri, and S. Succi, Explicit quantum circuit for simulating the advection–diffusion–reaction dynamics, IEEE Transactions on Quantum Engineering 6, 1 (2025).DOI
  29. [29]A. W. Harrow, A. Hassidim, and S. Lloyd, Quantum algorithm for linear systems of equations, Phys. Rev. Lett. 103, 150502 (2009).pubmed.ncbi.nlm.nih.gov/19905613
  30. [30]A. M. Childs, R. Kothari, and R. D. Somma, Quantum algorithm for systems of linear equations with exponentially improved dependence on precision, SIAM J. Comput. 46, 1920 (2017).DOI
  31. [31]L. Budinski, Quantum algorithm for the advection-diffusion equation simulated with the lattice boltzmann method, Quantum Information Processing 20, 57 (2021).DOI
  32. [32]B. Ljubomir, Quantum algorithm for the navier-stokes equations by using the streamfunction-vorticity formulation and the lattice boltzmann method, International Journal of Quantum Information 20, 2150039 (2022).DOI
  33. [33]W. Itani, K. R. Sreenivasan, and S. Succi, Quantum algorithm for lattice boltzmann (QALB) simulation of incompressible fluids with a nonlinear collision term, Physics of Fluids 36, 017112 (2024).arxiv.org/abs/2304.05915
  34. [34]D. Wawrzyniak, J. Winter, S. Schmidt, T. Indinger, C. F. Janßen, U. Schramm, and N. A. Adams, A quantum algorithm for the lattice-boltzmann method advection-diffusion equation, Computer Physics Communications 306, 109733 (2025).DOI
  35. [35]E. Madelung, Quantentheorie in hydrodynamischer form, Zeitschrift für Physik 40, 322 (1927).DOI
  36. [36]Z. Meng and Y. Yang, Quantum computing of fluid dynamics using the hydrodynamic Schrödinger equation, Phys. Rev. Res. 5, 033182 (2023).DOI
  37. [37]Z. Meng and Y. Yang, Quantum spin representation for the Navier-Stokes equation, Phys. Rev. Res. 6, 043130 (2024).DOI
  38. [38]L. Salasnich, S. Succi, and A. Tiribocchi, Quantum wave representation of dissipative fluids, International Journal of Modern Physics C 35, 2450100 (2024).DOI
  39. [39]P. Brearley and S. Laizet, Quantum algorithm for solving the advection equation using hamiltonian simulation, Phys. Rev. A 110, 012430 (2024).DOI
  40. [40]P. Over, S. Bengoechea, P. Brearley, S. Laizet, and T. Rung, Quantum algorithm for the advection-diffusion equation by direct block encoding of the time-marching operator, Phys. Rev. A 112, L010401 (2025).DOI
  41. [41]A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A variational eigenvalue solver on a photonic quantum processor, Nature Communications 5, 10.1038/ncomms5213 (2014).DOI
  42. [42]M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, H. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, Variational quantum algorithms, Nature Reviews Physics 3, 625 (2021).DOI
  43. [43]S. S. Bharadwaj and K. R. Sreenivasan, Hybrid quantum algorithms for flow problems, Proceedings of the National Academy of Sciences 120, e2311014120 (2023).DOI
  44. [44]M. Lubasch, J. Joo, P. Moinier, M. Kiffner, and D. Jaksch, Variational quantum algorithms for nonlinear problems, Phys. Rev. A 101, 010301 (2020).DOI
  45. [45]P. Over, S. Bengoechea, T. Clerici, L. Scandurra, E. de Villiers, and D. Jaksch, Boundary treatment for variational quantum simulations of partial differential equations on quantum computers, Computers & Fluids 288, 106508 (2025).
  46. [46]S. Bengoechea, P. Over, D. Jaksch, and T. Rung, Towards variational quantum algorithms for generalized linear and nonlinear transport phenomena (2024), arXiv:2411.14931 [quant-ph].DOI
  47. [47]U. Schollwöck, The density-matrix renormalization group in the age of matrix product states, Annals of Physics 326, 96 (2011).DOI
  48. [48]A. Termanova, A. Melnikov, E. Mamenchikov, N. Belokonov, S. Dolgov, A. Berezutskii, R. Ellerbrock, C. Mansell, and M. R. Perelshtein, Tensor quantum programming, New Journal of Physics 26, 1263019 (2024).arxiv.org/abs/2403.13486
  49. [49]P. Siegl, G. S. Reese, T. Hashizume, N.-L. van Hülst, and D. Jaksch, Tensor-programmable quantum circuits for solving differential equations (2025), arXiv:2502.04425 [quant-ph].DOI
  50. [50]M. D. García and A. Márquez Romero, Survey on computational applications of tensor-network simulations, IEEE Access 12, 193212 (2024).
