Stochastic Operator Network (SON)

• SON is a class of operator learning models that represents stochastic mappings between function spaces using probabilistic architectures.
• It leverages SDE-based branch networks, Hamiltonian-type loss functions, and latent-variable mechanisms to effectively capture uncertainty in outputs.
• The approach is applied to stochastic PDEs, random fields, and quantum dynamics, enabling robust operator learning under noisy and variable conditions.

Searching arXiv for papers explicitly using “Stochastic Operator Network” and adjacent operator-learning formulations on stochastic processes. Stochastic Operator Network (SON) denotes a class of operator-learning models intended to represent stochastic mappings between function spaces or process spaces. In the explicit SON formulation, the architecture combines the branch–trunk structure of DeepONet with a Stochastic Neural Network (SNN): the branch net is modeled as an SDE, training is performed through an adjoint BSDE, and the gradient of the loss is replaced by the gradient of a Hamiltonian from the Stochastic Maximum Principle (SMP) in the SGD update (Bausback et al., 10 Jul 2025). A subsequent SPDE-oriented formulation retains the same basic idea for probabilistic prediction of stochastic solution operators and introduces a Hamiltonian-type loss for uncertainty quantification of solution fields (Huynh et al., 16 May 2026). In a broader, non-standardized usage suggested by adjacent work, SON also refers to operator learners that act on stochastic processes, random fields, or conditional distributions over solution operators, even when the papers themselves use different names for the construction (Ma et al., 30 Sep 2025, Saini et al., 30 Dec 2025).

1. Definition and conceptual scope

In the explicit DeepONet-based SON literature, the target is a stochastic operator rather than a single deterministic map. The operator takes an input function $u$ and returns a random output function, so the learned object is not only $u \mapsto G(u)$ but $u \mapsto G(u)(y,\omega)$, with stochasticity encoded inside the network through the diffusion term of an SDE (Bausback et al., 10 Jul 2025). The SPDE variant makes the same point in application-specific terms: SON takes deterministic problem data such as forcing, initial conditions, or coefficients, and outputs a distribution over solution fields, with randomness generated by Brownian sampling in hidden feature dynamics (Huynh et al., 16 May 2026).

The literature also supports a wider operator-theoretic reading. UQ-SONet formulates a stochastic operator model

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$

where $\mathcal O$ is a variable-size set of sensor observations, $h(\mathcal O)$ is a permutation-invariant embedding, and $z \sim \mathcal N(0,I)$ is a latent random variable (Ma et al., 30 Sep 2025). SINNOs and K-SNNOs, by contrast, are constructive stochastic operators on random processes in $L^2(\Omega,\mathcal F,\mathbb P)$, built analytically rather than learned from data (Saini et al., 30 Dec 2025, Saini et al., 7 Jan 2026). This suggests that SON is best understood as an umbrella notion for operator-valued neural or neural-inspired constructions whose outputs are stochastic processes, random fields, or conditional laws, rather than as a single fixed architecture.

2. Architectural families

The term is not standardized across the literature, and several relevant papers explicitly note that they do not use the phrase “Stochastic Operator Network.” The following families nevertheless instantiate closely related ideas.

Family	Stochastic mechanism	Representative papers
DeepONet–SNN SON	Branch net as an SDE; learned diffusion; SMP/Hamiltonian training	(Bausback et al., 10 Jul 2025, Huynh et al., 16 May 2026)
UQ-SONet	Latent variable $z$, set transformer embedding, cVAE decoder	(Ma et al., 30 Sep 2025)
SINNO / K-SNNO	Random coefficients or stochastic neurons in interpolation/Kantorovich operators	(Saini et al., 30 Dec 2025, Saini et al., 7 Jan 2026)
Chaos / transfer-operator / operator-construction variants	Wiener chaos basis, RKHS transfer operators, stochastic time-evolution operators	(Eigel et al., 2024, Ke et al., 6 Jan 2025, Zhang et al., 1 Sep 2025)

The DeepONet–SNN lineage is the most literal use of the name SON. In that setting, the branch network output is the terminal state of a stochastic flow, and the trunk remains a deterministic coordinate encoder. The resulting model preserves the operator-learning viewpoint of DeepONet while promoting the branch features from deterministic coefficients to random variables or random fields (Bausback et al., 10 Jul 2025).

UQ-SONet modifies a different axis of the operator-learning problem. Instead of fixing the number and locations of sensors, it uses a set transformer to produce a permutation-invariant representation of sparse and variable observations, then employs a conditional variational autoencoder to approximate the conditional distribution of the output operator (Ma et al., 30 Sep 2025). The stochasticity here is latent-variable-based rather than SDE-based.

