DeepONet Operator: Neural Mapping of Function Spaces

Updated 21 April 2026

DeepONet operator is a neural architecture that learns nonlinear mappings between function spaces using a branch-trunk network structure.
It employs an inner-product fusion technique based on universal approximation theorems to ensure accurate operator representation.
DeepONet is applied in surrogate modeling for PDEs, real-time control, and uncertainty quantification in complex systems.

A DeepONet operator is a neural operator architecture designed to learn nonlinear mappings between function spaces—specifically, to approximate general solution operators arising from differential equations, dynamical systems, and other infinite-dimensional problems. It implements a theoretically grounded inner-product structure capable of approximating any continuous nonlinear operator, as guaranteed by a universal approximation theorem. Since its introduction, DeepONet has been extended to address practical, architectural, and training challenges across a range of applications in the sciences, engineering, and control.

1. Mathematical Foundations and Operator Representation

A DeepONet realizes an operator $G : \mathcal{V} \to \mathcal{F}$ mapping an input function $v(\cdot)$ to an output function $G[v](\cdot)$ via a decomposition

$G[v](y) \approx \hat{G}[v](y) = \sum_{i=1}^p b_i(v(x_1), \ldots, v(x_m))\, t_i(y)$

where:

$b_i$ are outputs of the branch network, taking as input $v$ evaluated at a set of sensor points $\{x_1,\ldots,x_m\}$ ,
$t_i$ are outputs of the trunk network, parametrized by the output query $y$ (which may represent space, time, or a multi-dimensional location),
$p$ is the latent dimension, determining the expressiveness of the learned operator basis.

Theoretical justification comes from the universal operator approximation theorem, originally due to Chen & Chen (1995), and further formalized by Lu et al. (2021): for any compact $v(\cdot)$ 0 and any continuous nonlinear operator $v(\cdot)$ 1, arbitrary accuracy is attainable given sufficient network capacity, number of sensors, and latent width (Lu et al., 2019).

2. Network Architecture and Variants

Standard DeepONet

Branch Network: Typically a multi-layer feed-forward neural network (MLP), ingesting sampled values of the input function.
Trunk Network: Another MLP, whose input is the evaluation coordinate $v(\cdot)$ 2.
Fusion: The outputs of the branch and trunk networks are combined in an inner product yielding the prediction $v(\cdot)$ 3.
Loss: Mean squared error over paired training data.

Advanced Variants

Multiple Branches: For multi-input operators—e.g., PDEs with multiple input functions—separate branch networks are constructed, and their outputs are fused with the trunk via element-wise products and inner sums (Tan et al., 2022, Kong et al., 2023).
Physics-Informed DeepONet: Additional loss terms enforce PDE residuals, boundary conditions, or physical constraints, integrating automatic differentiation for derivatives (Goswami et al., 2022, Sun et al., 22 Feb 2026).
Fourier-/Wavelet-/POD-Enhanced Trunks: For solution fields with sharp or multi-scale features, trunk inputs may be mapped via Fourier features or proper orthogonal decomposition (POD) bases, yielding enhanced expressive power for capturing complex solution manifolds (Choi et al., 14 Jul 2025, Sharma et al., 2024).
Ensemble and Mixture-of-Experts: Multiple trunks (ensemble of features) or partition-of-unity spatial mixtures improve error rates and enable local adaptivity (Sharma et al., 2024).
Resolution Independence: Learned implicit dictionary bases compatible with arbitrarily sampled input functions enable architecture-agnostic embedding of point-cloud data (Bahmani et al., 2024).

3. Training Methodologies and Loss Functions

Data-Driven Training

When paired input–output datasets are available, DeepONet is trained by empirical minimization of the mean squared error: $v(\cdot)$ 4 where $v(\cdot)$ 5 denotes the DeepONet with trainable parameters $v(\cdot)$ 6.

Physics-Informed and Derivative-Enhanced DeepONet

When the governing equations or physical constraints are available (but possibly no or little ground-truth data), terms enforcing the PDE residual, initial/boundary conditions, or variational energies can be included: $v(\cdot)$ 7 Automatic differentiation computes derivatives needed for PDE constraints or for enforcing parameter/spatial derivatives of the learned solution, as in DE-DeepONet (Qiu et al., 2024).

Efficient and Robust Optimization

Randomized/Linear DeepONets: Using randomized neural networks for branch/trunk (with only the output layer trained by least-squares), order-of-magnitude faster training is achieved with modest accuracy loss (Jiang et al., 1 Mar 2025).
Two-Step Training: Sequentially training the trunk and then branch sub-networks (with orthonormalization) improves stability and generalization bounds, decoupling errors due to input, output, and network size (Lee et al., 2023).
Joint Physics–Data Loss: Physics-informed terms can use surrogate ML models exploiting short-term dependencies (e.g., when ODE order is known but functional form unknown) to enforce derivative constraints, robustly improving learning from sparse and noisy data (Sun et al., 22 Feb 2026).