  51. [51]G. Rozza, G. Stabile, and F. Ballarin, Advanced Reduced Order Methods and Applications in Computational Fluid Dynamics, edited by G. Rozza, G. Stabile, and F. Ballarin (Society for Industrial and Applied Mathematics, Philadelphia, PA, 2022).DOI
  52. [52]J. Eisert, M. Cramer, and M. B. Plenio, Colloquium: Area laws for the entanglement entropy, Rev. Mod. Phys. 82, 277 (2010).DOI
  53. [53]B. N. Khomkskii, O(dlog n)-quantics approximation of n-d tensors in high-dimensional numerical modeling, Constructive Approximation 34, 257 (2011).
  54. [54]S. V. Dolgov, B. N. Khomkskii, and I. V. Oseledets, Fast solution of parabolic problems in the tensor train/quantized tensor train format with initial application to the Fokker–Planck equation, SIAM Journal on Scientific Computing 34, A3016 (2012).DOI
  55. [55]L. Richter, L. Sallandt, and N. Nüskен, Solving high-dimensional parabolic PDEs using the tensor train format, in Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 139, edited by M. Meila and T. Zhang (PMLR, 2021) pp. 8998–9009.arxiv.org/abs/2102.11830
  56. [56]M. Lubasch, P. Moinier, and D. Jaksch, Multigrid renormalization, Journal of Computational Physics 372, 587 (2018).
  57. [57]L. Hölscher, P. Rao, L. Müller, J. Klepsch, A. Luckow, T. Stollenwerk, and F. K. Wilhelm, Quantum-inspired fluid simulation of two-dimensional turbulence with GPU acceleration, Phys. Rev. Res. 7, 013112 (2025).DOI
  58. [58]M. Kiffner and D. Jaksch, Tensor network reduced order models for wall-bounded flows, Physical Review Fluids 8, 124101 (2023).DOI
  59. [59]E. Ye and N. Loureiro, Quantized tensor networks for solving the Vlasov-Maxwell equations (2024), arXiv:2311.07756 [physics.comp-ph].DOI
  60. [60]R. D. Peddinti, S. Pisoni, A. Marini, P. Lott, H. Argentieri, E. Tiunov, and L. Aolita, Quantum-inspired framework for computational fluid dynamics, Communications Physics 7, 135 (2024).DOI
  61. [61]A. A. Michailidis, C. Fenton, and M. Kiffner, Tensor train multiplication (2024), arXiv:2410.19747 [physics.comp-ph].arxiv.org/abs/2410.19747
  62. [62]M. E. Danis, D. Truong, I. Boureima, O. Korobkin, K. Ø. Rasmussen, and B. S. Alexandrov, Tensor-train WENO scheme for compressible flows, Journal of Computational Physics 529, 113891 (2025).arxiv.org/abs/2405.12301
  63. [63]S. Pisoni, R. D. Peddinti, E. Tiunov, S. E. Guzman, and L. Aolita, Compression, simulation, and synthesis of turbulent flows with tensor trains (2025), arXiv:2506.05477 [physics.flu-dyn].arxiv.org/abs/2506.05477
  64. [64]K. Lively, V. Pagni, and G. Camacho, A quantum-inspired algorithm for wave simulation using tensor networks (2025), arXiv:2504.11181 [quant-ph].DOI
  65. [65]L. Arenstein, M. Mikkelsen, and M. Kastoryano, Fast and flexible quantum-inspired differential equation solvers with data integration (2025), arXiv:2505.17046 [math.NA].arxiv.org/abs/2505.17046
  66. [66]N. Gourianov, P. Givi, D. Jaksch, and S. B. Pope, Tensor networks enable the calculation of turbulence probability distributions, Science Advances 11, eads5990 (2025).DOI
  67. [67]I. Demirdžić and S. Muzaferija, Numerical method for coupled fluid flow, heat transfer and stress analysis using unstructured moving meshes with cells of arbitrary topology, Computer Methods in Applied Mechanics and Engineering 125, 235 (1995).DOI
  68. [68]H. G. Weller, G. Tabor, H. Jasak, and C. Fureby, A tensorial approach to computational continuum mechanics using object-oriented techniques, Computer in Physics 12, 620 (1998).DOI
  69. [69]M.-C. Lai and C. S. Peskin, An immersed boundary method with formal second-order accuracy and reduced numerical viscosity, Journal of Computational Physics 160, 705 (2000).DOI
  70. [70]C. S. Peskin, The immersed boundary method, Acta Numerica 11, 479–517 (2002).
