Generative models in function spaces, situated at the intersection of generative modeling and operator learning, are attracting increasing attention due to their immense potential in diverse scientific and engineering applications. While functional generative models are theoretically domain- and discretization-agnostic, current implementations heavily rely on the Fourier Neural Operator (FNO), limiting their applicability to regular grids and rectangular domains. To overcome these critical limitations, we introduce the Mesh-Informed Neural Operator (MINO). By leveraging graph neural operators and cross-attention mechanisms, MINO offers a principled, domain- and discretization-agnostic backbone for generative modeling in function spaces. This advancement significantly expands the scope of such models to more diverse applications in generative, inverse, and regression tasks. Furthermore, MINO provides a unified perspective on integrating neural operators with general advanced deep learning architectures. Finally, we introduce a suite of standardized evaluation metrics that enable objective comparison of functional generative models, addressing another critical gap in the field.
Neural operators, which aim to approximate mappings between infinite-dimensional function spaces, have been widely applied in the simulation and prediction of physical systems. However, the limited representational capacity of network architectures, combined with their heavy reliance on large-scale data, often hinder effective training and result in poor extrapolation performance. In this paper, inspired by traditional numerical methods, we propose a novel physics guided multi-step neural operator (PMNO) architecture to address these challenges in long-horizon prediction of complex physical systems. Distinct from general operator learning methods, the PMNO framework replaces the single-step input with multi-step historical data in the forward pass and introduces an implicit time-stepping scheme based on the Backward Differentiation Formula (BDF) during backpropagation. This design not only strengthens the model's extrapolation capacity but also facilitates more efficient and stable training with fewer data samples, especially for long-term predictions. Meanwhile, a causal training strategy is employed to circumvent the need for multi-stage training and to ensure efficient end-to-end optimization. The neural operator architecture possesses resolution-invariant properties, enabling the trained model to perform fast extrapolation on arbitrary spatial resolutions. We demonstrate the superior predictive performance of PMNO predictor across a diverse range of physical systems, including 2D linear system, modeling over irregular domain, complex-valued wave dynamics, and reaction-diffusion processes. Depending on the specific problem setting, various neural operator architectures, including FNO, DeepONet, and their variants, can be seamlessly integrated into the PMNO framework.
A wide range of scientific problems, such as those described by continuous-time dynamical systems and partial differential equations (PDEs), are naturally formulated on function spaces. While function spaces are typically infinite-dimensional, deep learning has predominantly advanced through applications in computer vision and natural language processing that focus on mappings between finite-dimensional spaces. Such fundamental disparities in the nature of the data have limited neural networks from achieving a comparable level of success in scientific applications as seen in other fields. Neural operators are a principled way to generalize neural networks to mappings between function spaces, offering a pathway to replicate deep learning's transformative impact on scientific problems. For instance, neural operators can learn solution operators for entire classes of PDEs, e.g., physical systems with different boundary conditions, coefficient functions, and geometries. A key factor in deep learning's success has been the careful engineering of neural architectures through extensive empirical testing. Translating these neural architectures into neural operators allows operator learning to enjoy these same empirical optimizations. However, prior neural operator architectures have often been introduced as standalone models, not directly derived as extensions of existing neural network architectures. In this paper, we identify and distill the key principles for constructing practical implementations of mappings between infinite-dimensional function spaces. Using these principles, we propose a recipe for converting several popular neural architectures into neural operators with minimal modifications. This paper aims to guide practitioners through this process and details the steps to make neural operators work in practice. Our code can be found at https://github.com/neuraloperator/NNs-to-NOs
High-fidelity and efficient simulation of fluid dynamics drive progress in various scientific and engineering applications. Traditional computational fluid dynamics methods offer strong interpretability and guaranteed convergence, but rely on fine spatial and temporal meshes, incurring prohibitive computational costs. Physics-informed neural networks (PINNs) and neural operators aim to accelerate PDE solvers using deep learning techniques. However, PINNs require extensive retraining and careful tuning, and purely data-driven operators demand large labeled datasets. Hybrid physics-aware methods embed numerical discretizations into network architectures or loss functions, but achieve marginal speed gains and become unstable when balancing coarse priors against high-fidelity measurements. To this end, we introduce OmniFluids, a unified physics pre-trained operator learning framework that integrates physics-only pre-training, coarse-grid operator distillation, and few-shot fine-tuning, which enables fast inference and accurate prediction under limited or zero data supervision. For architectural design, the key components of OmniFluids include a mixture of operators, a multi-frame decoder, and factorized Fourier layers, which enable efficient and scalable modeling of diverse physical tasks while maintaining seamless integration with physics-based supervision. Across a broad range of two- and three-dimensional benchmarks, OmniFluids significantly outperforms state-of-the-art AI-driven methods in flow field reconstruction and turbulence statistics accuracy, delivering 10-100x speedups compared to classical solvers, and accurately recovers unknown physical parameters from sparse, noisy data. This work establishes a new paradigm for efficient and generalizable surrogate modeling in complex fluid systems under limited data availability.
