Stochastic Reduced-Order Method (SROM)

• SROM is a family of reduced-order strategies that embed stochasticity into surrogates, enabling the propagation of probability distributions rather than deterministic trajectories.
• It is applied across diverse contexts such as Fokker–Planck evolution, optimized discrete measures for uncertainty quantification, and adaptive PODs for time-dependent systems.
• SROM methods balance model fidelity with computational efficiency using data-driven parameter estimation, classification techniques, and stochastic operator inference to capture uncertainty.

Searching arXiv for the cited SROM papers to ground the article in current arXiv records. Stochastic Reduced-Order Method (SROM) denotes a family of reduced-order modeling strategies in which stochastic structure is built into the reduced representation so that the surrogate propagates distributions, latent stochastic states, or weighted discrete probability measures rather than only deterministic trajectories. In the recent arXiv literature, the term is explicitly polysemous: it has been used for a Fokker–Planck/Langevin reduction of statistical microstructure descriptors, for non-intrusive uncertainty quantification based on optimized discrete measures, for cluster- and basis-adaptive stochastic ROMs for time-dependent systems, and for data-driven latent stochastic models learned directly from observations (Tran et al., 2020, Altunay et al., 22 Jul 2025, Xiong et al., 2022, Naderi et al., 2022, Freitag et al., 2024, Ilersich et al., 15 Jan 2026).

1. Terminological scope and principal usages

The literature does not use the acronym in a single canonical sense. One strand defines SROM as a stochastic dynamical ROM whose state is a scalar descriptor or latent variable obeying an SDE, often accompanied by a Fokker–Planck equation for its PDF. A second strand uses SROM for a discrete probabilistic approximation of uncertain inputs, replacing a continuous law by a low-order weighted support. A third strand uses the label for stochastic closures, cluster-adaptive POD constructions, or latent-space stochastic surrogates for SPDEs and parametrized dynamical systems (Tran et al., 2020, Altunay et al., 22 Jul 2025, Naderi et al., 2022).

Usage	Reduced object	Representative arXiv id
Fokker–Planck descriptor evolution	PDF of a scalar microstructure descriptor	(Tran et al., 2020)
Optimized discrete measure for UQ	Support points and weights for random inputs	(Altunay et al., 22 Jul 2025)
Cluster-based stochastic ROM	Local POD basis selected by classification	(Xiong et al., 2022)
Time-dependent-basis stochastic ROM	Evolving spatial and stochastic modes	(Naderi et al., 2022)
Data-driven stochastic latent or operator ROM	Reduced SDEs, randomized projections, or stochastic closures	(Freitag et al., 2024, Ilersich et al., 15 Jan 2026, Yong et al., 2024)

This terminological split is not incidental. The microstructure paper explicitly distinguishes its Fokker–Planck-based SROM from the unrelated UQ usage in which SROM denotes optimized discrete probability measures (Tran et al., 2020). The TDB-ROM paper makes the same distinction, noting that classical SROM based on discrete probability representations is different in spirit from methods that reduce the solution field itself (Naderi et al., 2022). This suggests that “SROM” is best understood as a broad label for stochastic reduction principles rather than a single algorithmic template.

2. Fokker–Planck SROM for statistical microstructure descriptors

In integrated computational materials engineering, SROM has been defined as a continuous-time stochastic model for a scalar microstructure descriptor $x(t)$ whose PDF $f(x,t)$ evolves according to a Fokker–Planck equation. The descriptor follows a 1D nonlinear Langevin equation with Gaussian white noise,

$$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$

or, in Itô form,

$dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$

The associated PDF evolves according to

$$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$

Here the state is a scalar stochastic descriptor $x(t)$, the drift is $a(x,t)\equiv D^{(1)}(x,t)$, the diffusion is $D(x,t)\equiv D^{(2)}(x,t)$, and the stochastic forcing is a standard Wiener process (Tran et al., 2020).