Constructive stochastic operator models occupy another part of the design space. SINNOs define

$$\mathcal S_n(X_t,t) :=\sum_{k=0}^n X_{t_k}(\omega)\,\varphi_{\mathcal A(m)}\!\left(\frac{2m}{\delta}(t-t_k)\right),$$

so that randomness resides only in the coefficients $u \mapsto G(u)$0 and the activation and grid are deterministic (Saini et al., 30 Dec 2025). K-SNNOs instead place randomness in the neurons themselves through stochastic integrals, while the coefficients are deterministic cell averages of a kernel $u \mapsto G(u)$1 (Saini et al., 7 Jan 2026). In both cases, the operator acts directly on stochastic processes and admits mean-square analysis.

3. Stochastic mechanisms and optimization

The canonical SON construction of the DeepONet–SNN type models the branch hidden state $u \mapsto G(u)$2 by an SDE

$u \mapsto G(u)$3

or, in Euler–Maruyama form,

$u \mapsto G(u)$4

The branch output is the terminal state $u \mapsto G(u)$5, the trunk network returns $u \mapsto G(u)$6, and the operator output is

$u \mapsto G(u)$7

Because $u \mapsto G(u)$8 depends on Brownian noise, the operator output is random for fixed $u \mapsto G(u)$9 and $u \mapsto G(u)(y,\omega)$0 (Bausback et al., 10 Jul 2025).

Training in this formulation is cast as a stochastic optimal control problem with cost

$u \mapsto G(u)(y,\omega)$1

The Hamiltonian is written as

$u \mapsto G(u)(y,\omega)$2

and the gradient of the loss is replaced by $u \mapsto G(u)(y,\omega)$3, where $u \mapsto G(u)(y,\omega)$4 solve the adjoint BSDE. In the SPDE formulation, the Hamiltonian appears as

$u \mapsto G(u)(y,\omega)$5

with analogous forward–backward recursions in discrete time (Bausback et al., 10 Jul 2025, Huynh et al., 16 May 2026).

The SPDE paper also distinguishes a one-phase SON from a two-phase SON. In the one-phase version, the branch SNN, trunk network, and decoder are trained jointly. In the two-phase version, a deterministic Decoder-DeepONet is trained first to capture the mean solution structure, and a second SNN phase is then trained on the deterministic output to model uncertainty and mild refinement. The paper presents this two-phase design as the main practical contribution because it stabilizes training and reduces complexity in large operator models (Huynh et al., 16 May 2026).

Alternative probabilistic mechanisms differ substantially. UQ-SONet defines a conditional generative model

$u \mapsto G(u)(y,\omega)$6

with $u \mapsto G(u)(y,\omega)$7, and trains by minimizing the negative ELBO. Its uncertainty estimates are therefore obtained by Monte Carlo over latent draws $u \mapsto G(u)(y,\omega)$8, rather than by repeated Brownian sampling through a branch SDE (Ma et al., 30 Sep 2025).

4. Mathematical properties and approximation theory

The mathematically strongest guarantees in the surrounding literature come from constructive and chaos-based operator models. For SINNOs, the paper proves boundedness, exact interpolation at grid nodes in $u \mapsto G(u)(y,\omega)$9, reproduction of constants, and the estimate

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$0

where $$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$1 is the mean-square modulus of continuity. Under Hölder-type mean-square regularity, the error is $$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$2, and the analysis is carried out in mean square, in probability, and path-wise terms (Saini et al., 30 Dec 2025).

K-SNNOs provide a different constructive route. For processes with canonical representation

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$3

the K-SNNO approximation is

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$4

where the stochastic neurons $$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$5 are stochastic integrals against a process $$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$6 with orthogonal increments. The paper proves mean-square convergence and derives

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$7

for kernels in a Lipschitz-type class $$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$8 (Saini et al., 7 Jan 2026).

SDEONet contributes a DeepONet-inspired chaos factorization for stochastic differential equations. It approximates the SDE solution operator

$$\hat{\mathcal G}(h(\mathcal O), z)(y) = \sum_{n=1}^p b^n\big(h(\mathcal O), z\big)\, t^n(y),$$9

using a fixed Wiener chaos basis in the stochastic direction and neural networks for the deterministic chaos coefficients. The paper presents complete convergence and complexity analysis and frames the architecture as learning an optimal sparse truncation of the Wiener chaos expansion (Eigel et al., 2024).

A complementary theoretical line appears in the statistics-informed neural network literature. SINN establishes a universal approximation theorem for stochastic systems and interprets a recurrent neural network driven by i.i.d. noise as a learned stochastic evolution operator for Markovian and non-Markovian dynamics. The paper explicitly connects this viewpoint to the projection-operator formalism and coarse-grained stochastic modeling (Zhu et al., 2022).