Operator learning enables data-driven modeling of partial differential equations (PDEs) by learning mappings between function spaces. However, scaling transformer-based operator models to high-resolution, multiscale domains remains a challenge due to the quadratic cost of attention and its coupling to discretization. We introduce \textbf{Mondrian}, transformer operators that decompose a domain into non-overlapping subdomains and apply attention over sequences of subdomain-restricted functions. Leveraging principles from domain decomposition, Mondrian decouples attention from discretization. Within each subdomain, it replaces standard layers with expressive neural operators, and attention across subdomains is computed via softmax-based inner products over functions. The formulation naturally extends to hierarchical windowed and neighborhood attention, supporting both local and global interactions. Mondrian achieves strong performance on Allen-Cahn and Navier-Stokes PDEs, demonstrating resolution scaling without retraining. These results highlight the promise of domain-decomposed attention for scalable and general-purpose neural operators.
Neural operators have emerged as powerful tools for learning solution operators of partial differential equations. However, in time-dependent problems, standard training strategies such as teacher forcing introduce a mismatch between training and inference, leading to compounding errors in long-term autoregressive predictions. To address this issue, we propose Recurrent Neural Operators (RNOs)-a novel framework that integrates recurrent training into neural operator architectures. Instead of conditioning each training step on ground-truth inputs, RNOs recursively apply the operator to their own predictions over a temporal window, effectively simulating inference-time dynamics during training. This alignment mitigates exposure bias and enhances robustness to error accumulation. Theoretically, we show that recurrent training can reduce the worst-case exponential error growth typical of teacher forcing to linear growth. Empirically, we demonstrate that recurrently trained Multigrid Neural Operators significantly outperform their teacher-forced counterparts in long-term accuracy and stability on standard benchmarks. Our results underscore the importance of aligning training with inference dynamics for robust temporal generalization in neural operator learning.
Out-of-distribution (OOD) generalization remains a fundamental challenge in machine learning. Models trained on one data distribution often experience substantial performance degradation when evaluated on shifted or unseen domains. To address this challenge, the present paper studies the design of training data distributions that maximize average-case OOD performance. First, a theoretical analysis establishes a family of generalization bounds that quantify how the choice of training distribution influences OOD error across a predefined family of target distributions. These insights motivate the introduction of two complementary algorithmic strategies: (i) directly formulating OOD risk minimization as a bilevel optimization problem over the space of probability measures and (ii) minimizing a theoretical upper bound on OOD error. Last, the paper evaluates the two approaches across a range of function approximation and operator learning examples. The proposed methods significantly improve OOD accuracy over standard empirical risk minimization with a fixed distribution. These results highlight the potential of distribution-aware training as a principled and practical framework for robust OOD generalization.