The rationale is that microstructures are inherently random and that statistical descriptors such as grain size distributions or chord-length distributions are stochastic by nature and are often better represented as evolving PDFs than as single deterministic trajectories. The paper focuses on one scalar descriptor per ROM instance: grain area for kinetic Monte Carlo grain growth, chord-length distributions for phase-field spinodal decomposition, and total mean-square displacement and enthalpy for molecular dynamics of liquid Ar. It also states that the methodology applies to any scalar descriptor whose PDF can be estimated from simulation data, including descriptors such as two-point correlation $S_2$, lineal-path functions, cluster or particle size distributions, and power spectral density (Tran et al., 2020).

Calibration proceeds from descriptor time series generated by full-order ICME simulations. Two routes are described. The analytical route uses Kramers–Moyal estimators and moment relations when $D^{(1)}$ and $f(x,t)$0 depend only on time, so that linear regression of empirical mean and variance versus time yields the coefficients. The numerical route parameterizes $f(x,t)$1 and $f(x,t)$2 and minimizes a loss based on Kullback–Leibler divergence between predicted and training PDFs; the paper uses batch-parallel Bayesian optimization in the phase-field and MSD cases. Empirical initial PDFs are smoothed by Tikhonov regularization,

$f(x,t)$3

with typical choices $f(x,t)$4 and $f(x,t)$5, and the Fokker–Planck equation is discretized by finite differences on a uniform grid with explicit Runge–Kutta or implicit Crank–Nicolson time stepping (Tran et al., 2020).

The three demonstrations reported in the paper are summarized below.

ICME model	Descriptor	Reported speedup
Kinetic Monte Carlo	Grain area after log transform	$f(x,t)$6
Phase field	Chord-length distribution	$f(x,t)$7
Molecular dynamics	MSD and enthalpy	$f(x,t)$8

The kinetic Monte Carlo case uses training PDFs from 46.5 to 599.5 mcs and tests at 16,681.1 mcs, with calibrated coefficients $f(x,t)$9 and $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$0 after the log transform. The phase-field case uses training PDFs from 1400 to 1700 $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$1 and tests at 2405 $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$2, with $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$3 and $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$4. The molecular-dynamics MSD case fits time-dependent polynomials,

$$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$5

while the enthalpy case yields $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$6 and $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$7 (Tran et al., 2020).

The same paper is explicit about the framework’s assumptions and limits. It assumes Markovianity and white noise, adequate time-series sampling, and typically one scalar QoI per ROM. It also notes that purely data-driven estimation can yield negative values of $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$8 under noise, that higher-dimensional Fokker–Planck equations suffer from the curse of dimensionality, and that strongly skewed or heavy-tailed PDFs may require higher-order Kramers–Moyal terms or non-Gaussian noise models (Tran et al., 2020).

3. SROM as an optimized discrete probability measure for uncertainty quantification

A distinct usage defines SROM as a non-intrusive uncertainty quantification technique that replaces a continuous target probability law by a low-order, weighted discrete measure. In the personalized-drug design paper, the uncertain inputs are the dissolution rates $$\dot{x} = h(x,t) + g(x,t)\,\Gamma(t),\quad \langle \Gamma(t)\rangle=0,\quad \langle \Gamma(t)\Gamma(t')\rangle=\delta(t-t'),$$9 in a Noyes–Whitney-based dissolution model. The SROM is specified by support points $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$0 and probabilities $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$1 satisfying $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$2 and $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$3, and it is constructed by solving a constrained minimization problem that matches cumulative distribution functions, non-central moments, and correlations to those of the target distribution (Altunay et al., 22 Jul 2025).

The construction minimizes

$dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$4

where $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$5 matches analytical CDFs, $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$6 matches non-central moments up to order $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$7, and $dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$8 matches correlations. The paper emphasizes that this SROM is non-intrusive, uses a small number of carefully selected weighted samples, avoids sampling noise because every iteration reuses the same weighted samples, and does not require the design of an appropriate sparse quadrature with fixed bounds for the random input variables (Altunay et al., 22 Jul 2025).