  71. [71]K. Taira and T. Colonius, The immersed boundary method: A projection approach, Journal of Computational Physics 225, 2118 (2007).DOI
  72. [72]R. Verzicco, Immersed boundary methods: Historical perspective and future outlook, Annual Review of Fluid Mechanics 55, 129 (2023).DOI
  73. [73]S. Völkner, J. Brunswig, and T. Rung, Analysis of non-conservative interpolation techniques in overset grid finite-volume methods, Computers & Fluids 148, 39 (2017).DOI
  74. [74]A. Sharma, S. Ananthan, J. Sitaraman, S. Thomas, and M. A. Sprague, Overset meshes for incompressible flows: On preserving accuracy of underlying discretizations, Journal of Computational Physics 428, 109987 (2021).DOI
  75. [75]E. Kornev, S. Dolgov, K. Pinto, M. Pflitsch, M. Perelshtein, and A. Melnikov, Numerical solution of the incompressible Navier-Stokes equations for chemical mixers via quantum-inspired Tensor Train Finite Element Method (2023), arXiv:2305.10784 [physics.flu-dyn].arxiv.org/abs/2305.10784
  76. [76]A. J. Chorin, Numerical solution of the Navier-Stokes equations, Mathematics of Computation 22, 745 (1968).DOI
  77. [77]J. Kim and P. Moin, Application of a fractional-step method to incompressible Navier-Stokes equations, Journal of Computational Physics 59, 308 (1985).DOI
  78. [78]S. Patankar, Numerical heat transfer and fluid flow (CRC press, 2018).DOI
  79. [79]J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, Julia: A fresh approach to numerical computing, SIAM review 59, 65 (2017).arxiv.org/abs/1411.1607
  80. [80]I. V. Oseledets, Tensor-Train decomposition, SIAM Journal on Scientific Computing 33, 2295 (2011).DOI
  81. [81]I. V. Oseledets and E. E. Tyrtyshnikov, Breaking the curse of dimensionality, or how to use SVD in many dimensions, SIAM Journal on Scientific Computing 31, 3744 (2009).DOI
  82. [82]B. N. Khoromskij, Tensor numerical methods for high-dimensional pdes: Basic theory and initial applications (2014).arxiv.org/abs/1408.4053
  83. [83]R. Penrose et al., Applications of negative dimensional tensors, Combinatorial mathematics and its applications 1, 3 (1971).
  84. [84]I. Oseledets and E. Tyrtyshnikov, TT-cross approximation for multidimensional arrays, Linear Algebra and its Applications 432, 70 (2010).DOI
  85. [85]M. K. Ritter, Y. Núñez Fernández, M. Wallerberger, J. von Delft, H. Shinaoka, and X. Waintal, Quantics tensor cross interpolation for high-resolution parsimonious representations of multivariate functions, Phys. Rev. Lett. 132, 056501 (2024).DOI
  86. [86]I. V. Oseledets, Constructive representation of functions in low-rank tensor formats, Constructive Approximation 37, 1 (2012).DOI
  87. [87]R. Orus, A practical introduction to tensor networks: Matrix product states and projected entangled pair states, Annals of Physics 349 (2014) 117-158 349, 117 (2013).DOI
  88. [88]V. A. Kazeev and B. N. Khoromskij, Low-rank explicit qtt representation of the laplace operator and its inverse, SIAM Journal on Matrix Analysis and Applications 33, 742 (2012).DOI
  89. [89]J. D. Biamonte, S. R. Clark, and D. Jaksch, Categorical tensor network states, AIP Advances 1, 10.1063/1.3672009 (2011).DOI
  90. [90]S. V. Dolgov, TT-GMRES: solution to a linear system in the structured tensor format, Russian Journal of Numerical Analysis and Mathematical Modelling 28, 149 (2013).arxiv.org/abs/1206.5512
  91. [91]S. Holtz, T. Rohwedder, and R. Schneider, The alternating linear scheme for tensor optimization in the tensor train format, SIAM Journal on Scientific Computing 34, A683 (2012).DOI
  92. [92]I. V. Oseledets and S. V. Dolgov, Solution of linear systems and matrix inversion in the TT-format, SIAM Journal on Scientific Computing 34, A2718 (2012).DOI
  93. [93]S. V. Dolgov and D. V. Savostyanov, Alternating minimal energy methods for linear systems in higher dimensions, SIAM Journal on Scientific Computing 36, A2248 (2014).DOI
  94. [94]C. Grossmann, H.-G. Roos, and M. Stynes, Numerical Treatment of Partial Differential Equations (Springer Berlin Heidelberg, 2007).DOI
  95. [95]S. Paeckel, T. Köhler, A. Swoboda, S. R. Manmana, U. Schollwöck, and C. Hubig, Time-evolution methods for matrix-product states, Annals of Physics 411, 167998 (2019).
