Stochastic closure modeling is an emerging area in reduced-order modeling of large and complex (e.g., multiscale, multi-physics) systems, for which the computationally affordable reduced-order physical descriptions (e.g., digital twins for additive manufacturing, world models for AI agents) in numerical simulations inevitably lead to information losses that cannot be well addressed by classical deterministic closures. However, learning stochastic closures typically requires closure data that are difficult to obtain from measurements or often prohibitively expensive to compute using first-principles simulations. In this work, we develop a stochastic closure modeling framework that builds on score-based diffusion models, leveraging indirect offline training and online data assimilation without requiring the direct data of closures. Numerical experiments on two-dimensional turbulent flows demonstrate that indirect offline learning captures key closure statistics and improves predictions of system state. In addition, combining the learned closure with known physical dynamics leads to better generalization capability across different regimes than purely data-driven models that directly predict system state. Compared to standard data assimilation methods with a Gaussian assumption, the posterior-score assimilation better represents non-Gaussian features of the learned closure, and its combination with state-level data assimilation further improves prediction of the system state. The proposed framework enables stochastic closure modeling in realistic and challenging scenarios of complex dynamical systems in mechanics and engineering problems, where non-Gaussian statistics of model-form uncertainties are non-negligible (e.g., due to the nonlinearities of the system dynamics and/or the observations) and true closure data are not available, for both offline learning and online assimilation.
Figures & tables
Figure 1: Schematic of the proposed framework. (a) Offline indirect training compares simulated resolved trajectories with temporally sparse reference states through a trajectory energy score loss. (b) Reverse diffusion generates closure fields conditioned on the evolving resolved state. (c) Online posterior-score DA refines closures at the model-form uncertainty level, while complementary state EnKF corrects the resolved prediction. Black curves illustrate reference trajectories, and colored curves and shaded bands represent ensemble predictions and their uncertainty.
Figure 2: Spatial filtering and closure formation: (a) DNS vorticity ω on the 512×512 grid; (b) filtered vorticity ω on the 64×64 resolved grid; (c) the corresponding subfilter-scale closure Π , defined in Eq. ( 37 ); and (d) kinetic energy spectra E(k) of the DNS and FDNS fields. The dashed curve shows the squared filter transfer function. The vertical dotted line marks the LES cutoff at kN=32 .
Model
Single sample
Ensemble mean
Score-D
6.815e-02
4.335e-02
Score-I
3.642e-01
1.517e-01
Table 1: Relative errors of Score-D and Score-I closure predictions against paired FDNS closures at held-out resolved states. Columns report errors for individual closure samples and ensemble means. Bold values indicate the lowest error in each column.
Figure 3: A priori comparison of closure fields at two resolved states. Each row shows, from left to right, the FDNS reference closure Π , the Score-D and Score-I ensemble means, and their corresponding prediction errors relative to FDNS. Each ensemble mean is obtained from 32 closure samples conditioned on the same resolved state.
Figure 4: Spectral comparison of FDNS, Score-D, and Score-I closures: closure power spectrum CΠ(k) , SGS energy transfer TE(k) , and SGS enstrophy transfer TZ(k) , from left to right. Positive and negative transfer values indicate addition to and removal from the resolved energy or enstrophy, respectively.
Figure 5: Joint probability densities of resolved vorticity ω and closure Π for FDNS, Score-D, and Score-I. Densities are estimated from local state–closure pairs collected across spatial points and held-out resolved states, using individual generated closure samples for the learned models. All panels share a common logarithmic density scale.
Figure 6: Resolved vorticity predictions using DSM, Score-D, and Score-I at t=5 and 10s in the training regime. At each time, the vorticity fields are followed by their corresponding prediction errors relative to FDNS. Score-D and Score-I fields show individual ensemble realizations.
t (s)
DSM
Score-D
Score-I
2.5
5.259e-01
5.730e-02
1.079e-01
5.0
8.258e-01
1.288e-01
2.394e-01
7.5
1.070e+00
2.460e-01
4.303e-01
10.0
1.172e+00
3.897e-01
5.816e-01
Table 2: Relative state prediction errors of DSM, Score-D, and Score-I against FDNS in the training regime. Errors for the stochastic models use ensemble-mean predictions. Bold values indicate the lowest error at each time.
Figure 7: A posteriori diagnostics of DSM, Score-D, and Score-I in the training regime: relative state prediction errors, kinetic energy spectra E(k) at t=10s , and RMSE-to-spread ratios, from left to right. Prediction errors use ensemble means for the stochastic models. The dotted horizontal line in the right panel marks equal RMSE and ensemble spread.
Model
RE at t=0.2s
RE at t=0.4s
Score-D
9.889e-03
1.932e-02
Score-I
9.076e-03
1.722e-02
FNO
9.987e-03
1.942e-02
State-DM
1.066e-02
2.043e-02
Table 3: Short-window relative state prediction errors of Score-D, Score-I, FNO, and State-DM in the training regime. The selected models provide comparable accuracy at t=0.2 and 0.4s for the subsequent regime-transfer comparisons.
Figure 8: State prediction error growth in the training regime and under forcing and viscosity shifts. The hybrid models, Score-D and Score-I, are compared with the purely data-driven FNO and State-DM predictors and the unclosed coarse solver (No SGS). Curves show relative errors averaged over test cases, using ensemble-mean predictions for stochastic models.