The forward model is geometric. Dissolution is represented by the Noyes–Whitney law

$dx(t) = D^{(1)}(x,t)\,dt + D^{(2)}(x,t)\,dW(t).$9

and by an Eikonal equation for a front moving with position-dependent speed $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$0,

$$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$1

The remaining mass at time $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$2 is

$$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$3

with interpolated material properties

$$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$4

Robust design minimizes an expectation–standard deviation objective,

$$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$5

where the expectation and variance are approximated by the SROM support points and weights (Altunay et al., 22 Jul 2025).

The numerical examples assume independent Gamma laws for $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$6 and $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$7 and use $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$8 weighted samples in both a linear target-release example and a pulsatile target-release example. The paper reports that the SROM closely matches the target Gamma CDFs, that SROM-based robust topology optimization yields release profiles with reduced variance relative to deterministic designs, and that the 95–5% uncertainty bands are narrower, especially at later dissolution times or during the pulsatile interval from $$\frac{\partial f}{\partial t} = -\partial_x\!\left(D^{(1)} f\right) + \partial_x^2\!\left(D^{(2)} f\right).$$9 to $x(t)$0 (Altunay et al., 22 Jul 2025). This usage of SROM is therefore a quadrature-like probabilistic surrogate for uncertain inputs, not a reduced stochastic dynamics for state evolution.

4. Basis-adaptive and multiscale stochastic ROMs for time-dependent systems

One line of work uses SROM to denote reduced models that adapt basis functions or stochastic closures to heterogeneity in the solution ensemble. The pre-classification-based SROM combines a time-dependent generalized centroidal Voronoi tessellation aligned with POD optimality, cluster-specific POD bases, and a naive Bayes pre-classifier trained on random inputs. Instead of one global POD basis, the solution ensemble is partitioned into $x(t)$1 clusters using a distance defined by time-integrated projection residual. For a new input $x(t)$2, the classifier predicts a cluster label and the corresponding cluster-based POD basis is used online. The paper gives an error bound in the ideal-classification case in terms of per-cluster POD tail energies, plus sampling and time-discretization errors, and reports improved accuracy over standard POD in stochastic Navier–Stokes experiments (Xiong et al., 2022).

The same work also makes the trade-off between local adaptivity and classification error explicit. In a trigonometric random-inflow case, predicted-label test error decreased from 0.6256 for $x(t)$3 to 0.5038 for $x(t)$4, while the estimated classifier error increased from the trivial single-cluster case to about 9.22% for $x(t)$5 and 20.10% for $x(t)$6. In a second hat-type inflow case with white noise, the classifier error rose further, to about 15.80% for $x(t)$7 and 31.99% for $x(t)$8, but the clustered ROM still improved on standard POD (Xiong et al., 2022). The method is therefore a stochastic reduced model in the sense that it learns a distribution-dependent basis-selection mechanism from parameter randomness.

A related but more intrusive class of methods uses time-dependent bases to evolve a low-rank stochastic manifold for SPDEs. In the adaptive sparse interpolation paper, stochastic reduced-order modeling with time-dependent bases is formulated through a decomposition

$x(t)$9

with orthonormality constraints on $a(x,t)\equiv D^{(1)}(x,t)$0 and $a(x,t)\equiv D^{(1)}(x,t)$1 and evolution equations obtained by minimizing the residual of the governing SPDE. The paper adopts the dynamically bi-orthonormal formulation and addresses the nonlinearity barrier by replacing the full right-hand side with an online DEIM-type low-rank interpolant $a(x,t)\equiv D^{(1)}(x,t)$2 constructed without offline snapshots. This yields compressed updates whose nominal cost is $a(x,t)\equiv D^{(1)}(x,t)$3, rather than the full-order scaling required by naive evaluation of generic nonlinearities, and the reported Burgers and compressible Navier–Stokes examples show orders-of-magnitude reduction in computational cost (Naderi et al., 2022).