We consider the problem of learning the evolution operator for the time-dependent Schr\"{o}dinger equation, where the Hamiltonian may vary with time. Existing neural network-based surrogates often ignore fundamental properties of the Schr\"{o}dinger equation, such as linearity and unitarity, and lack theoretical guarantees on prediction error or time generalization. To address this, we introduce a linear estimator for the evolution operator that preserves a weak form of unitarity. We establish both upper and lower bounds on the prediction error that hold uniformly over all sufficiently smooth initial wave functions. Additionally, we derive time generalization bounds that quantify how the estimator extrapolates beyond the time points seen during training. Experiments across real-world Hamiltonians -- including hydrogen atoms, ion traps for qubit design, and optical lattices -- show that our estimator achieves relative errors $10^{-2}$ to $10^{-3}$ times smaller than state-of-the-art methods such as the Fourier Neural Operator and DeepONet.
Operator learning seeks to approximate mappings from input functions to output solutions, particularly in the context of partial differential equations (PDEs). While recent advances such as DeepONet and Fourier Neural Operator (FNO) have demonstrated strong performance, they often rely on regular grid discretizations, limiting their applicability to complex or irregular domains. In this work, we propose a Graph-based Operator Learning with Attention (GOLA) framework that addresses this limitation by constructing graphs from irregularly sampled spatial points and leveraging attention-enhanced Graph Neural Netwoks (GNNs) to model spatial dependencies with global information. To improve the expressive capacity, we introduce a Fourier-based encoder that projects input functions into a frequency space using learnable complex coefficients, allowing for flexible embeddings even with sparse or nonuniform samples. We evaluated our approach across a range of 2D PDEs, including Darcy Flow, Advection, Eikonal, and Nonlinear Diffusion, under varying sampling densities. Our method consistently outperforms baselines, particularly in data-scarce regimes, demonstrating strong generalization and efficiency on irregular domains.
The very challenging task of learning solution operators of PDEs on arbitrary domains accurately and efficiently is of vital importance to engineering and industrial simulations. Despite the existence of many operator learning algorithms to approximate such PDEs, we find that accurate models are not necessarily computationally efficient and vice versa. We address this issue by proposing a geometry aware operator transformer (GAOT) for learning PDEs on arbitrary domains. GAOT combines novel multiscale attentional graph neural operator encoders and decoders, together with geometry embeddings and (vision) transformer processors to accurately map information about the domain and the inputs into a robust approximation of the PDE solution. Multiple innovations in the implementation of GAOT also ensure computational efficiency and scalability. We demonstrate this significant gain in both accuracy and efficiency of GAOT over several baselines on a large number of learning tasks from a diverse set of PDEs, including achieving state of the art performance on a large scale three-dimensional industrial CFD dataset.
We introduce an encoder-only approach to learn the evolution operators of large-scale non-linear dynamical systems, such as those describing complex natural phenomena. Evolution operators are particularly well-suited for analyzing systems that exhibit complex spatio-temporal patterns and have become a key analytical tool across various scientific communities. As terabyte-scale weather datasets and simulation tools capable of running millions of molecular dynamics steps per day are becoming commodities, our approach provides an effective tool to make sense of them from a data-driven perspective. The core of it lies in a remarkable connection between self-supervised representation learning methods and the recently established learning theory of evolution operators. To show the usefulness of the proposed method, we test it across multiple scientific domains: explaining the folding dynamics of small proteins, the binding process of drug-like molecules in host sites, and autonomously finding patterns in climate data. Code and data to reproduce the experiments are made available open source.
Many architectures for neural PDE surrogates have been proposed in recent years, largely based on neural networks or operator learning. In this work, we derive and propose a new architecture, the Neural Functional, which learns function to scalar mappings. Its implementation leverages insights from operator learning and neural fields, and we show the ability of neural functionals to implicitly learn functional derivatives. For the first time, this allows for an extension of Hamiltonian mechanics to neural PDE surrogates by learning the Hamiltonian functional and optimizing its functional derivatives. We demonstrate that the Hamiltonian Neural Functional can be an effective surrogate model through improved stability and conserving energy-like quantities on 1D and 2D PDEs. Beyond PDEs, functionals are prevalent in physics; functional approximation and learning with its gradients may find other uses, such as in molecular dynamics or design optimization.
