Deep Neural-Operator Framework

• The DNO framework is a deep learning architecture that parameterizes branch and trunk nets to learn operators mapping between infinite-dimensional function spaces.
• It employs a sum-of-products structure, as seen in DeepONet, RINO, and other variants, to guarantee universal approximation of nonlinear operators.
• Applications include real-time surrogate modeling, digital twins, and accelerated PDE solvers, achieving high accuracy with reduced training time.

A deep neural-operator (DNO) framework is a class of machine learning methodology for learning mappings between infinite-dimensional function spaces—operators—using deep neural network architectures. This approach generalizes classical supervised learning, which targets finite-dimensional input–output mappings, by instead parameterizing nonlinear operators that act on functions or fields and can be trained on paired input–output functions derived from scientific simulations, experiments, or real-world dynamical systems.

1. Mathematical Foundations and Operator Parameterization

The core goal is to approximate an operator $\mathcal{G} : \mathcal{U} \to \mathcal{Y}$, where $\mathcal{U}$ and $\mathcal{Y}$ are spaces of functions, frequently Banach spaces defined on spatial or spatiotemporal domains. The DNO framework is motivated by the universal approximation theorem for operators, which states that for suitable pairs of function spaces and continuous operators, any such operator may be uniformly approximated by sum-of-products representations: $$\mathcal{G}(u)(y) \approx \sum_{k=1}^K a_k(u) \, b_k(y)$$ where $a_k$ are continuous functionals on $\mathcal{U}$ and $b_k$ are continuous basis functions on the output domain, with $K$ large enough for the target accuracy (Zhang, 2024, Liu et al., 2024).

In Deep Operator Networks (DeepONets), $a_k$ and $b_k$ are parameterized as neural networks—the branch net encodes the input function, while the trunk net provides basis functions for the output space. For input function $u$, sensor-sampled at points $\{x_j\}_{j=1}^{N_s}$, and output location $y$, the canonical neural operator output is: $$\widehat{\mathcal{G}}_\theta(u)(y) = \sum_{k=1}^K a_k(\hat{u};\alpha) \, b_k(y;\beta)$$ with $\hat{u} = [u(x_1), ..., u(x_{N_s})]^T$, and $\alpha$, $\beta$ are network parameters for branch and trunk nets, respectively.

2. Canonical Architectures and Variants in DNO

The DNO paradigm encompasses several principal model classes, each targeting different facets of operator learning:

• Deep Operator Network (DeepONet): Composed of a branch net (processes input function sampled at sensor points) and a trunk net (processes output coordinates). The sum-of-products structure ensures universality for a wide class of nonlinear operators (Goswami et al., 2022, Liu et al., 2024).
• Resolution-Independent Neural Operator (RINO): Overcomes the fixed-sensor limitation by projecting input/output functions onto learned continuous dictionaries (e.g., SIRENs), yielding coordinate embeddings compatible with DeepONet-style architectures independent of sensor arrangements (Bahmani et al., 2024).
• Physics-Informed and Energy-Dissipative Architectures: Incorporate prior physical knowledge through composite losses (including PDE residuals, energy laws), constraints, or time-evolution of network parameters (e.g., EDE-DeepONet embeds a scalar auxiliary variable to guarantee energy dissipation in learned dynamics) (Goswami et al., 2022, Zhang et al., 2023).
• Distributed and Multi-Operator DNOs: Architectures such as MODNO share a common branch net across distinct operators (multi-task/family learning), while the trunk net is localized to each operator, enabling cross-operator representation learning and parameter efficiency (Zhang, 2024). D2NO partitions heterogeneous input spaces, allocating separate branch nets per class and a shared trunk net to achieve universal approximation for highly variable input families (Zhang et al., 2023).
• Backpropagation-Free Operator Training: ELM-DeepONet replaces iterative gradient descent with a closed-form least-squares fit by freezing random features in both branch and trunk nets and training only the linear mixing weights (Son, 16 Jan 2025).

3. Theoretical Guarantees and Scaling Laws

DNO frameworks are grounded in proven universal approximation properties; for DeepONet-type models, it is shown that for every continuous nonlinear operator $\mathcal{G}$, there exist neural nets for $a_k$ and $b_k$ yielding arbitrarily close approximations on compacta (Liu et al., 2024, Goswami et al., 2022). These results extend to settings with finite discretization and variable input dimension via dictionary learning (Bahmani et al., 2024).

Scaling laws for approximation error and generalization have been rigorously established. For general Lipschitz operators where the input space is $d_1$-dimensional and output space is $d_2$-dimensional:

• The minimal achievable operator approximation error with $N_\#$ total network parameters scales as $$\epsilon \sim (\log N_\# / \log \log N_\# )^{-1/d_1}$$ (Liu et al., 2024).
• When input functions reside on a lower-dimensional linear subspace of dimension $b_U$, scaling improves to $\epsilon \sim N_\#^{-1/((d_2+1) b_U + d_2)}$.
• Generalization error over $M$ input–output examples scales as $M^{-1/(2 + (d_2+1) b_U + d_2)}$ under low-dimensional input structure.