Another pair of works locates stochasticity in the reduced coefficient dynamics rather than in the basis. One S-ROM for Burgers augments the POD-Galerkin drift by data-driven linear and quadratic closure terms plus additive noise, trains on multiple trajectories from random initial conditions, and proves convergence of both the estimated POD modes and the parameter estimates at the canonical Monte Carlo rate $a(x,t)\equiv D^{(1)}(x,t)$4. The same paper reports accurate trajectory-wise predictions from new initial conditions and prediction times far beyond the training range, while quantifying uncertainty due to unresolved scales through ensemble forecasting (Lu et al., 2022). A multiscale framework goes further by partitioning the state into large, medium, and small scales and building a conditional-Gaussian SROM in which medium-scale variables are linear Gaussian conditional on the history of the large scales,

$a(x,t)\equiv D^{(1)}(x,t)$5

with closed ODEs for $a(x,t)\equiv D^{(1)}(x,t)$6 and $a(x,t)\equiv D^{(1)}(x,t)$7. Because this structure admits an analytic nonlinear data-assimilation solution, the model can recover unobserved states without ensemble sampling and, in the reported turbulent regimes, outperforms a deterministic Galerkin ROM in both free-run statistics and data assimilation (Mou et al., 2022).

5. Data-driven stochastic operator inference, latent SDEs, and randomized projections

A nonintrusive operator-inference variant of SROM has been developed for bilinear stochastic differential equations with additive noise. The reduced model has the same algebraic form as an intrusive POD-Galerkin ROM,

$a(x,t)\equiv D^{(1)}(x,t)$8

but the reduced operators are inferred directly from state observations by least squares. Drift operators are fitted from empirical expectation dynamics, and diffusion is recovered from residuals in the covariance Lyapunov equation. The paper also proposes two stochastic generalizations of snapshot-based subspace construction, one using state snapshots and one using moment snapshots, and proves that the moment-snapshot left-singular-vector subspace is contained in that of the state-snapshot matrix. Under mild assumptions, the inferred operators converge almost surely to those of the intrusive POD ROM, and numerical tests on heat and convection–reaction systems show that the nonintrusive SROM matches POD-level errors down to the noise floor of about $a(x,t)\equiv D^{(1)}(x,t)$9 (Freitag et al., 2024).

A more recent development learns continuous-time latent stochastic ROMs jointly with probabilistic autoencoders. In that framework, the high-dimensional state $D(x,t)\equiv D^{(2)}(x,t)$0 is encoded into a latent process $D(x,t)\equiv D^{(2)}(x,t)$1 governed by

$D(x,t)\equiv D^{(2)}(x,t)$2

while decoding is probabilistic,

$D(x,t)\equiv D^{(2)}(x,t)$3

The central methodological device is an amortized stochastic variational formulation with a Markov Gaussian process variational family, which eliminates forward SDE solves inside the training loop. The paper states that the per-iteration training cost scales as $D(x,t)\equiv D^{(2)}(x,t)$4 and is independent of the number of trajectories, the number of time steps per trajectory, and the stiffness of the SDE. Reported numerical studies on a reaction–diffusion system, forced parametrized Burgers, and controlled cylinder wake show strong generalization to unseen parameters and forcings, with lower mean errors and much shorter training times than PNODE and PNSDE baselines (Ilersich et al., 15 Jan 2026).

A different uncertainty-aware OpInf formulation places the stochasticity not in latent diffusion but in the projection itself. Here the projection matrix is randomized on a constrained Stiefel subset,

$D(x,t)\equiv D^{(2)}(x,t)$5

so that linear physical constraints are preserved almost surely. Given anchor projection matrices $D(x,t)\equiv D^{(2)}(x,t)$6, a global basis $D(x,t)\equiv D^{(2)}(x,t)$7, and a Dirichlet random weight vector $D(x,t)\equiv D^{(2)}(x,t)$8, the random projection is constructed as

$D(x,t)\equiv D^{(2)}(x,t)$9

The Dirichlet concentration parameters are selected by a quadratic program enforcing a Fréchet-mean condition at $S_2$0, and each sample is associated with the closest anchor operator by the index of the largest weight. The method is presented as a non-intrusive SROM for model-form uncertainty in ROM projection spaces and is demonstrated on Burgers and two-dimensional Navier–Stokes benchmarks (Yong et al., 2024).