4. Extensions to Complex Inputs, Uncertainty Quantification, and Applications

Multi-Input, Temporal, and Geometric Extensions

Multiple Functional Inputs: Enhanced DeepONet (EDeepONet) and MIONet architectures allow several input functions via parallel branches and elementwise/trunk fusion (Tan et al., 2022, Kong et al., 2023).
Temporal and Variable-Length Inputs: Sequential DeepONet integrates recurrent structures (GRU/LSTM) for sequence modeling, capturing long-range temporal dependence and enabling real-time, variable-length predictions (He et al., 2023, Kong et al., 2023).
Geometric Awareness: Extensions such as Geom-DeepONet fuse shape parameters, signed-distance functions, and SIREN-based trunk representations for predictive modeling on complex 3D geometries with point-cloud data (He et al., 2024).
Parametric/Interface Problems: XI-DeepONet incorporates varying geometries via level-set/implicit functions as network inputs, enabling generalization across families of parametric interface PDEs (Bi et al., 2024).

Robustness, Generalization, and Uncertainty

Bayesian/Probabilistic DeepONet: Posterior sampling over parameters (via SGHMC or similar) and probabilistic output heads enable uncertainty quantification, yielding well-calibrated confidence intervals for time series and system trajectories (Moya et al., 2022).
Physics/Noise Robustness: Surrogate-augmented and physics-informed loss structures provide 2–3 $v(\cdot)$ 8 lower errors under sparse, noisy, or out-of-distribution inputs relative to pure data-driven DeepONet (Sun et al., 22 Feb 2026).

Surrogacy, Control, and Real-World Applications

DeepONet serves as a fast, generalizable surrogate in:

Real-time model predictive control (via MS-DeepONet for block horizon prediction) (Jong et al., 23 May 2025)
Subsurface field inversion (acoustic FWI), electromagnetic/plasma simulation acceleration via reduced-order surrogates (Nath et al., 14 Apr 2025, Lv et al., 27 Apr 2025)
Large-scale parameterized geometry predictions, design optimization, uncertainty quantification (He et al., 2024, Moya et al., 2022, Sharma et al., 2024)

5. Performance, Design Insights, and Best Practices

Empirical Benchmarks

In a broad suite of PDE and ODE operators, DeepONet achieves relative errors in the 1–5% range with moderate data and parameter budget, often matching or exceeding Fourier Neural Operators (FNOs), especially on irregular domains or with non-uniform sampling (Goswami et al., 2022, Choi et al., 14 Jul 2025).
Physics-informed and ensemble/PoU-DeepONet architectures yield further error reductions (up to $v(\cdot)$ 9– $G[v](\cdot)$ 0) over vanilla/standard configurations (Sharma et al., 2024, Choi et al., 14 Jul 2025).
Sequential/trunk-feature enhanced variants excel where solution fields have high-frequency, sharp, or interface-dominated structure (He et al., 2024, Choi et al., 14 Jul 2025, Sharma et al., 2024).

Design and Implementation Recommendations

Align PDE parameters that modulate solution bases (e.g., diffusion coefficients) with trunk-net inputs to exploit known analytical structure (Choi et al., 14 Jul 2025).
For multiple or disentangled inputs, implement parallel branch networks and consider explicit architectural separation (Tan et al., 2022).
For temporal or history-dependent systems, employ GRU/LSTM-based sequence encoders or hybrid feed-forward–recurrent branch/trunk structures (He et al., 2023, Kong et al., 2023).
To avoid over-parameterization, use POD/feature dictionaries for branch/trunk compression or hybridize with randomized neural layers to accelerate training (Sharma et al., 2024, Jiang et al., 1 Mar 2025).
Always ensure training sets broadly span the functional/parametric variability likely in deployment; DeepONet is a powerful interpolator, but extrapolation beyond training support is unreliable (Choi et al., 14 Jul 2025).

6. Limitations, Open Challenges, and Future Directions

Structural and Data Limitations

DeepONets require manual or prior-informed sensor selection for input sampling; signal resolution below the input length-scale can degrade performance (Bahmani et al., 2024).
Robust extrapolation outside the training parameter/function space remains limited; physics-informed or hybrid regularization only partially mitigates this (Choi et al., 14 Jul 2025, Sun et al., 22 Feb 2026).
Surrogate/physics-informed methods depend critically on surrogate generalization; errors in the auxiliary model propagate into operator loss (Sun et al., 22 Feb 2026).

Open Problems and Prospects

Further efficiency and generalization improvements may arise from optimal sensor selection, adaptive gnoring, or hybrid FNO–DeepONet architectures.
Physics-informed dictionary or manifold learning for both branches and trunks could yield higher-order generalization.
Extensions to high-dimensional, multi-physics, or implicit data regimes (e.g., meshless methods, implicit geometric representations) are ongoing challenges (He et al., 2024, Bi et al., 2024, Bahmani et al., 2024).
Operator learning for systems with latent, partially observed, or incomplete inputs remains largely unexplored terrain.

DeepONet has rapidly become a central architecture for operator learning, offering mathematically rigorous, highly flexible, and computationally efficient surrogacy for a wide class of complex systems. Through further architectural, training, and application-driven innovation, DeepONet and its variants are expected to play an increasingly prominent role in computational modeling, control, and scientific machine learning.