In this paper
AbstractIntroductionProblem settingMethodologyMotivation: from time-delay augmentation to spectral modesSpectral Proper Orthogonal DecompositionOverview of the SPOD-GPR frameworkStage 1: parametric PODRegression of orthonormal bases on a Grassmann chartStage 2: parametric SPOD-GPR on the POD coordinatesGaussian process regressionPredictionBaseline: POD-GPRNumerical experimentsSet-up and error measuresTwo-dimensional Kuramoto–Sivashinsky equationPeriodic regimeTravelling-wave regimeFlow past a cylinder: Reynolds-number parametrisationFlow past an elliptic obstacle: geometric parametrisationDiscussionConclusionsReferencesWorksPhysics and Astronomy
Parametric spatio-temporal reduced-order modelling via Spectral Proper Orthogonal Decomposition and Gaussian Process Regression
Abstract
We present SPOD-GPR, a non-intrusive, parametric, spatio-temporal reduced-order model (ROM) that combines Proper Orthogonal Decomposition (POD), Spectral POD (SPOD), and Gaussian Process Regression (GPR). A parametric POD first projects the high-dimensional state onto a low-dimensional subspace. The reduced trajectories are then divided into windowed sections, transformed into the frequency domain, and truncated to the most energetic frequencies. SPOD is then performed independently at each retained frequency, yielding temporally coherent modes and coefficients. The parametric dependence of the POD and SPOD bases is modelled by GPR on a local chart of the Grassmann manifold after Procrustes alignment. The SPOD modal coefficients are learned both as explicit functions of the parameters and through a parameter-conditioned autoregressive model that allows for temporal extrapolation beyond the training window. Gaussian processes are sampled jointly and pathwise via Matheron conditioning, ensuring that each predictive realization carries uncertainty quantification consistent with the posterior distribution. The method is tested on four parametric examples where the parameter values and the prediction horizon lie outside the training data: the two-dimensional anisotropic Kuramoto–Sivashinsky equation in a time-periodic and a travelling-wave regime, each parametrised by one of its two coefficients; flow past a cylinder parametrised by the Reynolds number; and flow past an elliptic obstacle parametrised by its semi-axes. In all cases, SPOD-GPR reduces the normalised mean absolute error compared to a POD-GPR baseline, and improves the calibration of the predictive uncertainty, especially in the fluid flow cases, where POD-GPR is severely overconfident. This opens the way to spectral, uncertainty-aware reduced-order models for many-query and real-time applications, such as digital twins.
Introduction
Applications that involve a large number of queries or require real-time performance, such as design optimisation, control, uncertainty quantification, and the use of digital twins, rely on solving parametrised, time-dependent partial differential equations (PDEs) repeatedly. Reduced-order models (ROMs) meet this demand by replacing the high-fidelity full-order model (FOM) with a low-dimensional surrogate [1, 2]. Unlike projection-based (intrusive) ROMs, which require access to the discretised operators and are often impractical when the FOM is provided as legacy software or in the form of a commercial package, nonintrusive and data-driven ROMs determine the reduced dynamics directly from snapshot data [3, 4, 5]. A typical procedure involves first reducing the snapshots using Proper Orthogonal Decomposition (POD) [6, 7], and then fitting the POD coefficients as functions of the parameters and/or time; in this case, Gaussian Process Regression (GPR) [8] is especially attractive since it is data-efficient, offers strong approximation guarantees in the relevant reproducing kernel Hilbert space [9, 10, 11, 12], and provides a predictive variance without having to add any extra modelling effort [3, 4].
Gaussian processes have been widely used in the field of learning dynamical systems from data, whether for modelling one-step transition functions [13, 14], as a way of representing vector fields for ordinary differential equations whose forms are unknown [15, 16, 17], or as priors that are designed to take into account the governing equations [18, 19, 20]. The kernel-based methods for learning dynamics have also been investigated from the point of view of cross-validation and kernel flows [21, 22]. Other linear and sparse-regression methods include Dynamic Mode Decomposition (DMD) and various related variants [23, 24, 25, 26], as well as sparse identification of nonlinear dynamical systems [27].
A major disadvantage of most POD-based nonintrusive ROMs is that the POD modes are arranged purely according to their energy and do not show how the dynamics are organised in time. In flow situations characterised by travelling waves, quasi-periodic motions, or the presence of several coherent time scales, one POD mode can combine features associated with different frequencies; consequently, its modal coefficient shows a complex, multiscale temporal behaviour which is harder to learn and, in particular, more difficult to extrapolate. Spectral Proper Orthogonal Decomposition (SPOD), which was first introduced by Lumley [6, 28] and later considered again by Towne, Schmidt and Colonius [29, 30], overcomes this problem by building an optimal, energy-ranked basis for each frequency. The SPOD modes are coherent both in space and in time, correspond to optimally averaged DMD modes for statistically stationary flows, and agree with resolvent modes under certain assumptions [29]. SPOD is now widely employed to study jets [31], wall-bounded turbulence [32] and multiphase flows [33], and has given rise to a wide range of algorithms, including streaming [34], multitaper [35], gappy [36] and robust [37] versions, as well as methods for low-rank reconstruction and denoising [38]. It should be noted that a separate filter-based method with the same name was put forward in [39]. In the context of reduced-order modelling, the SPOD coefficients have been predicted by means of neural networks [40], showing that the frequency-resolved latent representation can be easier to learn than the POD version.
In this paper we develop these concepts and present a parametric, probabilistic spatio-temporal ROM which combines SPOD with GPR. The main contributions are:
We justify our use of SPOD for the purpose of learning dynamics by means of a time-delay (block) augmentation of the state: even though such an augmentation is capable of converting a non-autonomous scalar signal into an autonomous system, when POD is applied to the augmented snapshot matrix (this is known as block POD) it entangles the spatial and temporal scales; carrying out a windowed Fourier transform along the delay coordinate separates these scales and yields SPOD (§ 3.1).
We present a two-stage reduction procedure, first POD and then SPOD, enabling the application of large SPOD windows to high-dimensional datasets, together with a frequency truncation that applies to all the training parameters (§ 3.4 and § 3.6).
We make all the bases parametric by solving separate (S)POD problems at each value of the parameter, aligning the resulting bases using orthogonal/unitary Procrustes rotations, and then performing the regression on a chart of the Grassmann manifold [41], centred at the
Karcher mean of the training subspaces; in this way the interpolated modes remain exactly orthonormal (§ 3.5).
We use three ridge-GP surrogates to model the SPOD modes and coefficients: one for mapping the parameters to the modes, another for mapping the parameters to the trajectories, and a third which is a parameter-conditioned autoregressive model for time extrapolation. By applying joint pathwise (Matheron) sampling [42] across all of the GPs, uncertainty is propagated to the physical field, and the variance scale of each GP is adjusted using the leave-one-out residuals (§ 3.7).
We assess the method on four parametric benchmark problems, looking at test parameters and prediction horizons which were not encountered during training, and we compare it with a POD-GPR baseline that employs the same regression structure (§ 4).
The rest of the paper is organised as follows: § 2 describes the problem setting, § 3 sets out the SPOD-GPR methodology, § 4 gives the details of the numerical experiments, § 6 looks at the limitations, and § 7 summarises the conclusions.
Problem setting
Let represent a spatial domain and a parameter set. We consider a time-dependent parametric PDE with solution , which is spatially discretised by means of a FOM with degrees of freedom and evaluated at uniformly spaced time instants . The discrete state at time is denoted by , and we introduce a symmetric positive-definite weighting matrix , such as a finite-element mass matrix, defining the discrete inner product .
The training set comprises snapshots for each of training parameter instances belonging to the set ,
The aim is to predict the system’s state for a new parameter value over a future time horizon extending beyond the training time interval, together with an estimate of the associated predictive uncertainty. This involves carrying out parametric interpolation and temporal extrapolation at the same time, a problem that is considerably more difficult than tackling either of these tasks separately.
Methodology
Motivation: from time-delay augmentation to spectral modes
It is simplest to deduce a dynamical law from data when the state is autonomous, that is, when its time derivative (or its one-step increment) depends only on the state at that moment. In reality, reduced coordinates are almost never autonomous since projecting onto a small number of modes eliminates the information needed to close the evolution equations. A common solution, one that is based on delay-embedding theorems [43], is to extend the state by adding to it time-lagged copies of itself,
This kind of augmentation is useful in situations where a system is not originally autoregressive but does become autoregressive after time-augmented variables have been incorporated. A simple example is the case of : since , the time derivative cannot be given as a function of alone and therefore no first-order autonomous scalar equation can generate this signal. On the other hand, if the state is augmented with a shift of in time, then the pair obeys a linear autonomous system.
A simple way of reducing the augmented state (2) is to apply to the stacked snapshot matrix the same subspace (Grassmannian) framework used in POD, that is to treat the delayed copies as a single extended vector. The result, however, is that each mode becomes a space–time structure combining both spatial and temporal scales: one mode may contain both slow, large-scale and fast, small-scale features, and the temporal coherence of the underlying structures is thereby lost.
The way to proceed is to diagonalise the delay axis before carrying out the compression; for this purpose, we apply a windowed discrete Fourier transform (DFT) over the lag index,
where is a tapering window. This process splits the augmented state into separate frequency components. We then perform an independent POD for each . This is nothing other than Spectral Proper Orthogonal Decomposition (SPOD) [6, 29]: it produces temporally coherent modes, each linked to a single oscillation frequency, while at the same time maintaining POD optimality within each frequency band. As a result, the coefficients correspond to one coherent structure and one characteristic frequency each, which can make their evolution smoother and easier to learn than that of conventional POD coefficients. This insight is the basis of the approach presented below.
Spectral Proper Orthogonal Decomposition
We give a brief summary of SPOD in the formulation adopted here [29, 30]. Let , , denote a (mean-removed) vector-valued time series and let be a symmetric positive-definite weighting matrix defining the inner product and the corresponding norm . The data are divided into overlapping segments each containing samples (with odd), using a hop size , which results in . In this study a fully sliding approach is adopted with (), which maximises the number of segments and gives one segment for each admissible window position, that is . A Hamming window is applied to each segment, after which the segment is Fourier transformed,
where, since the data are real-valued, only the non-negative frequencies are kept. At each frequency, the block realisations are then assembled into .
The SPOD modes at frequency are the eigenvectors of the weighted cross-spectral density ; in practice, however, the snapshots method is used,
where denotes the conjugate transpose, consists of the eigenvectors and of the non-zero SPOD eigenvalues, and the modes are reconstructed as . An equivalent method, which we follow in our implementation, obtains by means of a truncated (randomised) singular value decomposition of the weighted Fourier snapshots. More specifically, if is a Cholesky factorisation, we compute and recover the first modes as . By retaining the most important modes at each frequency, the SPOD coefficients are
In each row relates to one SPOD mode and each column to a block; consequently, at any given frequency and mode one has a sequence of complex coefficients which show how the relevant coherent structure evolves from block to block.
Inverse transform for the sliding window. When , each interior sample appears in overlapping blocks; the sample at index is recovered by using the central point of block , this being the point at which the Hamming window reaches its maximum () and where the reconstruction is affected as little as possible by frequency and rank truncation:
with denoting the set of frequency bins that are retained, and and when , these weights accounting for the contribution of the conjugate negative-frequency bins (there is no Nyquist bin because is odd). Unlike the method of taking the last sample from each block, sampling at the centre avoids dividing by the small values of the tapered window near the ends and thus reduces the Gibbs-type oscillations at the boundaries between blocks.
Overview of the SPOD-GPR framework
The proposed model approximates the parametric solution by
where the temporal mean is denoted by , is a parametric POD basis having orthonormal columns in weighted coordinates, and are the reduced coordinates. These coordinates are in turn represented by means of a parametric SPOD of rank for each of the frequency bins. The approach is divided into two stages: in stage 1 a rank- parametric POD carries out the spatial compression (§ 3.4); in stage 2 a parametric SPOD-GPR is employed to model the temporal behaviour of the POD coordinates (§ 3.6). The reason for this two-stage procedure is computational efficiency: since blocking, Fourier transforms, and per-frequency SVDs depend on the spatial dimension, applying SPOD to coordinates allows the use of large time windows and thus enables a finer frequency resolution at a reasonable cost. As the POD coordinates are defined with respect to the -inner product, the reduced energy corresponds to the -weighted physical energy captured by the POD, and no further weighting is necessary in stage 2 (). A summary of the SPOD-GPR step is given in Figure 1: the data obtained from the FOM (after having been reduced by POD) are divided into overlapping time blocks, multiplied by a moving temporal window, and transformed into the frequency domain; GPR models are then trained on the resulting SPOD coefficient sequences and are used to predict the spatio-temporal dynamics for new parameters and times.

Figure 1. Sketch of the Gaussian Process Regression stage based on SPOD (when applied to the POD coordinates of stage 1). The high-fidelity data are obtained from the full-order model (1), divided over time into overlapping blocks (2), weighted using a temporal window and transformed into frequency space (3). GPR models are then trained on the resulting sequences of SPOD coefficients and are used to reconstruct the spatio-temporal dynamics (4).
Stage 1: parametric POD
For every training parameter, from the set , we first evaluate the temporal average and then, with , assemble the weighted fluctuation matrix
The approximation is obtained by applying a truncated randomized SVD, where has orthonormal columns and contains the POD coefficients, one column per time instant. The rank is the same for all the parameters and, similarly to the common frequency set of § 3.6, is selected from the energies averaged over the training set: denoting by the singular values of , we compute and take as the smallest rank for which exceeds a prescribed fraction of the averaged total energy (alternatively, can be specified explicitly). It should be noted that a single global POD is not constructed; rather, each parameter is linked to its individual basis, which allows the modes to vary with .
Since an SVD determines its singular subspaces only up to an orthogonal change of basis (for example, due to sign ambiguities and rotations when the singular values are nearly identical), the unprocessed bases are not smooth functions of . In order to ensure consistency, each basis is aligned with a common reference frame . This frame is a representative of the Karcher (Riemannian) mean of the training subspaces on the Grassmann manifold [44, 45, 46], that is, the minimiser of , with the geodesic distance defined by the principal angles; it is computed by the standard fixed-point iteration that alternates the Grassmann logarithm and exponential maps. Placing the reference at the centre of the data keeps the principal angles to all the training subspaces small, and hence away from the singularity of the chart of § 3.5. The alignment is performed by means of an orthogonal Procrustes problem,
which yields the rotated basis and coefficients and ; these preserve the reconstruction, since . For complex frames the same formula holds with replaced by . The mapping is modelled using the manifold-aware GPR as described in § 3.5, while the mean field is learned by means of a separate GP in physical space. It is important to note that the mean is not included in the POD coordinates, because if the two independently interpolated terms were multiplied together accuracy would be degraded between training points when the individual parameter-dependent bases contain low-energy directions that are specific to each parameter.
Regression of orthonormal bases on a Grassmann chart
Simply regressing the individual components of would generally violate the orthonormality condition. Instead, we carry out the regression on coordinates that are derived from a local chart of the Grassmann manifold which is based at a reference basis [41, 47]. Orthonormality is understood in the Euclidean inner product, since the physical weighting has already been absorbed ( in stage 1, in stage 2). Let (or when considering SPOD modes) represent the reference frame, and let be the aligned frame. We parametrise the subspace spanned by ; its orientation within that subspace is fixed by the Procrustes alignment, as shown below. We define and
so that the subspace associated with can be expressed as the graph over the reference subspace; this gives rise to an affine chart on the Grassmann manifold and is well defined so long as is nonsingular (equivalently, none of the principal angles between the two subspaces equals ). The Procrustes alignment fixes the in-subspace orientation but does not change this nonsingularity condition. Since , the matrix has only independent parameters. In order to determine these parameters without having to explicitly construct an orthogonal complement to , we select rows of by means of the maxvol algorithm [48, 49], initialised with partially pivoted Gaussian elimination; this gives a permutation such that the subblock in has (locally) maximal volume , which guarantees that all the entries of are bounded by one in modulus and hence that is well conditioned; we then define as and, upon writing , define
It is possible to verify that , thus demonstrating that gives a true and unconstrained parameterisation of the subspace.
No additional coordinate is needed for the orientation of the frame within its subspace, because it is fixed by the Procrustes alignment. Indeed, by (11), the aligned frame satisfies , which is Hermitian positive definite; since and , it follows that , i.e. . From the predicted values of , the frame is therefore reconstructed as
Conversely, for any this construction yields an exactly orthonormal frame () with Hermitian positive definite, i.e. a frame that is Procrustes-aligned to . Every predicted or sampled basis is therefore a valid orthonormal frame, expressed in the same aligned convention as the training frames and the rotated modal coefficients, so that predicted modes and coefficients remain consistent. This requires the Procrustes reference to coincide with the base of the chart and no further rotation to be applied after the alignment. Since lies in a vector space, its entries can be regressed directly. The procedure is sketched in Figure 2.
Figure 2. Schematic of the regression of orthonormal frames on the Grassmann chart, in stage-1 notation (in stage 2, and for each bin ). The aligned training frames are mapped to unconstrained coordinates relative to the reference frame , which corresponds to . The GP predicts , which (14) maps back to an exactly orthonormal frame , Procrustes-aligned to . Every point of the coordinate space corresponds to a valid frame.
Specialisation. In stage 1 the chart is applied with real frames, and , so that . In stage 2 it is applied separately at each retained bin with complex frames , the Karcher-mean reference of bin (§ 3.6), and , i.e. with and , so that , whose real and imaginary parts are fitted separately. The cost is per frame in stage 1 and per frame and bin in stage 2.
Stage 2: parametric SPOD-GPR on the POD coordinates
The aligned, zero-mean POD coordinates are used as input in stage 2, i.e. is the -th column of ; they form the time series of § 3.2 with and .
Common frequency set. For each training parameter we calculate the Parseval-weighted DFT energy in each bin, that is , then average the energies over the training set and choose the bins having the highest energy, calling this set . This common set of bins is used for all the parameters, so there is no need to interpolate the frequencies in and the computations can be limited to the selected bins (by means of a direct DFT).
Per-parameter SPOD and alignment. For each training parameter, a separate rank- SPOD is carried out on the bins in , resulting in complex modes and coefficients . At each bin, a reference frame is computed, as in stage 1, as a representative of the Karcher mean of the training subspaces , here on the complex Grassmann manifold [44, 45]. The modes corresponding to each parameter are then rotated so as to coincide with by solving the unitary Procrustes problem, that is, the complex version of (11); the coefficients are then rotated accordingly, and , with unitary, this rotation preserving the reconstruction . No further rotation (for instance, an alignment across frequencies) is applied after this step, as required in § 3.5. The aligned modes are finally transformed, on a bin-by-bin basis, into coordinates as explained in § 3.5. For each retained frequency, the reference and aligned frames used in that construction are and , respectively. The corresponding chart coordinates are therefore , and the reconstruction (14) is applied separately at each .
Three surrogates. We define to be the vector consisting of the real and imaginary parts of . Writing , we fix the ordering
We train three Gaussian process surrogates:
the SPOD-basis GP , a single multi-output GP with real outputs, whose outputs are mapped back to orthonormal SPOD modes using (14) with the frequency-specific reference frame ;
the trajectory GP , which jointly predicts the complete coefficient trajectory over the training window, yielding a reconstruction that is truly parameter-dependent within that window;
the dynamics GP , which is trained using pairs that have been aggregated across all training parameters and blocks and serves to establish a single evolution rule that is conditioned on the parameters, being employed thereafter for extrapolation beyond the training window.
In the dynamics GP the parameter appears only as an input feature. By predicting the increment rather than the next state, the learning target is kept smaller and more nearly stationary when the trajectory changes smoothly from one block to the next, as is the case with the sliding window . During extrapolation, this GP is used autoregressively. The three surrogates and the way in which they are combined at prediction time are sketched in Figure 3.

Figure 3. Sketch of the three stage-2 Gaussian process surrogates and of their combination at prediction time (§ 3.8). For a query parameter , the SPOD-basis GP predicts the chart coordinates of the SPOD modes, which are mapped back to orthonormal modes by (14); the trajectory GP predicts the spectral coefficients over the training window, and its last block initialises the autoregressive rollout of the dynamics GP. Modes and coefficients are combined through the central inverse (8), and the physical field is then reconstructed with the stage-1 quantities via (9). The dashed arrow indicates the in-window reconstruction obtained from the trajectory GP alone. All the GPs are sampled jointly, in lock-step, by means of Matheron’s rule.
Gaussian process regression
Each of the surrogate models (that is, the POD-mode GP and the stage 1 mean GP, along with the three stage 2 GPs) makes use of the same regression backend. The inputs are transformed onto the unit hypercube and the outputs are standardised. A Matérn-9/2 unit-variance correlation kernel is used, this function having a length scale , where is the fill distance of a quasi-uniform set of inducing points obtained by farthest-point sampling within a ball that contains the inputs, and is a constant that is specified by the user. The resulting model is
and its posterior mean is identical to the kernel ridge regression predictor [10]. In order to achieve scalability when dealing with large data sets (in particular for the dynamics GP, which involves observations), we approximate the kernel using Nyström features [50]. More specifically, by selecting landmarks from (a subset of) the training samples together with the inducing points, we can write , thus preventing any matrix from being formed. If we denote the feature matrix by , then the posterior-mean weights are given by .
Pathwise sampling. The posterior draws are obtained by using Matheron’s rule (pathwise conditioning) [42],
with . Since each draw is an entire function, it can be evaluated consistently at a number of different input points. This feature is important for the autoregressive rollout, since the output of each step is used as the input for the following step. We sample functions from all of the GPs in a synchronized (lock-step) way; specifically, sample 0 has , and therefore matches the posterior mean for each GP. The draw simplifies: with , one has , hence , whose perturbation is distributed as , i.e. the exact weight-space posterior. We therefore draw , with , and , so that a single factorisation of , whose size is the number of Nyström features, serves all outputs and samples.
Calibration of the variance. The variance hyperparameter has no effect on the posterior mean and is therefore adjusted separately for each output using the closed-form leave-one-out (LOO) residuals. If we let and , the LOO residual at point is given by and has a unit-scale standard deviation of . By the Woodbury identity, , so that and ; both are computed from the factorisation of , and no matrix is formed. The value of is selected so that the nominal 95% interval includes 95% of the standardised LOO residuals. In the case of the dynamics GP, it is too optimistic to omit just one step since adjacent blocks from the same trajectory are almost identical; instead, entire training trajectories are left out (leave-one-parameter-out), as this more accurately represents prediction for an unseen parameter. For this last model, the variance calibration is therefore based on the leave-one-parameter-out residuals rather than the pointwise LOO residuals.
Prediction
For a given query parameter and a prediction horizon of steps beyond the training interval, and for each sample :
the trajectory GP is queried at , and its final block is taken as the parameter-dependent initial condition;
the dynamics GP is propagated according to for blocks, where the additional blocks offset the delay induced by the central inverse (8);
the SPOD modes are drawn from the SPOD-basis GP and then decoded using (14), while the POD coordinates are recovered via (8);
the POD basis and the mean are sampled from the stage-1 GPs and the complete field is then reconstructed using (9).
The estimate is given by sample 0 (the posterior mean in the latent GP outputs, followed by nonlinear reconstruction), and the pointwise predictive variance is obtained as the empirical mean squared deviation of samples from sample 0. As a result, all the sources of uncertainty (the mean field, the POD basis, the SPOD modes, the initial condition, and the dynamics) are propagated together and in a nonlinear manner into the physical field. If the same method is applied but steps 1–2 are omitted—that is, if the trajectory GP is used for each block—a parameter-dependent reconstruction is obtained over the training interval.
Baseline: POD-GPR
As a reference, we consider a parametric POD-GPR that employs exactly the same regression pipeline: for each parameter, the weighted POD bases of rank are aligned by means of a Procrustes analysis and then mapped or regressed onto a Grassmann chart, after which a mean GP, a trajectory GP defined by , and a dynamics GP given by are used, with the same kernels, calibration procedure, and Matheron sampling. Here , the -th column of the aligned rank- coefficient matrix, denotes the POD reduced state of the baseline, distinct from the spectral latent state of the proposed method. For a query parameter , the baseline prediction reads
where is reconstructed from the chart via (14), and the rollout starts from the last training state given by the trajectory GP. Compared with (9), the reduced coordinates are propagated directly in time, one snapshot per step, instead of through the windowed SPOD blocks and the central inverse (8). The only difference from SPOD-GPR is the choice of spectral latent representation.
Numerical experiments
Set-up and error measures
In every experiment the parameters are selected uniformly across their respective ranges. For each simulation the first 95% of the snapshots (with rounding so that the blocking procedure is exact) are used for training and the remaining ones for testing. Since the test parameters are not seen during training, each test involves both parametric interpolation and temporal extrapolation. We apply SPOD with a window of , a stride of , and a Hamming taper; a stage-1 POD rank of ; and GPs with a Matérn-9/2 kernel, a ridge value of , 50 inducing points for the dynamics GP, and Matheron samples. The case-specific hyperparameters are provided in Table 1. In all the cases, the truncation of both methods (the POD modes of POD-GPR, and the frequency bins with modes each of SPOD-GPR) preserves approximately 95% of the training energy. Let and stand for the actual and predicted test arrays over the test times, degrees of freedom, and test parameters; let be the overall range of , and let represent the predictive variance. We give the normalised mean absolute error (NMAE) and the normalised root-mean-square error (NRMSE),
| Case | parameter | ||||||
| KS, periodic | 80/20 | 950/50 | 1 | 3 | 2 | 5 | |
| KS, travelling wave | 80/20 | 950/50 | 4 | 20 | 6 | 5 | |
| Cylinder, Reynolds | 80/20 | 950/50 | 2 | 3 | 2 | 50 | |
| Elliptic obstacle | 80/20 | 950/50 | 2 | 3 | 2 | 50 |
Table 1.
as well as the normalised predicted standard deviation NSD , which for a well-calibrated model should be directly comparable to the NRMSE; their ratio NSD/NRMSE (spread–skill ratio) is therefore close to one for a calibrated model, and smaller (larger) than one for an overconfident (conservative) model. We also give the normalised 95% interval half-width 1.96 NSD and the empirical coverage COV, defined as the proportion of entries for which . The time histories are obtained by averaging only over the degrees of freedom and the test parameters, while the spatial maps are derived by averaging over time and the test parameters. Furthermore, we provide the spatial average of , that is, the mean over the test parameters of the pointwise magnitude, as an indication of the amplitude of the predicted fields.
The results for the four cases are summarised in Table 2 and discussed below.
| Case | Method | NMAE | NRMSE | HW95 | NSD/NRMSE | COV | |
| KS, periodic | SPOD-GPR | 0.160 | 0.65 | 0.774 | 2.76 | ||
| POD-GPR | 0.762 | 2.7 | 0.989 | 2.79 | |||
| KS, travelling wave | SPOD-GPR | 0.238 | 0.295 | 0.520 | 0.90 | 0.876 | 1.81 |
| POD-GPR | 0.275 | 0.327 | 1.01 | 1.6 | 0.926 | 2.08 | |
| Cylinder, Reynolds | SPOD-GPR | 0.50 | 0.723 | 1.024 | |||
| POD-GPR | 0.006 | 0.999 | |||||
| Elliptic obstacle | SPOD-GPR | 0.327 | 2.9 | 0.992 | 1.000 | ||
| POD-GPR | 0.043 | 0.999 |
Table 2.
Table 1: Case-dependent settings. is the number of SPOD modes per frequency (SPOD-GPR) and the number of POD modes (POD-GPR); is the number of frequency bins retained; is the length-scale coefficient of all the GPs; are the numbers of training and test parameters, and the numbers of training and test snapshots for each parameter.
Table 2: Space–time averaged results for the test window (unseen parameters and future times). HW95: the normalised 95% half-width NSD; COV: the empirical coverage of the 95% interval; : the space–time average of the mean magnitude ; NSD/NRMSE: the spread–skill ratio, close to one for a calibrated model. The best errors are given in bold.
Two-dimensional Kuramoto–Sivashinsky equation
We study the anisotropic two-dimensional Kuramoto–Sivashinsky equation in the rescaled formulation of [51],
posed on the periodic domain , with a zero spatial average and the initial condition . The parameters and indicate the dimensions of the original domain; by varying these parameters one obtains steady, time-periodic, quasi-periodic or chaotic behaviour [51]. The PDE is discretised using a Fourier pseudospectral scheme and the time advancement is carried out with the exponential time-differencing fourth-order Runge–Kutta method (ETDRK4) using a time step of . The weights W are taken to be the uniform quadrature weights corresponding to the periodic grid, and every fifth parameter setting is kept as test data. Two different regimes are studied, these being distinguished by the values of the parameters and by the resulting dynamical characteristics.
Periodic regime
In the first regime we take the value and select from the interval . The dynamics that result show a spatially stationary cellular pattern with an amplitude which undergoes periodic temporal oscillations (as can be seen in Figure 4 and Figure 6, top row). We integrate the equation until time , eliminate the first transients, and store one snapshot every 20 steps, thus obtaining 1000 snapshots with a time step of ; the test interval is , with time measured from the end of the discarded transient. Since the time dependence is simple, SPOD-GPR needs only one mode for each frequency () and frequency bins, whereas POD-GPR uses POD modes.
Figure 4. Kuramoto–Sivashinsky equation, periodic regime: full-order solution for three values of (rows) at three time instants (columns), with time measured from the first stored snapshot.
Figure 6. Kuramoto–Sivashinsky equation, periodic regime (): time-averaged maps covering the test window and test parameters; NSD denotes the normalised predicted standard deviation and COV the empirical 95% coverage. The colour scale shown below each row is shared by the maps in that row.
The time traces shown in Figure 5 show the main qualitative difference between the two surrogate models. The space-averaged magnitude from SPOD-GPR stays phase-aligned with the true values for the whole of the prediction period, whereas the output from POD-GPR quickly settles down after just a few steps to a level that is essentially constant , apart from a few step-like jumps between plateaus: it seems that the autoregressive propagation of the POD coefficients is approaching a fixed point and thus the oscillating behaviour is eliminated. Both methods slightly underpredict the amplitude (the space–time means are 2.76 and 2.79 as compared to 2.84). As stated in Table 2, SPOD-GPR reduces the NMAE by about 22% and the NRMSE by around 13%; its error varies periodically with the solution. The spatial fields in Figure 6 show that the errors of both models reflect the cellular structure, with that of SPOD-GPR being smaller almost everywhere.
Figure 5. Kuramoto–Sivashinsky equation in the periodic regime (), test window. On the left: the NMAE (solid), the NRMSE (dashed) and the predicted normalised 95% half-width (dotted) for SPOD-GPR (orange) and POD-GPR (blue); for a calibrated Gaussian predictive distribution, the half-width is about 1.96 times the NRMSE. In the centre: the space average of the mean magnitude , with the reference solution in black. On the right: the empirical coverage of the 95% predictive interval.
In this case, the uncertainty quantification is not very convincing. The prediction interval from the SPOD-GPR method is too narrow (the normalised half-width is 0.16 as against an NRMSE of 0.13, i.e. NSD/NRMSE ), the coverage therefore being 0.77; it tends towards one mainly along the nodal lines of the pattern, where the error is small, and is lower in the other areas. POD-GPR shows the opposite behaviour: its interval is very wide (half-width 0.76, about five times its NRMSE, i.e. NSD/NRMSE ), so that its coverage of 0.99 is obtained at the price of a weakly informative band. Although the spatial distribution of the POD-GPR standard deviation follows that of its error, its amplitude is overestimated.
Travelling-wave regime
In the second regime we take the value and select from the interval . For values in this range the solution forms quasi-periodic, travelling-wave-type coherent structures having a more complicated temporal spectrum. We integrate the equation until ; after having eliminated the initial transient for we take one snapshot every 40 steps, which gives a total of 1000 snapshots with a time step . The test time interval is , in absolute time. We use frequency bins and retain modes in each bin, while the POD-GPR is constructed from POD modes. The configurations at different times and parameters are shown in Figure 7.
Figure 7. Kuramoto–Sivashinsky equation, travelling-wave regime: full-order solution for three values of (rows) at three time instants (columns), with time measured from the first stored snapshot.
Figure 8 illustrates how the errors, the predicted uncertainty, the coverage, and the mean magnitude develop over time. Throughout the entire test period, SPOD-GPR has a smaller NMAE and NRMSE than POD-GPR—by about 13% and 10% respectively (see Table 2). Both methods underestimate the field amplitude, that is, their posterior means are too damped, which is to be expected when dealing with quasi-periodic travelling waves predicted at unseen parameters and beyond the training window: once phase coherence is lost, the prediction which minimises the error usually moves back towards the mean. Interestingly, although SPOD-GPR predicts an even smaller amplitude than POD-GPR it still manages to achieve a lower pointwise error, which suggests a better phase preservation, although pointwise error measures also tend to favour more damped predictions once the phase is lost. The uncertainty from POD-GPR increases rapidly and reaches a 95% half-width comparable to the full data range, consistent with its high coverage (0.93), while SPOD-GPR predicts an interval that is roughly half as wide and has a slightly lower coverage (0.88); in terms of the spread–skill ratio, SPOD-GPR is close to calibration (NSD/NRMSE ),
Figure 8. Kuramoto–Sivashinsky equation in the travelling-wave regime (), test window. On the left: the NMAE (solid), the NRMSE (dashed) and the predicted normalised 95% half-width (dotted) for SPOD-GPR (orange) and POD-GPR (blue); for a calibrated Gaussian predictive distribution, the half-width is about 1.96 times the NRMSE. In the centre: the space average of the mean magnitude , with the reference solution in black. On the right: the empirical coverage of the 95% predictive interval.
whereas POD-GPR is conservative (). The time-averaged maps in Figure 9 confirm that SPOD-GPR gives consistently smaller errors and that its predicted standard deviation follows the structure of the travelling patterns, rather than the smooth, weakly informative envelope produced by POD-GPR.
Figure 9. Kuramoto–Sivashinsky equation, travelling-wave regime (): time-averaged maps covering the test window and test parameters; NSD denotes the normalised predicted standard deviation and COV the empirical 95% coverage. The colour scale shown below each row is shared by the maps in that row; note the narrow range of the NMAE colour scale.
Flow past a cylinder: Reynolds-number parametrisation
A second benchmark involves incompressible flow past a circular cylinder of diameter which is situated at in the channel . The flow is given by the Navier–Stokes system
together with a parabolic inflow of unit mean speed, with , no-slip boundary conditions on both the channel walls and the cylinder surface, and zero pressure at the outlet. The Reynolds number , defined in terms of the mean inflow velocity and the cylinder diameter, is taken as the parameter. The Reynolds numbers are selected uniformly across this range, the total number of samples being 100, 80 of which are used for training and 20 for testing.
The PDEs are advanced by means of an incremental pressure-correction method using Adams–Bashforth/Crank–Nicolson time stepping and Q2–Q1 finite elements on a curved quadrilateral mesh which is refined in the wake region, the implementation being carried out in DOLFINx [52], with a time step . After 1000 initial steps have been taken to enable the vortex street to form, an additional 1000 steps () are stored and then interpolated in space to a coarser, quasi-uniform Q1 mesh that is used by the ROMs. The weight matrix is taken as the lumped Q1 mass matrix on this mesh, and this matrix is duplicated for each velocity component. The testing interval corresponds to approximately one vortex-shedding cycle, while the SPOD window is equivalent to about two cycles; only the highest-energy frequency
bins—given the window resolution (), these are concentrated near the shedding frequency—and modes per bin are retained, while POD-GPR uses POD modes. Representative snapshots of the velocity magnitude are shown in Figure 10.
Figure 10. Flow past a cylinder, Reynolds-number parametrisation: full-order velocity magnitude for three values of (rows) at three time instants (columns), with time measured from the first stored snapshot.
The results are given in Figure 11 and Figure 12. Both surrogate models reproduce the mean flow, but SPOD-GPR is markedly more accurate: its value of is within 0.2% of that of the reference, compared with 2.6% for POD-GPR, and it reduces the NMAE and NRMSE by about 76% and 72% respectively. In both cases, the errors are mostly found in the near wake, the region where vortex shedding takes place and where sensitivity to the parameters is greatest. The uncertainty estimates also differ sharply: the prediction interval of POD-GPR is far too narrow (the half-width is as compared with an NRMSE of , i.e. NSD/NRMSE ) and includes only 0.6% of the test cases, which shows that it is severely overconfident. On the other hand, the 95% half-width of SPOD-GPR is comparable to its NRMSE ( in both cases), i.e. its predicted standard deviation is about half the actual error (NSD/NRMSE ), and its interval reaches 72% coverage: below the nominal 95%, but far better calibrated than POD-GPR. Its predicted standard deviation is concentrated in the near wake, exactly where the errors are located (Figure 12), so it serves as a practical indicator of error. It should be noted, however, that the space-averaged coverage also includes the regions outside the wake, where both the error and the predicted variance are small and the coverage is close to one; within the wake, and in particular in its downstream part, the coverage is lower (Figure 12, bottom row).
Figure 11. Flow past a cylinder, Reynolds-number parametrisation, test window. On the left: the NMAE (solid), the NRMSE (dashed) and the predicted normalised 95% half-width (dotted) for SPOD-GPR (orange) and POD-GPR (blue); for a calibrated Gaussian predictive distribution, the half-width is about 1.96 times the NRMSE. In the centre: the space average of the mean velocity magnitude , with the reference solution in black. On the right: the empirical coverage of the 95% predictive interval.
Figure 12. Flow past a cylinder, Reynolds-number parametrisation: time-averaged maps covering the test window and the test parameters; NSD denotes the normalised predicted standard deviation and COV the empirical 95% coverage. The colour scale shown below each row is shared by the maps in that row (and, for the first row, by the top map).
Flow past an elliptic obstacle: geometric parametrisation
The third test case uses the same channel shape and numerical solver at a fixed , but substitutes the circular cylinder of radius with an ellipse which has semi-axes . A two-parameter vector is thus defined. The values of are selected uniformly from this square; altogether 100 different geometries are generated, of which 80 are used for training and 20 are kept for testing. The change in geometry is applied to both the computational mesh
and the output mesh by means of a smooth radial transformation,
where represents the centre of the obstacle and is a quintic smoothstep function. The function has a value of 1 at the boundary of the obstacle (), so that the circle is mapped precisely onto the ellipse, and reaches 0 when , where the mapping reduces to the identity. As the mesh connectivity is preserved by this mapping, the velocity field calculated in each deformed domain is stored at the nodes of the reference mesh, which ensures that all the snapshots have the same degrees of freedom and possess the same weight matrix (that is, the lumped mass matrix of the reference mesh). This is in line with the standard approach for geometrically parametrised ROMs, wherein the solution manifold is represented on a fixed reference configuration. The channel set-up and four of the test geometries are shown in Figure 13, and representative snapshots of the velocity magnitude in Figure 14.
Figure 13. Flow past an elliptic obstacle. Top: the computational domain together with the boundary conditions, the reference cylinder, and the deformation region (, dotted). Bottom: four of the test configurations, including the elliptic obstacle (red), the reference circle (dashed), and a Cartesian grid pushed forward by the deformation map. To improve visibility, the deviation of the ellipse from the circle, , is magnified five times in the drawings; the actual semi-axes are listed above each panel.
Figure 14. Flow past an elliptic obstacle, geometric parametrisation: full-order velocity magnitude on the reference configuration for three values of (rows) at three time instants (columns), with time measured from the first stored snapshot; top: close-up of the near-wake region, bottom: full domain.
The results are presented in Figure 15 and Figure 16 and Table 2. This set-up is more difficult than the previous one since the geometry changes the near-wake topology and the shape of the obstacle influences both the shedding frequency and phase; the errors of SPOD-GPR are about four times greater than in the Reynolds case (those of POD-GPR are comparable) and are mostly concentrated in the near wake located downstream of the obstacle. Both models slightly underestimate the velocity magnitude (by around 2%). SPOD-GPR reduces the NMAE and NRMSE by approximately 5% and 4% respectively. As in the Reynolds case, the POD-GPR uncertainty is highly overconfident (with a coverage of 4%, NSD/NRMSE ), while the SPOD-GPR interval widens as the forecast horizon increases, as is expected for an autoregressive extrapolation, and is conservative (half-width 0.33 against an NRMSE of 0.058, i.e. NSD/NRMSE ), with an average coverage of 0.99; by the end of the test window, however, the half-width approaches half of the data range, so that the band becomes only weakly informative.

Figure 15. Flow past an elliptic obstacle, geometric parametrisation, test window. On the left: the NMAE (solid), the NRMSE (dashed) and the predicted normalised 95% half-width (dotted) for SPOD-GPR (orange) and POD-GPR (blue); for a calibrated Gaussian predictive distribution, the half-width is about 1.96 times the NRMSE. In the centre: the space average of the mean velocity magnitude , with the reference solution in black. On the right: the empirical coverage of the 95% predictive interval.
Figure 16. Flow past an elliptic obstacle, geometric parametrisation: the time-averaged maps over the test window and test parameters, presented on the reference configuration; NSD denotes the normalised predicted standard deviation and COV the empirical 95% coverage. The colour scale shown below each row is shared by the maps in that row (and, for the first row, by the top map).
Discussion
Accuracy. On the four benchmark problems, SPOD-GPR reduces the NMAE by 5–76% and the NRMSE by 4–72% compared to POD-GPR. The gains are largest for the Reynolds-parametrised cylinder flow and smallest for the elliptic obstacle. In the periodic Kuramoto–Sivashinsky scenario, the benefit is mainly qualitative: the time-domain autoregressive model of POD-GPR suppresses the oscillations after just a few steps, while the spectral formulation—where a periodic waveform is represented by slowly varying coefficients at a small number of fixed frequencies—keeps the oscillations going for the whole prediction period. This is in agreement with the mechanism described in § 3.1: a single oscillatory scalar coefficient is not autonomous, but its windowed Fourier coefficients are. It should be noted that, with the set-up in Table 1, the dynamics GP in SPOD-GPR has a latent dimension of (6 and 160 in the two KS cases and 12 for the flows) as opposed to for POD-GPR (2, 6, and 2), so the spectral approach also makes use of a higher-dimensional latent state. Nevertheless, the comparison is still valid since the latent dimension of POD-GPR cannot be increased without limit: the POD coefficients at higher indices become more irregular in time and hence harder to predict.
Uncertainty. The key difference between the two methods is the way in which they quantify uncertainty. In both flow cases, the prediction bands from POD-GPR are between about one and two orders of magnitude narrower than the actual error and cover only 0.6–4% of the test samples. A possible explanation is as follows. In both methods, the variance of the dynamics GP is calibrated on the leave-one-parameter-out residuals of single-step increments; in the time domain, the changes between consecutive snapshots are very small and therefore easy to predict, so the calibration fails to account for the accumulation of phase error over a long rollout. On the other hand, in the spectral case each autoregressive step moves the entire windowed block to the next position; the coefficients change more slowly and smoothly, and the same calibration more accurately represents the true extrapolation error. Moreover, the uncertainty associated with the SPOD modes—those that contain phase information—is passed on to the reconstructed field. As a result, SPOD-GPR produces intervals that range from moderately overconfident (cylinder, 72%) to conservative (elliptic obstacle, 99%), with spread–skill ratios between 0.5 and 2.9 (against and for POD-GPR), and exhibit a spatial pattern that matches the error. With regard to the KS equation, SPOD-GPR is only slightly overconfident in the travelling-wave regime (88% coverage, with intervals about half the width of those obtained from POD-GPR) and more overconfident in the periodic regime (77%), where POD-GPR is instead conservative (99%) but with intervals about five times wider than those of SPOD-GPR. Therefore, improving the calibration of the dynamics GP for long, near-periodic rollouts should be a major focus of future work.
Limitations. The test windows employed in this study are short compared with the training window (5% of the snapshots), and the ranges of parameters investigated are only moderately wide. For longer prediction horizons, the autoregressive rollout will eventually drift in phase and the posterior mean will tend to revert to the mean flow, as is already indicated by the amplitude underestimation observed in all of the KS tests. The method also relies on a number of hyperparameters (the POD rank , the window length , the number of frequencies , the SPOD rank , the kernel length scales), which have been selected manually in this case. Although using a sliding window with gives the highest number of training pairs for the dynamics GP, it also results in a considerable amount of redundancy and thus increases the computational cost of the Nyström regression. Finally, the separate SVDs and Procrustes alignments for each parameter assume that the dominant subspaces change smoothly as the parameter varies, an assumption which can fail near or across bifurcations.
Conclusions
We have developed SPOD-GPR, a non-intrusive, parametric, probabilistic spatio-temporal reduced-order model which makes use of the frequency structure of the data: the regression is performed on frequency-resolved SPOD coefficients rather than on time-domain POD coefficients. SPOD is motivated by augmenting the state with time delays: applying a Fourier transform along the delay axis separates the temporal scales that would otherwise be mixed up by block POD, and as a result temporally coherent modes are obtained, the coefficients of which are easier to learn. A first step of POD reduction is carried out in order to make the subsequent spectral analysis computationally feasible. The parameter dependence of all the bases is then learned on a Grassmann manifold chart following a Procrustes alignment, thus making sure that the predicted modes stay exactly orthonormal. The modal coefficients are obtained by means of a Gaussian process that maps parameters to trajectories and a parameter-conditioned autoregressive Gaussian process, and the uncertainty is passed on to the physical field through pathwise sampling of all the GPs.
The method was tested on four benchmark problems: the two-dimensional Kuramoto–Sivashinsky equation in both periodic and travelling-wave regimes, flow past a cylinder parametrised by the Reynolds number, and flow past an elliptic obstacle parametrised by its semi-axes. In all of them, for unseen parameters and future times, it improved on the accuracy of a POD-GPR model using the same regression architecture. Most importantly, for the flow cases it also produced uncertainty estimates that were far better calibrated than those of POD-GPR, which was strongly overconfident. For the periodic Kuramoto–Sivashinsky case, SPOD-GPR kept the temporal oscillations that POD-GPR failed to capture, although its intervals were overconfident, while those of POD-GPR were conservative but much wider. In the travelling-wave regime, instead, the predicted standard deviation of SPOD-GPR closely matched its actual error, whereas POD-GPR was conservative. Future work will concentrate on the automated tuning of the spectral hyperparameters, extending the forecast horizons, improving the calibration of the dynamics GP for long, near-periodic rollouts, dealing with higher-dimensional parameter spaces, and making use of the calibrated uncertainty in uncertainty-aware ROMs and in real-time digital-twin applications.
References
- [1]Jan S Hesthaven, Gianluigi Rozza, and Benjamin Stamm. Certified Reduced Basis Methods for Parametrized Partial Differential Equations. SpringerBriefs in Mathematics. Springer International Publishing, Cham, 2016. ISBN 9783319224695, 9783319224701. doi: 10.1007/978-3-319-22470-1. URL https://link.springer.com/10.1007/978-3-319-22470-1.DOI
- [2]Guglielmo Padula, Michele Girfoglio, and Gianluigi Rozza. A brief review of reduced order models using intrusive and non-intrusive techniques. PAMM, 24(4), December 2024. ISSN 1617-7061. doi: 10.1002/pamm.202400210.DOI
- [3]Mengwu Guo and Jan S. Hesthaven. Reduced order modeling for nonlinear structural analysis using Gaussian process regression. Computer Methods in Applied Mechanics and Engineering, 341:807–826, November 2018. ISSN 00457825. doi: 10.1016/j.cma.2018.07.017. URL https://linkinghub.elsevier.com/retrieve/pii/S0045782518303487.DOI
- [4]Mengwu Guo and Jan S. Hesthaven. Data-driven reduced order modeling for time-dependent problems. Computer Methods in Applied Mechanics and Engineering, 345:75–99, March 2019. ISSN 00457825. doi: 10.1016/j.cma.2018.10.029. URL https://linkinghub.elsevier.com/retrieve/pii/S0045782518305334.DOI
- [5]Stefania Fresca and Andrea Manzoni. POD-DL-ROM: Enhancing deep learning-based reduced order models for nonlinear parametrized PDEs by proper orthogonal decomposition. Computer Methods in Applied Mechanics and Engineering, 388:114181, January 2022. ISSN 00457825. doi: 10.1016/j.cma.2021.114181. URL https://linkinghub.elsevier.com/retrieve/pii/S0045782521005120.DOI
- [6]John L. Lumley. Stochastic Tools in Turbulence. Academic Press, New York, 1970.
- [7]G Berkooz, P Holmes, and J L Lumley. The Proper Orthogonal Decomposition in the Analysis of Turbulent Flows. Annual Review of Fluid Mechanics, 25(1):539–575, January 1993. ISSN 0066-4189, 1545-4479. doi: 10.1146/annurev.fl.25.010193.002543. URL https://www.annualreviews.org/doi/10.1146/annurev.fl.25.010193.002543.DOI
- [8]Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning. The MIT Press, November 2005. ISBN 9780262256834. doi: 10.7551/mitpress/3206.001.0001. URL https://direct.mit.edu/books/book/2320/Gaussian-Processes-for-Machine-Learning.DOI
- [9]Holger Wendland. Scattered Data Approximation. Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press, Cambridge, 2004. ISBN 9780521843355. doi: 10.1017/CBO9780511617539. URL https://www.cambridge.org/core/books/scattered-data-approximation/980EEC9DBC4CA7 11D089187818135E3.DOI
- [10]Motonobu Kanagawa, Philipp Hennig, Dino Sejdinovic, and Bharath K. Sriperumbudur. Gaussian Processes and Kernel Methods: A Review on Connections and Equivalences, July 2018. URL http://arxiv.org/abs/1807.02582. arXiv:1807.02582.
- [11]Charles A. Micchelli, Yuesheng Xu, and Haizhang Zhang. Universal Kernels. Journal of Machine Learning Research, 7(95):2651–2667, 2006. ISSN 1533-7928. URL http://jmlr.org/papers/v7/micchelli06a.html.
- [12]Emilio Tanowe Maddalena, Paul Scharnhorst, and Colin N. Jones. Deterministic error bounds for kernel-based learning techniques under bounded noise. Automatica, 134:109896, December 2021. ISSN 00051098. doi: 10.1016/j.automatica.2021.109896. URL https://linkinghub.elsevier.com/retrieve/pii/S0005109821004192.DOI
- [13]J.M. Wang, D.J. Fleet, and A. Hertzmann. Gaussian Process Dynamical Models for Human Motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):283–298, February 2008. ISSN 0162-8828. doi: 10.1109/TPAMI.2007.1167. URL http://ieeexplore.ieee.org/document/4359316/.DOI
- [14]Marc Peter Deisenroth, Dieter Fox, and Carl Edward Rasmussen. Gaussian Processes for Data-Efficient Learning in Robotics and Control. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(2):408–423, February 2015. ISSN 0162-8828, 2160-9292. doi: 10.1109/TPAMI.2013.218. URL http://ieeexplore.ieee.org/document/6654139/.DOI
- [15]Markus Heinonen, Cagatay Yildiz, Henrik Mannerström, Jukka Intosalmi, and Harri Lähdesmäki. Learning unknown ODE models with Gaussian processes, 2018. URL https://arxiv.org/abs/1803.04303.
- [16]Katharina Ensinger, Nicholas Tagliapietra, Sebastian Ziesche, and Sebastian Trimpe. Exact Inference for Continuous-Time Gaussian Process Dynamics. Proceedings of the AAAI Conference on Artificial Intelligence, 38(11):11883–11891, March 2024. ISSN 2374-3468, 2159-5399. doi: 10.1609/aaai.v38i11.29074. URL https://ojs.aaai.org/index.php/AAAI/article/view/29074.DOI
- [17]Guglielmo Padula, Michele Girfoglio, and Gianluigi Rozza. A Quadratic Order Reduction – Gaussian Process Ordinary Differential Equation framework for the inference of Large Continuous Dynamical Systems, June 2026. URL http://arxiv.org/abs/2606.13063. arXiv:2606.13063 [math.NA].
- [18]Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Machine learning of linear differential equations using Gaussian processes. Journal of Computational Physics, 348: 683–693, November 2017. ISSN 00219991. doi: 10.1016/j.jcp.2017.07.050. URL https://linkinghub.elsevier.com/retrieve/pii/S0021999117305582.DOI
- [19]Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Numerical Gaussian Processes for Time-Dependent and Nonlinear Partial Differential Equations. SIAM Journal on Scientific Computing, 40(1):A172–A198, January 2018. ISSN 1064-8275, 1095-7197. doi: 10.1137/17M1120762. URL https://epubs.siam.org/doi/10.1137/17M1120762.DOI
- [20]Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M. Stuart. Solving and learning nonlinear PDEs with Gaussian processes. Journal of Computational Physics, 447: 110668, December 2021. ISSN 00219991. doi: 10.1016/j.jcp.2021.110668. URL https://linkinghub.elsevier.com/retrieve/pii/S0021999121005635.DOI
- [21]Houman Owhadi and Gene Ryan Yoo. Kernel Flows: From learning kernels from data into the abyss. Journal of Computational Physics, 389:22–47, July 2019. ISSN 00219991. doi: 10.1016/j.jcp.2019.03.040. URL https://linkinghub.elsevier.com/retrieve/pii/S0021999119302232.DOI
- [22]Boumediene Hamzi and Houman Owhadi. Learning dynamical systems from data: A simple cross-validation perspective, part I: Parametric kernel flows. Physica D: Nonlinear Phenomena, 421:132817, July 2021. ISSN 01672789. doi: 10.1016/j.physd.2020.132817. URL https://linkinghub.elsevier.com/retrieve/pii/S0167278920308186.DOI
- [23]Matthew O. Williams, Ioannis G. Kevrekidis, and Clarence W. Rowley. A Data–Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition. Journal of Nonlinear Science, 25(6):1307–1346, December 2015. ISSN 0938-8974, 1432-1467. doi: 10.1007/s00332-015-9258-5. URL http://link.springer.com/10.1007/s00332-015-9258-5.DOI
- [24]Diya Sashidhar and J. Nathan Kutz. Bagging, optimized dynamic mode decomposition for robust, stable forecasting with spatial and temporal uncertainty quantification. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 380 (2229):20210199, August 2022. ISSN 1364-503X, 1471-2962. doi: 10.1098/rsta.2021.0199. URL https://royalsocietypublishing.org/doi/10.1098/rsta.2021.0199.DOI
- [25]Peter J. Baddoo, Benjamin Herrmann, Beverley J. McKeon, and Steven L. Brunton. Kernel learning for robust dynamic mode decomposition: linear and nonlinear disambiguation optimization. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 478(2260):20210830, April 2022. ISSN 1364-5021, 1471-2946. doi: 10.1098/rspa.2021.0830. URL https://royalsocietypublishing.org/doi/10.1098/rspa.2021.0830.DOI
- [26]Francesco Andreuzzi, Nicola Demo, and Gianluigi Rozza. A Dynamic Mode Decomposition Extension for the Forecasting of Parametric Dynamical Systems. SIAM Journal on Applied Dynamical Systems, 22(3):2432–2458, September 2023. ISSN 1536-0040. doi: 10.1137/22M1481658. URL https://epubs.siam.org/doi/10.1137/22M1481658.DOI
- [27]Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, April 2016. ISSN 0027-8424, 1091-6490. doi: 10.1073/pnas.1517384113. URL https://pnas.org/doi/full/10.1073/pnas.1517384113.DOI
- [28]H. K. Moffatt. Stochastic Tools in Turbulence, by John L. Lumley (Academic Press, 1970) [book review]. Journal of Fluid Mechanics, 67(2):413–415, January 1975. ISSN 0022-1120, 1469-7645. doi: 10.1017/S0022112075210389. URL https://www.cambridge.org/core/product/identifier/S0022112075210389/type/journal_article.
- [29]Aaron Towne, Oliver T. Schmidt, and Tim Colonius. Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis. Journal of Fluid Mechanics, 847:821–867, July 2018. ISSN 0022-1120, 1469-7645. doi: 10.1017/jfm.2018.283. URL https://www.cambridge.org/core/product/identifier/S0022112018002835/type/journal_article.DOI
- [30]Oliver T. Schmidt and Tim Colonius. Guide to Spectral Proper Orthogonal Decomposition. AIAA Journal, 58(3):1023–1033, March 2020. ISSN 0001-1452, 1533-385X. doi: 10.2514/1.J058809. URL https://arc.aiaa.org/doi/10.2514/1.J058809.DOI
- [31]Oliver T. Schmidt, Aaron Towne, Georgios Rigas, Tim Colonius, and Guillaume A. Brès. Spectral analysis of jet turbulence. Journal of Fluid Mechanics, 855:953–982, November 2018. ISSN 0022-1120, 1469-7645. doi: 10.1017/jfm.2018.675. URL https://www.cambridge.org/core/product/identifier/S0022112018006754/type/journal_article.DOI
- [32]Leandra I. Abreu, André V. G. Cavalieri, Philipp Schlatter, Ricardo Vinuesa, and Dan S. Henningson. Spectral proper orthogonal decomposition and resolvent analysis of near-wall coherent structures in turbulent pipe flows. Journal of Fluid Mechanics, 900:A11, October 2020. ISSN 0022-1120, 1469-7645. doi: 10.1017/jfm.2020.445. URL https://www.cambridge.org/core/product/identifier/S0022112020004450/type/journal_article.DOI
- [33]Antonio Colanera, Alessandro Della Pia, Matteo Chiatto, Luigi De Luca, and Francesco Grasso. Modal decomposition analysis of unsteady viscous liquid sheet flows. Physics of Fluids, 33(9):092117, September 2021. ISSN 1070-6631, 1089-7666. doi: 10.1063/5.0056583. URL https://pubs.aip.org/aip/pof/article/33/9/092117/1064407/Modal-decomposition-analysis-of-unsteady-viscous.DOI
- [34]Oliver T. Schmidt and Aaron Towne. An efficient streaming algorithm for spectral proper orthogonal decomposition. Computer Physics Communications, 237:98–109, April 2019. ISSN 00104655. doi: 10.1016/j.cpc.2018.11.009. URL https://linkinghub.elsevier.com/retrieve/pii/S0010465518304016.DOI
- [35]Oliver T. Schmidt. Spectral proper orthogonal decomposition using multitaper estimates. Theoretical and Computational Fluid Dynamics, 36(5):741–754, October 2022. ISSN 0935-4964, 1432-2250. doi: 10.1007/s00162-022-00626-x. URL https://link.springer.com/10.1007/s00162-022-00626-x.DOI
- [36]Akhil Nekkanti and Oliver T. Schmidt. Gappy spectral proper orthogonal decomposition. Journal of Computational Physics, 478:111950, April 2023. ISSN 00219991. doi: 10.1016/j.jcp.2023.111950. URL https://linkinghub.elsevier.com/retrieve/pii/S0021999123000451.DOI
- [37]Antonio Colanera, Oliver T. Schmidt, and Matteo Chiatto. Robust spectral proper orthogonal decomposition. Computer Physics Communications, 307:109432, February 2025. ISSN 00104655. doi: 10.1016/j.cpc.2024.109432. URL https://linkinghub.elsevier.com/retrieve/pii/S0010465524003552.DOI
- [38]Akhil Nekkanti and Oliver T. Schmidt. Frequency–time analysis, low-rank reconstruction and denoising of turbulent flows using SPOD. Journal of Fluid Mechanics, 926:A26, November 2021. ISSN 0022-1120, 1469-7645. doi: 10.1017/jfm.2021.681. URL https://www.cambridge.org/core/product/identifier/S0022112021006819/type/journal_article.DOI
- [39]Moritz Sieber, C. Oliver Paschereit, and Kilian Oberleithner. Spectral proper orthogonal decomposition. Journal of Fluid Mechanics, 792:798–828, April 2016. ISSN 0022-1120, 1469-7645. doi: 10.1017/jfm.2016.103. URL https://www.cambridge.org/core/product/identifier/S0022112016001038/type/journal_article.DOI
- [40]Andrea Lario, Romit Maulik, Oliver T. Schmidt, Gianluigi Rozza, and Gianmarco Mengaldo. Neural-network learning of SPOD latent dynamics. Journal of Computational Physics, 468:111475, November 2022. ISSN 00219991. doi: 10.1016/j.jcp.2022.111475. URL https://linkinghub.elsevier.com/retrieve/pii/S002199912200537X.DOI
- [41]Alan Edelman, Tomás A. Arias, and Steven T. Smith. The Geometry of Algorithms with Orthonormality Constraints. SIAM Journal on Matrix Analysis and Applications, 20(2):303–353, June 1998. ISSN 0895-4798, 1095-7162. doi: 10.1137/S0895479895290954. URL https://epubs.siam.org/doi/10.1137/S0895479895290954.DOI
- [42]James T. Wilson, Viacheslav Borovitskiy, Alexander Terenin, Peter Mostowsky, and Marc Peter Deisenroth. Efficient Sampling Functions from Gaussian Process Posteriors. 2020. doi: 10.48550/ARXIV.2002.09309. URL https://arxiv.org/abs/2002.09309.
- [43]Floris Takens. Detecting strange attractors in turbulence. In David Rand and Lai-Sang Young, editors, Dynamical Systems and Turbulence, Warwick 1980, volume 898, pages 366–381. Springer Berlin Heidelberg, Berlin, Heidelberg, 1981. ISBN 9783540111719 9783540389453. doi: 10.1007/BFb0091924. URL http://link.springer.com/10.1007/BFb0091924.DOI
- [44]H. Karcher. Riemannian center of mass and mollifier smoothing. Communications on Pure and Applied Mathematics, 30(5):509–541, September 1977. ISSN 0010-3640, 1097-0312. doi: 10.1002/cpa.3160300502. URL https://onlinelibrary.wiley.com/doi/10.1002/cpa.3160300502.DOI
- [45]P.-A. Absil, R. Mahony, and R. Sepulchre. Riemannian Geometry of Grassmann Manifolds with a View on Algorithmic Computation. Acta Applicandae Mathematicae, 80(2):199–220, January 2004. ISSN 0167-8019, 1572-9036. doi: 10.1023/B:ACAP.0000013855.14971.91. URL https://link.springer.com/10.1023/B:ACAP.0000013855.14971.91.DOI
- [46]David Amsallem and Charbel Farhat. Interpolation Method for Adapting Reduced-Order Models and Application to Aeroelasticity. AIAA Journal, 46(7):1803–1813, July 2008. ISSN 0001-1452, 1533-385X. doi: 10.2514/1.35374. URL https://arc.aiaa.org/doi/10.2514/1.35374.DOI
- [47]Rasmus Jensen and Ralf Zimmermann. Canonically centered coordinates for Grassmann interpolation: Lagrange, Hermite, and errors. BIT Numerical Mathematics, 66(2):27, April 2026. ISSN 1572-9125. doi: 10.1007/s10543-026-01123-x. URL https://doi.org/10.1007/s10543-026-01123-x.
- [48]S. Goreinov and E. Tyrtyshnikov. The maximal-volume concept in approximation by low-rank matrices. In Vadim Olshevsky, editor, Contemporary Mathematics, volume 280, pages 47–51. American Mathematical Society, Providence, Rhode Island, 2001. ISBN 9780821819210 9780821856161 9780821878705. doi: 10.1090/conm/280/4620. URL https://www.ams.org/comn/280.DOI
- [49]S. A. Goreinov, I. V. Oseledets, D. V. Savostyanov, E. E. Tyrtyshnikov, and N. L. Zamarashkin. How to Find a Good Submatrix. In Matrix Methods: Theory, Algorithms and Applications, pages 247–256. WORLD SCIENTIFIC, April 2010. ISBN 9789812836014 9789812836021. doi: 10.1142/9789812836021_0015. URL http://www.worldscientific.com/doi/abs/10.1142/9789812836021_0015.DOI
- [50]Christopher Williams and Matthias Seeger. Using the Nyström Method to Speed Up Kernel Machines. In Advances in Neural Information Processing Systems, volume 13. MIT Press, 2000. URL https://papers.nips.cc/paper_files/paper/2000/hash/19de10adbaa1b2ee13f77f679fa1483a-Abstract.html.
- [51]A. Kalogirou, E. E. Keaveny, and D. T. Papageorgiou. An in-depth numerical study of the two-dimensional Kuramoto–Sivashinsky equation. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 471(2179):20140932, July 2015. ISSN 1364-5021, 1471-2946. doi: 10.1098/rspa.2014.0932. URL http://royalsocietypublishing.org/rspa/article/57634.DOI
- [52]Igor A. Baratta, Joseph P. Dean, Jørgen S. Dokken, Michal Habera, Jack S. Hale, Chris N. Richardson, Marie E. Rognes, Matthew W. Scroggs, Nathan ORCID icon Sime, and Garth N. Wells. DOLFINx: The next generation FEniCS problem solving environment. DOLFINx: The next generation FEniCS problem solving environment., 2023. ISSN 0098-3500. doi: 10.1145/3524456.