5. Application domains

The explicit SON literature is anchored in stochastic PDEs and noisy operator regression. The original SON paper studies noisy operators in 2D and 3D and emphasizes recovery of uncertainty from noisy operator outputs (Bausback et al., 10 Jul 2025). The SPDE extension evaluates SON on 2D reaction–diffusion, 2D advection–diffusion, the 2D heat equation, and 2D Burgers’ equation, under multiple uncertainty sources including additive solution noise, random forcing, random coefficients, and space–time noise (Huynh et al., 16 May 2026).

UQ-SONet broadens the scope to deterministic and stochastic PDEs with sparse and variable sensor locations, including the 1D deterministic diffusion equation, 2D deterministic Poisson, 1D and 2D stochastic elliptic problems, and the Navier–Stokes equation. Its purpose is not only probabilistic prediction but also uncertainty arising from incomplete measurements or from operators with inherent randomness (Ma et al., 30 Sep 2025).

Several adjacent domains instantiate SON-like operator learning without using the label uniformly. In structural engineering, neural operators have been used for stochastic modeling of nonlinear structural system response to earthquakes and wind, with DeepONet, FNO, self-adaptive FNO, and a hybrid DeepFNOnet proposed for nonlinear time-history prediction under stochastic excitation (Goswami et al., 16 Feb 2025). In open quantum systems, a neural network is used as a universal generator of a stochastic time-evolution operator for non-Markovian quantum state diffusion, enabling reconstruction of reduced density matrices, absorption spectra, and extended-time dynamics (Zhang et al., 1 Sep 2025). In latent stochastic dynamics, operator-based latent Markov representations built from embedded latent transfer operators and observable operators support sequential state-estimation and operator-based eigen-mode decomposition in RKHS embeddings (Ke et al., 6 Jan 2025).

Stochastic operator learning also appears in reduced-order and time-series settings. SINNOs are proposed for approximation of stochastic processes with potential applications in COVID-19 case prediction (Saini et al., 30 Dec 2025). SINN is used for coarse-graining and transition dynamics, and the paper shows that the resulting reduced-order model can be trained on temporally coarse-grained data and is well suited for rare-event simulations (Zhu et al., 2022).

6. Terminological status, misconceptions, and open directions

A common misconception is that SON refers to a single settled architecture. The literature does not support that reading. “Stochastic Operator Network” is explicit in the DeepONet–SNN papers and in UQ-SONet, but several closely related works state that they do not use the phrase even though they construct stochastic operator-valued neural models or operator-like approximants on random processes (Bausback et al., 10 Jul 2025, Ma et al., 30 Sep 2025, Saini et al., 30 Dec 2025, Saini et al., 7 Jan 2026). A plausible implication is that SON currently functions more as a research direction than as a fully standardized technical term.

A second misconception is that stochasticity must reside in network weights. The papers collectively exhibit several distinct mechanisms: learned diffusion in a branch SDE (Bausback et al., 10 Jul 2025, Huynh et al., 16 May 2026), latent variables in a cVAE (Ma et al., 30 Sep 2025), random coefficients taken from process samples (Saini et al., 30 Dec 2025), stochastic neurons driven by stochastic integrators (Saini et al., 7 Jan 2026), and deterministic networks driven by noise sequences whose statistics are matched at the output level (Zhu et al., 2022). The choice is therefore architectural rather than definitional.

The main limitations are likewise heterogeneous. The explicit SON papers note the complexity of forward–backward SDE training, the Brownian assumption in the branch dynamics, and the problem-specific character of some diffusion parameterizations (Bausback et al., 10 Jul 2025, Huynh et al., 16 May 2026). The SPDE paper further remarks that computational cost remains higher than that of purely deterministic DeepONet and that rigorous error bounds for approximation of full solution distributions are not provided (Huynh et al., 16 May 2026). In the constructive SINNO literature, the strongest rigorously derived quantitative mode of convergence is uniform-in-$\mathcal O$0 $\mathcal O$1-convergence, even though the paper asserts almost sure convergence for fixed $\mathcal O$2 (Saini et al., 30 Dec 2025). In SDEONet, automatic discovery of the optimal sparse chaos subset is identified as nontrivial in practice (Eigel et al., 2024). In ELTO-based latent transfer-operator models, finite-rank cross-covariance assumptions and the absence of explicit finite-sample error bounds delimit the current theory (Ke et al., 6 Jan 2025).

The current trajectory of the field points toward broader stochastic operator learning rather than a single canonical SON design. This suggests continued convergence among SDE-based branch dynamics, latent-variable operator models, constructive mean-square approximation theory, RKHS transfer operators, and application-specific stochastic neural operators for PDEs, quantum dynamics, structural hazards, and reduced-order stochastic modeling (Huynh et al., 16 May 2026, Ma et al., 30 Sep 2025, Zhang et al., 1 Sep 2025, Goswami et al., 16 Feb 2025).