Operator learning is a rapidly growing field that aims to approximate nonlinear operators related to partial differential equations (PDEs) using neural operators. These rely on discretization of input and output functions and are, usually, expensive to train for large-scale problems at high-resolution. Motivated by this, we present a Multi-Level Monte Carlo (MLMC) approach to train neural operators by leveraging a hierarchy of resolutions of function dicretization. Our framework relies on using gradient corrections from fewer samples of fine-resolution data to decrease the computational cost of training while maintaining a high level accuracy. The proposed MLMC training procedure can be applied to any architecture accepting multi-resolution data. Our numerical experiments on a range of state-of-the-art models and test-cases demonstrate improved computational efficiency compared to traditional single-resolution training approaches, and highlight the existence of a Pareto curve between accuracy and computational time, related to the number of samples per resolution.
Real-world systems, from aerospace to railway engineering, are modeled with partial differential equations (PDEs) describing the physics of the system. Estimating robust solutions for such problems is essential. Deep learning-based architectures, such as neural PDE solvers, have recently gained traction as a reliable solution method. The current state of development of these approaches, however, primarily focuses on improving accuracy. The environmental impact of excessive computation, leading to increased carbon emissions, has largely been overlooked. This paper introduces a carbon emission measure for a range of PDE solvers. Our proposed metric, EcoL2, balances model accuracy with emissions across data collection, model training, and deployment. Experiments across both physics-informed machine learning and operator learning architectures demonstrate that the proposed metric presents a holistic assessment of model performance and emission cost. As such solvers grow in scale and deployment, EcoL2 represents a step toward building performant scientific machine learning systems with lower long-term environmental impact.
Solving partial or ordinary differential equation models in cardiac electrophysiology is a computationally demanding task, particularly when high-resolution meshes are required to capture the complex dynamics of the heart. Moreover, in clinical applications, it is essential to employ computational tools that provide only relevant information, ensuring clarity and ease of interpretation. In this work, we exploit two recently proposed operator learning approaches, namely Fourier Neural Operators (FNO) and Kernel Operator Learning (KOL), to learn the operator mapping the applied stimulus in the physical domain into the activation and repolarization time distributions. These data-driven methods are evaluated on synthetic 2D and 3D domains, as well as on a physiologically realistic left ventricle geometry. Notably, while the learned map between the applied current and activation time has its modelling counterpart in the Eikonal model, no equivalent partial differential equation (PDE) model is known for the map between the applied current and repolarization time. Our results demonstrate that both FNO and KOL approaches are robust to hyperparameter choices and computationally efficient compared to traditional PDE-based Monodomain models. These findings highlight the potential use of these surrogate operators to accelerate cardiac simulations and facilitate their clinical integration.
Recent advances in Neural Fields have enabled powerful, discretization-invariant methods for learning neural operators that approximate solutions of Partial Differential Equations (PDEs) on general geometries. Building on these developments, we introduce enf2enf, an encoder--decoder methodology for predicting steady-state Partial Differential Equations with non-parameterized geometric variability, based on recently proposed Equivariant Neural Field architectures. In enf2enf, input geometries are encoded into latent point cloud embeddings that inherently preserve geometric grounding and capture local phenomena. The resulting representations are then combined with global parameters and directly decoded into continuous output fields, thus efficiently modeling the coupling between geometry and physics. By leveraging the inductive biases of locality and translation invariance, our approach is able to capture fine-scale physical features as well as complex shape variations, thereby enhancing generalization and physical compliance. Extensive experiments on a high-fidelity aerodynamic dataset, a hyper-elastic material benchmark, and multi-element airfoil geometries, demonstrate that the proposed model achieves superior or competitive performance compared to state-of-the-art graph based, operator learning, and neural field methods. Notably, our method supports real time inference and zero-shot super-resolution, enabling efficient training on low-resolution meshes while maintaining high accuracy on full-scale discretizations.