Model
Training regime
Forcing shift
Viscosity shift
No SGS
1.222e+00
1.071e+00
1.070e+00
Score-D
7.165e-01
8.007e-01
5.823e-01
Score-I
5.816e-01
6.554e-01
3.826e-01
FNO
7.632e-01
1.035e+00
1.264e+00
State-DM
6.888e-01
9.691e-01
9.707e-01
Table 4: Relative state prediction errors against FDNS at t=10s in the training regime and under forcing and viscosity shifts. Errors are averaged over test cases, using ensemble-mean predictions for stochastic models. Bold values indicate the lowest error in each regime.
Figure 9: Resolved vorticity fields at t=10s . Rows correspond to the training regime, forcing shift, and viscosity shift, from top to bottom. Columns compare FDNS with Score-D, Score-I, FNO, and State-DM.
Forcing shift
Viscosity shift
Model
E
Z
E
Z
Score-D
-4.520e-02
+9.900e-02
+2.550e-02
+7.750e-02
Score-I
-4.100e-02
+1.163e-01
+2.400e-02
+3.740e-02
FNO
-1.217e-01
-6.085e-01
+6.394e-01
+7.659e-01
State-DM
-2.401e-01
-6.823e-01
+2.832e-01
+3.502e-01
Table 5: Relative deviations of resolved kinetic energy E and enstrophy Z from FDNS at t=10s under forcing and viscosity shifts. Deviations are evaluated for each test case and then averaged. Bold values indicate the smallest absolute deviation in each column.
Figure 10: Resolved kinetic energy spectra E(k) at t=10s in the training regime and under forcing and viscosity shifts, from left to right. Curves compare the model predictions with FDNS, with spectra averaged over realizations and test cases.
Figure 11: Sequential data assimilation in the training, larger-scale forcing, and smaller-scale forcing regimes, from left to right. Upper panels show relative errors of ensemble-mean state predictions against FDNS; lower panels show the ratio of RMSE to ensemble spread Sωˉ . The dotted horizontal line marks equal RMSE and spread. Sparse vorticity observations are assimilated every 0.05s .
Method
Training regime
Larger-scale forcing
Smaller-scale forcing
Score-I only
8.628e-01
9.528e-01
9.315e-01
Score DA
4.072e-01
4.147e-01
7.648e-01
Closure EnKF
9.214e-01
9.789e-01
1.225e+00
State EnKF
2.156e-01
1.468e-01
4.491e-01
Score DA + state EnKF
1.356e-01
8.683e-02
3.491e-01
Gaussian closure + state EnKF
9.016e-01
1.007e+00
1.147e+00
Table 6: Relative errors of ensemble-mean state predictions against FDNS at t=20s for different DA strategies. Columns compare the training, larger-scale forcing, and smaller-scale forcing regimes. Score-I supplies the closure model except for the Gaussian-closure baseline. Bold values indicate the lowest error in each regime.
Figure 12: Standardized closure probability densities during the early ( 0 – 5s ) and late ( 15 – 20s ) stages of training-regime assimilation. Panels compare FDNS closures with samples from Score-I using posterior-score DA or closure EnKF. The standard Gaussian density is included for comparison with the non-Gaussian closure distributions.
Early: 0 – 5s
Late: 15 – 20s
Method
σΠ
Excess kurtosis
σΠ
Excess kurtosis
FDNS
0.477
2.00
0.453
4.25
Score-I + score DA
0.510
1.52
0.620
5.23
Score-I + closure EnKF
0.505
1.48
0.974
1.52
Table 7: Closure standard deviations σΠ and excess kurtosis for FDNS and Score-I with posterior-score DA or closure EnKF. Statistics are compared over the early ( 0 – 5s ) and late ( 15 – 20s ) stages of training-regime assimilation.
Figure 13: Resolved vorticity predictions and their errors relative to FDNS at t=20s for different DA strategies. Each pair of rows shows the vorticity fields followed by their corresponding errors. Row pairs represent the training, larger-scale forcing, and smaller-scale forcing regimes, from top to bottom. Columns compare FDNS with the five assimilation configurations, using ensemble-mean fields for the predictions.
Figure 14: Comparison of posterior-score DA and its combination with state EnKF at unobserved locations. Columns correspond to the training, larger-scale forcing, and smaller-scale forcing regimes. The first three rows show vorticity time series at three fixed unobserved points; the bottom row shows spatial profiles along an unobserved grid row at t=20s . Colored curves show ensemble means, shaded bands indicate central 95% ensemble intervals, and black curves show FDNS.
Figure 15: Joint probability densities of the normalized forecast velocity components ux and uy at the speed probe. Columns compare the prior, Bayesian reference, posterior-score DA, and closure EnKF. The upper and lower rows correspond to low- and high-speed observations, respectively. Both observation cases use the same Score-I prior, repeated in the first column for comparison.
Figure 16: Prior and posterior marginal densities under nonlinear speed observations. Columns show the forecast velocity components ux and uy and the normalized speed at the probe, from left to right. The upper and lower rows correspond to low- and high-speed observations, respectively. Curves compare the prior, Bayesian reference, posterior-score DA, and closure EnKF; vertical lines mark the observed speeds.
School of Mechanical Engineering Purdue University West Lafayette, IN 47907 · College of Information Sciences and Technology Pennsylvania State University University Park, PA 16802