4. Distributed, Multi-Task, and Resolution-Invariant Training

Modern DNO research emphasizes scalability to large families of operators, heterogeneous datasets, and arbitrary input/output samplings.

• MODNO (Zhang, 2024) shares the branch net (input encoder) across all operators, while trunk nets (output bases) are operator-specific and decoupled for local updates. This distributed scheme enables more efficient data use, parameter sharing, and cross-task learning, especially beneficial when some operators have limited data.
• D2NO (Zhang et al., 2023) handles input heterogeneity by allocating each function class a custom branch net and sensor set, while maintaining a global trunk. This allows efficient, universal learning without oversampling smooth functions or underfitting rough ones.
• RINO (Bahmani et al., 2024) and similar architectures employ dictionary learning (e.g., SIRENs) to project arbitrarily-sampled input data onto finite-dimensional embeddings for compatibility with DeepONet structure, thus handling variable-resolution data robustly.

5. Specialized Enhancements and Physics Integration

Extensions of the DNO paradigm include architectural, algorithmic, and physics-based innovations:

• Hybrid Decoder-DeepONet: Addresses unaligned observation data by replacing the rigid dot product in DeepONet with a feedforward decoder network, allowing efficient operator learning for scattered or unstructured observations. Multi-Decoder-DeepONet incorporates mean-field statistics to improve accuracy (Chen et al., 2023).
• Dual-Path Neural Operator (DPNO): Arranges each operator block in parallel ResNet- and DenseNet-style streams, fusing features at the output; empirically yields substantial improvements in solution accuracy for standard PDE benchmarks with only moderate parameter overhead (Wang et al., 17 Jul 2025).
• Stability Analysis via B-Spline DNO: Represents outputs as B-spline expansions parameterized by neural networks, enabling rigorous post-training spectral analysis via Dynamic Mode Decomposition and linkage to the Koopman framework for nonlinear stability characterization (Romagnoli et al., 22 Dec 2025).
• Derivative-Enhanced DeepONet (DE-DeepONet): Adds derivative information (parameter and spatial) to the loss function and performs linear dimension reduction on high-dimensional inputs via KLE or active subspaces, yielding significant generalization gains in low-data regimes (Qiu et al., 2024).
• Operator Learning for Probabilistic Models: DNOs for path-dependent functionals (e.g., stochastic differential equations and option pricing) admit theoretical guarantees under technical conditions (moment bounds, Lipschitz continuity) and provide uniform network-size error bounds for pricing operators in both European and American option contexts (Bayraktar et al., 10 Nov 2025).

6. Applications and Performance Characteristics

DNO frameworks enable direct emulation of parametric PDE solvers, real-time physics surrogacy, large-scale uncertainty quantification, and acceleration of complex scientific and engineering workflows.

• Real-Time Surrogate Modeling: DNOs deliver sequence-to-sequence forecasts for systems with fast transients, e.g., gas flow dynamics in engine models, achieving $\mathcal{L}_2$ errors of 6.5–12% and robust uncertainty quantification via ensembling (Kumar et al., 2023).
• Complex Physics and Multi-Geometry Generalization: Diffeomorphism neural operators employ geometric mapping and FNOs to learn solution operators across variable domains and parameters, with rigorous quantification of out-of-domain generalization error (Zhao et al., 2024).
• Accelerated High-Dimensional Surrogates: ELM-DeepONets provide backpropagation-free fitting, achieving two to four orders of magnitude reduction in training time with accuracy competitive or superior to gradient-trained DeepONets on ODE and PDE tasks (Son, 16 Jan 2025).
• Integrated Digital Twins and Real-Time Control: DNO surrogates embedded in digital twin architectures enable uncertainty-aware, closed-loop control and online calibration of additive manufacturing processes, with millisecond inference and sub-percent field errors (Liu et al., 2024).
• Scientific Forecasting and Data Assimilation: DNOs and their variants (e.g., latent DeepONet, FCN-based operator networks) support real-time ocean forecasting and large-scale geophysical predictions, maintaining phase accuracy over long time horizons and competitive root-mean-squared errors (Rajagopal et al., 2023).

7. Prospects and Limitations

DNO methodologies, underpinned by exhaustive theoretical analysis and empirical validation across scientific domains, provide a robust framework for learning complex operators, incorporating physical constraints, distributing model capacity efficiently, and scaling to real-world, high-dimensional data. Limitations persist in data efficiency for certain regimes, the interpretability and regularization of learned representations, and deployment to highly multimodal, multimodal or regime-switching problems, which remain ongoing areas of research within the community.

Papers that exemplify various facets of the DNO paradigm include works on MODNO (Zhang, 2024), neural scaling laws in DeepONet (Liu et al., 2024), dual-path architectures (Wang et al., 17 Jul 2025), B-spline DNO stability (Romagnoli et al., 22 Dec 2025), distributed training (Zhang et al., 2023), ELM-based training (Son, 16 Jan 2025), derivative-enhanced frameworks (Qiu et al., 2024), and digital twins for complex manufacturing (Liu et al., 2024).