  96. [96]E. M. Stoudenmire and S. R. White, Minimally entangled typical thermal state algorithms, New Journal of Physics 12, 055026 (2010).DOI
  97. [97]J. H. Ferziger, M. Perić, and R. L. Street, Computational Methods for Fluid Dynamics (Springer International Publishing, 2020).
  98. [98]J. F. Thompson, F. C. Thames, and C. Mastin, Automatic numerical generation of body-fitted curvilinear coordinate system for field containing any number of arbitrary two-dimensional bodies, Journal of Computational Physics 15, 299 (1974).DOI
  99. [99]J. F. Thompson, General curvilinear coordinate systems, Applied Mathematics and Computation 10-11, 1 (1982).DOI
  100. [100]W. J. Gordon and C. A. Hall, Construction of curvilinear co-ordinate systems and applications to mesh gener- ation, International Journal for Numerical Methods in Engineering 7, 461 (1973).
  101. [101]W. J. Gordon and C. A. Hall, Transfinite element methods: Blending-function interpolation over arbitrary curved element domains, Numerische Mathematik 21, 109 (1973).
  102. [102]U. M. Ascher, S. J. Ruuth, and R. J. Spiteri, Implicit-explicit Runge-Kutta methods for time-dependent partial differential equations, Applied Numerical Mathematics 25, 151 (1997), special Issue on Time Integration.DOI
  103. [103]B. Koc, S. Rubino, T. Chacón Rebollo, and T. Iliescu, Residual-based data-driven variational multiscale reduced order models for parameter-dependent problems, Computational and Applied Mathematics 44 (2025).DOI
  104. [104]S. Mittal and B. Kumar, Flow past a rotating cylinder, Journal of Fluid Mechanics 476, 303–334 (2003).DOI
  105. [105]S. Kang, H. Choi, and S. Lee, Laminar flow past a rotating circular cylinder, Physics of Fluids 11, 3312–3321 (1999).DOI
  106. [106]M. Visbal and D. Rizzetta, Large-eddy simulation on curvilinear grids using compact differencing and filtering schemes, Journal of Fluids Engineering-transactions of The Asme - J FLUID ENG 124 (2002).DOI
  107. [107]S. Deck, P.-E. Weiss, and N. Renard, A rapid and low noise switch from rans to wmles on curvilinear grids with compressible flow solvers, Journal of Computational Physics 363, 231 (2018).DOI
  108. [108]N. Gourianov, M. Lubasch, S. Dolgov, Q. Y. van den Berg, H. Babaee, P. Givi, M. Kiffner, and D. Jaksch, A quantum-inspired approach to exploit turbulence structures, Nature Computational Science 2, 30 (2022).DOI
  109. [109]E. Ye and N. F. G. Loureiro, Quantum-inspired method for solving the Vlasov-Poisson equations, Physical Review E 106, 035208 (2022).DOI
  110. [110]P. Secular, N. Gourianov, M. Lubasch, S. Dolgov, S. R. Clark, and D. Jaksch, Parallel time-dependent variational principle algorithm for matrix product states, Phys. Rev. B 101, 235123 (2020).DOI
  111. [111]L. Prandtl, Applications of modern hydrodynamics to aeronautics, Tech. Rep. (1923).
  112. [112]N.-L. van Hülst, P. Siegl, P. Over, S. Bengoechea, T. Hashizume, M. G. Cecile, T. Rung, and D. Jaksch, A software example for “quantum-inspired fractional step method in curvilinear coordinates” (2025).
  113. [113]ENGYS Ltd, Helix (2025).
  114. [114]B. Fornberg, Steady viscous flow past a circular cylinder up to reynolds number 600, Journal of Computational Physics 61, 297 (1985).
  115. [115]A. E. Hamielec and J. D. Raal, Numerical studies of viscous flow around circular cylinders, The Physics of Fluids 12, 11 (1969).DOI

Paper details

vlnc:W_a6m6kec9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents