Reduced-order models (ROMs) learned from trusted numerical simulations enable the real-time simulation of complex physical systems, making computationally prohibitive tasks-such as uncertainty quantification and optimal control-tractable. In particular, data-driven ROMs are opening the opportunity to learn from both numerical simulations and experimental data. However, integrating these sources of information is far from straightforward. Real-world measurements often involve multiple sources of uncertainty, including non-Gaussian noise, latent heterogeneities, unknown or hidden system parameters, and multiscale phenomena occurring below the resolution of the data. Consequently, real-world data are better described by parametrized stochastic partial differential equations, for which the development of effective ROMs remains relatively unexplored. In this work, we introduce POD-CNF, a data-driven generative ROM for efficient uncertainty quantification of parametrized high-dimensional systems. The approach combines projection-based model reduction and deep generative modeling. First, Proper Orthogonal Decomposition (POD) is employed to obtain a low-dimensional representation of the solution manifold. A Conditional Normalizing Flow (CNF) is then trained to approximate the density of the reduced coefficients conditionally on the input parameters. This formulation allows the surrogate to capture complex, multimodal output distributions, going beyond standard regression-based ROMs and least-squares predictions. We assess the effectiveness of POD-CNF through numerical experiments, demonstrating its ability to accurately reproduce output distributions while substantially reducing computational costs. Finally, a proof-of-concept application illustrates how the generative paradigm facilitates integration into Bayesian inverse problems, highlighting promising directions for future research.
Figures & tables
Figure 1: Problem domain for the advection-diffusion problem (left) and corresponding discretization (right).
Figure 2: Advection diffusion problem. Two random solutions associated with the same operating conditions μ=[−1,2,−5] . The variability is induced by the hidden parameters, θ , whose values are assumed to be unknown (to the user, and to the surrogate).
Figure 3: Advection-diffusion problem, § 4.2 . Mean relative reconstruction error (solid blue line) and cumulative energy (dashed black line) versus the number of POD bases n . Truncating at n=20 guarantees a projection error in euclidean norm equal to 1.05% .
Figure 4: Advection-diffusion problem, § 4.2 . Solutions fields across different parametric configurations: (a) μ0=[−1,4,−5] , (b) μ1=[5,5,5] , and (c) μ2=[−5,0,−5] . For each row, the leftmost panel shows a ground-truth snapshot, computed with the numerical solver, while the three panels to its right show independent realizations sampled from the POD-CNF surrogate model. Configuration (a) corresponds to the parameter μ0∈P chosen for the local analysis.
Figure 5: Reduced log-likelihood sensitivity of the POD-CNF surrogate evaluated at μ0 against solutions belonging to different parameter instances μk .
Figure 6: Advection-diffusion problem, § 4.2 . Comparison between FOM and POD-CNF marginal densities (bottom panel) at eight selected monitoring points (top panel).
Figure 7: Advection-diffusion problem, § 4.2 . Global analysis over 50 unseen test configurations. The boxplots show the distribution of the Wasserstein-1 error over the considered test set, comparing FOM and surrogate outputs. From left to right: POD-CNF (red, mean =0.5673 ), POD-GP (green, mean =1.7970 ) and deterministic baseline (blue, mean =1.4369 ).
W~1 at μ0
Average W~1 on P
NLLtest
Training time
Inference time
Deterministic baseline
1.5284
1.4369 ± 0.1103
–
–
–
POD-GP
2.3819
1.7970 ± 0.1605
4.5315 ± 0.9112
54s
0.29s
POD-CNF
0.6735
0.5691 ± 0.0997
-20.7059 ± 7.6416
10m 33s
0.06s
Table 1: Advection-diffusion problem, § 4.2 . Performance metrics and training times for POD-CNF, POD-GP and the deterministic baseline (cf. Section 4.1 ). First column: surrogate under study. Second column: empirical Wasserstein distance between FOM-generated and surrogate-generated samples ( mgen=1000 ). Third column: average Wasserstein distance, computed across 50 random parametric configurations ( mgen=200) , and corresponding standard deviation., Fourth column: negative loglikelihood, averaged across the whole test set (320 joint samples in μ,θ ), reported with a 95% confidence interval. Fifth column: fitting time, excluding hyperparameter tuning. Last column: sequential time required to generate a single scenario for a given parametric configuration.
% Data
Ntrain
Nval
Average W~1 on P
W~1 at μ0
100%
4800
1280
0.5691 ± 0.0997
0.6735
75%
3600
960
0.5102 ± 0.0872
0.6675
50%
2400
640
0.5487 ± 0.0943
0.8187
25%
1200
320
0.6607 ± 0.1287
0.9352
10%
480
128
1.5019 ± 0.4024
1.8266
5%
240
64
2.0273 ± 0.4023
2.3104
Table 2: Advection-diffusion problem, § 4.2 . Sensitivity of POD-CNF performance to training dataset size. Evaluation metrics read as in Table 1 .
Figure 8: Linear elasticity problem, § 4.3 . Physical domain (left) and corresponding discretization (right).
Figure 9: Linear elasticity problem, § 4.3 . Two random solutions associated with the same operating conditions μ=[1.0,0.25] . The variability is induced by the hidden parameters, θ , whose values are assumed to be unknown (to the user, and to the surrogate).
Figure 10: Linear elasticity problem, § 4.3 . POD truncation analysis: mean relative reconstruction error on the validation set (blue solid line) and cumulative retained energy (dashed black line) as a function of the number of POD modes n .
Figure 11: Linear elasticity problem, § 4.3 . Solutions fields across different parametric configurations: (a) μ0=[1.5,0.20] , (b) μ1=[2.0,0.25] , and (c) μ2=[2.0,0.25] . For each row, the leftmost panel shows a ground-truth snapshot, computed with the numerical solver, while the three panels to its right show independent realizations sampled from the POD-CNF surrogate model. Configuration (a) corresponds to the parameter μ0∈P chosen for the local analysis.
Figure 12: Reduced log-likelihood sensitivity of the POD-CNF surrogate evaluated at μ0 against solutions belonging to different parameter instances μk .
Figure 13: Linear elasticity problem, § 4.3 . Top left: computational mesh and locations of the four monitoring points (P1–P4). Bottom and right: joint probability density contours of the horizontal and vertical displacements (ux,uy) at P1–P4 for μ0 , comparing the FOM reference with the POD-CNF generated distribution (left and right of each subplot, respectively).
Figure 14: Linear elasticity problem, § 4.3 . Global analysis over 50 unseen test configurations. The boxplots show the distribution of the Wasserstein-1 error over the considered test set, comparing FOM and surrogate outputs. From left to right: POD-CNF (red, mean =0.5117 ), POD-GP (green, mean =1.8402 ) and deterministic baseline (blue, mean =2.6443 ).
W~1 at μ0
Average W~1 on P
NLLtest
Training time
Inference time
Deterministic baseline
2.8098
2.6443 ± 0.3308
–
–
–
POD-GP
1.6707
1.8402 ± 0.3343
0 -27.7385 ± 0.4076
55.57s
0.29s
POD-CNF
0.4561
0.5117 ± 0.1259
-101.9474 ± 2.3082
12m 24s
0.08s
Table 3: Linear elasticity problem, § 4.3 . Quantitative analysis comparing POD-CNF, POD-GP and the deterministic baseline, cf. § 4.1 . Table entries read as in Table 1 .
% Data
Ntrain
Nval
Average W~1 on P
W~1 at μ0
100%
4800
1280
0.5117 ± 0.1259
0.4561
75%
3600
960
0.5110 ± 0.1624
0.4135
50%
2400
640
0.5062 ± 0.1327
0.4936
25%
1200
320
0.6761 ± 0.1799
0.6768
10%
480
128
1.2932 ± 0.2403
1.4152
5%
240
64
2.3952 ± 0.2498
1.3998
Table 4: Linear elasticity problem, § 4.3 . Sensitivity of POD-CNF accuracy to training sample size. Table entries read as in Table 2 .
Label & QoI & Expression & Continuous counterpart
Q1 & Average Concentration Gradient & 1Nc∑j=1cN∥∇u(j)μ∥2 & 1∣Ω∣∫Ω∥∇uμ∥dx
Q2 & Spatial Standard Deviation & [1Nh∑j=1hN∣u(j)μ−uμ∣2]1/2 & [1∣Ω∣∫Ω∣uμ−uμ∣2dx]1/2
Table 5: QoIs for the Forward UQ analysis. ∣Ω∣= area of the spatial domain. Nh= number of vertices in the mesh. Nc= number of cells in the mesh. u(j)=jth component of uμ. ∇u(j)= value of the gradient of the FEM interpolant at the barycenter of the jth cell. uμ= the nodal average (resp. integral) average of the concentration field (third and fourth columns, respectively). Γ_c_i and I_c_i = continuous boundary and discrete degrees of freedom associated with the i-th cylinder obstacle, respectively. Ωright=x∈Ω:x1>2.25. Iright= degrees of freedom in the mesh belonging to Ωright.
Figure 15: Comparison of the 99 confidence intervals obtained using the true FOM solver and the proposed POD-CNF surrogate for the evaluated Quantities of Interest (QoIs).
QoI & Est. Mean & 99% CI & N∗ & Time ∗ (s)
Q1 & 0.6360 & 0.0138 & 472 & 0.8349
Q2 & 0.2061 & 0.0040 & 376 & 0.6651
Q3 & 0.3275 & 0.0073 & 501 & 0.8862
Q4 & 0.1553 & 0.0248 & 25500 & 45.1065
Q5 & 0.5700 & 0.1310 & 52564 & 92.9796
Table 6: Summary of the Forward UQ analysis for POD-CNF (left) and FOM (right). Est. Mean = value of the Monte Carlo estimator with N=100 samples. 99% CI = semi-amplitude of the 99% confidence interval with N=100 Monte Carlo samples. N∗= estimated sample size required to reduce the relative uncertainty below 1%. Time∗ = running time (in seconds) required by the model to generate N∗ samples.
Figure 16: Linear elasticity problem, § 4.3 . For both panels, the first row reports the MCMC traces and posterior distribution p(μ∣usens) from the high-fidelity numerical solver (FOM baseline), while the second row shows the corresponding results obtained via the POD-CNF surrogate.
True m & Est. m & 95% CI & Error & & True δ & Est. δ & 95% CI & Error & Time
Non-intrusive reduced-order models (NIROMs) have become a standard tool for approximating parametric partial differential equations from computer design of experiments while significantly reducing computational costs. However, assessing the reliability of their predictions remains a major challenge, particularly in extrapolation regimes or under limited training data. In this work, we introduce a framework for quantifying model-form uncertainty in NIROMs by combining a perturbative stochastic representation of reduced bases with distribution-free conformal-type methods. Starting from a deterministic reduced basis constructed from snapshot matrices, we model uncertainty through random perturbations defined on the Stiefel manifold, directed along the discarded modes, yielding stochastic reduced-order approximations whose induced variance reflects the basis-truncation error. A transport approximation gives a closed-form posterior variance that separates basis-induced from regression-induced uncertainty, without re-training the underlying Gaussian processes. We include this posterior variance within a conformal risk control calibration framework, that provides prediction sets with coordinate miscoverage guarantees. The calibration factor produced by this framework is itself an interpretable, scalar diagnostic of the quality of the uncertainty estimate. The methodology is evaluated on parametric PDE benchmarks and an industrial tire-manufacturing calendering process. Numerical experiments demonstrate reliable, locally informative uncertainty quantification that goes beyond the Gaussian predictive variance.
Edgar Jaber, Rémy Vallot, Thibault Dairay +1
CB, ENS Paris Saclay · Université Paris-Saclay, CNRS, ENS Paris-Saclay, Centre Borelli, Gif-sur-Yvette, 91190, France · CB, Michelin +3
Deep Learning Reduced Order Models (ROMs) are becoming increasingly popular as surrogate models for parametric partial differential equations (PDEs) due to their ability to handle high-dimensional data, approximate highly nonlinear mappings, and utilize GPUs. Existing approaches typically learn evolution either on the full solution field, which requires capturing long-range spatial interactions at high computational cost, or on compressed latent representations obtained from autoencoders, which reduces the cost but often yields latent vectors that are difficult to evolve, since they primarily encode spatial information. Moreover, in parametric PDEs, the initial condition alone is not sufficient to determine the trajectory, and most current approaches are not evaluated on jointly predicting multiple solution components with differing magnitudes and parameter sensitivities. To address these challenges, we propose a joint model consisting of a convolutional encoder, a transformer operating on latent representations, and a decoder for reconstruction. The main novelties are joint training with multi-stage parameter injection and coordinate channel injection. Parameters are injected at multiple stages to improve conditioning. Physical coordinates are encoded to provide spatial information. This allows the model to dynamically adapt its computations to the specific PDE parameters governing each system, rather than learning a single fixed response. Experiments on the Advection-Diffusion-Reaction equation and Navier-Stokes flow around the cylinder wake demonstrate that our approach combines the efficiency of latent evolution with the fidelity of full-field models, outperforming DL-ROMs, latent transformers, and plain ViTs in multi-field prediction, reducing the relative rollout error by approximately 5 times.
Iva Mikuš, Boris Muha, Domagoj Vlah
University of Zagreb Faculty of Electrical Engineering and Computer Science · University of Zagreb Faculty of Science Department of Mathematics
Reduced-order models (ROMs) provide an efficient surrogate for complex multiscale systems, but their predictive accuracy is often compromised by truncation errors and the inadequate representation of interactions between resolved and unresolved scales. The missing effect of truncated (unresolved) scales on ROM (resolved) scales is often denoted as the closure problem. In this work, we formulate ROM closure modeling as a multi-fidelity (MF) learning problem and propose an uncertainty-aware MF framework based on conditional normalizing flow to enhance ROM predictive accuracy. The proposed approach learns a probabilistic mapping from low-fidelity (LF) ROM coefficients to high-fidelity (HF) coefficients, thereby improving predictive fidelity while quantifying the uncertainty associated with the learned closure. Two correction strategies are investigated: direct learning, in which HF coefficients are predicted directly from LF inputs, and residual learning, which learns the discrepancy between LF and HF coefficients and uses it to recover the corrected HF solution. The framework is demonstrated on a vortex merging problem governed by the two-dimensional Navier Stokes equations. Results show that both correction strategies improve ROM accuracy over uncorrected ROM, with residual learning achieving consistently better performance than direct learning. Moreover, the two proposed deep generative model-based strategies provide uncertainty quantification for the corrected ROM coefficients, which is critical for assessing prediction confidence and supporting the reliable use of ROMs in practical applications.
Jice Zeng, Shady E. Ahmed, David Barajas-Solano +1
Advanced Computing, Mathematics, and Data Division, Pacific Northwest National Laboratory, Richland, 99352, WA, USA