Finite element (FE) modeling is essential for structural analysis but remains computationally intensive, especially under dynamic loading. While operator learning models have shown promise in replicating static structural responses at FEM level accuracy, modeling dynamic behavior remains more challenging. This work presents a Multiple Input Operator Network (MIONet) that incorporates a second trunk network to explicitly encode temporal dynamics, enabling accurate prediction of structural responses under moving loads. Traditional DeepONet architectures using recurrent neural networks (RNNs) are limited by fixed time discretization and struggle to capture continuous dynamics. In contrast, MIONet predicts responses continuously over both space and time, removing the need for step wise modeling. It maps scalar inputs including load type, velocity, spatial mesh, and time steps to full field structural responses. To improve efficiency and enforce physical consistency, we introduce a physics informed loss based on dynamic equilibrium using precomputed mass, damping, and stiffness matrices, without solving the governing PDEs directly. Further, a Schur complement formulation reduces the training domain, significantly cutting computational costs while preserving global accuracy. The model is validated on both a simple beam and the KW-51 bridge, achieving FEM level accuracy within seconds. Compared to GRU based DeepONet, our model offers comparable accuracy with improved temporal continuity and over 100 times faster inference, making it well suited for real-time structural monitoring and digital twin applications.
Neural operators, particularly the Deep Operator Network (DeepONet), have shown promise in learning mappings between function spaces for solving differential equations. However, standard DeepONet requires input functions to be sampled at fixed locations, limiting its applicability in scenarios with variable sensor configurations, missing data, or irregular grids. We introduce the Set Operator Network (SetONet), a novel architecture that integrates Deep Sets principles into the DeepONet framework to address this limitation. The core innovation lies in the SetONet branch network, which processes the input function as an unordered \emph{set} of location-value pairs. This design ensures permutation invariance with respect to the input points, making SetONet inherently robust to variations in the number and locations of sensors. SetONet learns richer, spatially-aware input representations by explicitly processing spatial coordinates and function values. We demonstrate SetONet's effectiveness on several benchmark problems, including derivative/anti-derivative operators, 1D Darcy flow, and 2D elasticity. Results show that SetONet successfully learns operators under variable input sampling conditions where standard DeepONet fails. Furthermore, SetONet is architecturally robust to sensor drop-off; unlike standard DeepONet, which requires methods like interpolation to function with missing data. Notably, SetONet can achieve comparable or improved accuracy over DeepONet on fixed grids, particularly for nonlinear problems, likely due to its enhanced input representation. SetONet provides a flexible and robust extension to the neural operator toolkit, significantly broadening the applicability of operator learning to problems with variable or incomplete input data.
Energy-efficient ventilation control plays a vital role in reducing building energy consumption while ensuring occupant health and comfort. While Computational Fluid Dynamics (CFD) simulations offer high-fidelity modeling of airflow for building HVAC design, their high computational cost makes them impractical for practical adoption in real-time building management system. In this work, we present a data-driven framework that combines the physical accuracy of CFD with the computational efficiency of machine learning to enable energy-efficient building ventilation control. Our method jointly optimizes airflow supply rates and vent angles to reduce energy use and adhere to air quality constraints. We train a neural operator transformer to learn the mapping from building control actions to airflow field distributions using high-resolution CFD data. This learned operator enables a gradient-based control framework capable of optimal decision-making. Experimental results demonstrate that our approach achieves substantial energy savings compared to maximum airflow rate control, rule-based control, and data-driven control based on regional average CO2 predictions, while consistently maintaining safe indoor air quality. These results highlight the practicality and scalability of our method for enabling safe and energy-efficient building management.
Koopman analysis of a general dynamics system provides a linear Koopman operator and an embedded eigenfunction space, enabling the application of standard techniques from linear analysis. However, in practice, deriving exact operators and mappings for the observable space is intractable, and deriving an approximation or expressive subset of these functions is challenging. Programmatic methods often rely on system-specific parameters and may scale poorly in both time and space, while learning-based approaches depend heavily on difficult-to-know hyperparameters, such as the dimension of the observable space. To address the limitations of both methods, we propose a hybrid framework that uses semidefinite programming to find a representation of the linear operator, then learns an approximate mapping into and out of the space that the operator propagates. This approach enables efficient learning of the operator and explicit mappings while reducing the need for specifying the unknown structure ahead of time.