6. Assumptions, limitations, and recurrent misconceptions

A recurrent misconception is that SROM refers to one established formalism. The arXiv record shows the opposite. One paper’s SROM may evolve a scalar descriptor PDF with a Fokker–Planck equation, another may optimize weighted support points for uncertain inputs, another may switch among local POD bases by classification, and another may learn latent SDEs or randomize projection matrices on a Stiefel manifold (Tran et al., 2020, Altunay et al., 22 Jul 2025, Xiong et al., 2022, Yong et al., 2024). This suggests that precise contextual definition is indispensable whenever the acronym is used.

The assumptions are correspondingly heterogeneous. The microstructure Fokker–Planck SROM assumes Markovianity, white noise, adequate time-series sampling, and a scalar descriptor per ROM instance (Tran et al., 2020). The discrete-measure SROM for robust drug design assumes that only the dissolution rates are random, that they follow specified Gamma laws, and that independence holds between $S_2$1 and $S_2$2 in the examples (Altunay et al., 22 Jul 2025). The pre-classification SROM assumes that input features can predict cluster labels with sufficient fidelity under a naive Bayes model (Xiong et al., 2022). The time-dependent-basis and conditional-Gaussian ROMs assume a low-rank evolving manifold or conditional linearity sufficient to make the reduced stochastic dynamics tractable (Naderi et al., 2022, Mou et al., 2022). The nonintrusive operator-inference SROM is tailored to bilinear drift with additive Gaussian noise, while the latent autoencoding framework uses Gaussian likelihoods and highlights possible identifiability issues between decoder noise and latent diffusion (Freitag et al., 2024, Ilersich et al., 15 Jan 2026).

The limitations reported in the papers are equally method-specific. High-dimensional descriptor PDFs lead to the curse of dimensionality in Fokker–Planck formulations, and strongly skewed or heavy-tailed PDFs may not be captured well by a second-order truncation (Tran et al., 2020). Too few weighted support points may under-represent tails or higher-order statistics in discrete-measure SROMs (Altunay et al., 22 Jul 2025). Increasing the number of clusters can improve local energy capture but also increase misclassification rates in cluster-based ROMs (Xiong et al., 2022). Generic nonlinear SPDEs defeat the nominal efficiency of time-dependent-basis methods unless the nonlinear right-hand side is hyper-reduced, which motivates adaptive DEIM constructions (Naderi et al., 2022). In latent stochastic ROMs, phase drift over long horizons and heavy-tailed noise remain open issues (Ilersich et al., 15 Jan 2026). In randomized-projection SROMs, Riemannian logarithm and exponential evaluations and manifold clustering can become computationally heavy at large scale (Yong et al., 2024).

A plausible implication is that the appropriate SROM architecture depends first on what is being reduced. If the reduced object is the PDF of a scalar descriptor, Fokker–Planck dynamics are natural. If the main goal is non-intrusive propagation of input uncertainty through a deterministic solver, optimized discrete measures are natural. If the dominant challenge is variability across parameter regimes or transient solution classes, cluster-specific or time-dependent bases are natural. If full uncertainty-aware state prediction is required directly from data, operator-inference SDEs, latent stochastic autoencoders, or randomized-projection formulations are natural. Under that broader view, SROM is less a single method than a research program: the systematic insertion of stochastic structure into reduced-order modeling so that reduced surrogates preserve not only low-dimensional dynamics but also uncertainty, statistical variability, and model-form ambiguity.