Neural PDE operator learning

Especially forward operators. Image-to-image regression, where the images encode a physical process.

October 15, 2019 — October 19, 2023

dynamical systems
Hilbert space
how do science
Lévy processes
machine learning
neural nets
signal processing
stochastic processes
time series
Figure 1

Neural operators are a particular way of using statistical or machine learning approaches to solve PDEs, and maybe even to perform inference through them, by learning to predict one step ahead dynamics of the model.

We can learn a variety of operators (by which I mean function-to-function mappings) from data.

General notes about identifying forward operators are in recursive identification, notably the pushforward trick.

1 Background

NB: There are a lot of function spaces and operators in this section. I try making them simple and concrete here for the sake of sanity, but I promise everything can be made generic and abstract and fancy.

Recall the setup from the ML-for-PDEs page. We are not too far from that here i.e. the states of solutions of PDEs we are interested are functions, mapping some spatiotemporal coordinate space \(C\) to some vector-domain \(O\). Informally, we say the state of the system is a function \(f: C\to O.\) We assume there is a function space \(\mathcal{F}\) that contains all \(f\) of interest. We might also have a forcing/boundary function \(u: C\to O'\) from a possibly different function space, \(\mathcal{U}\).

The function values for all time are given in terms of solutions in terms differential operator \(\mathcal{D}_{\phi}\) for some parameter \(\phi\), \[\mathcal{D}_{\phi}[f]=u.\]

Here’s a concrete example: the Burgers’ equation \[ \begin{gathered} \partial_t f(x, t)+\partial_x\left(f^2(x, t) / 2\right)=\nu \partial_{x x} f(x, t), \quad x \in(0,1), t \in(0,1] \\ f(x, 0)=f_0(x), \quad x \in(0,1) \end{gathered} \] From this perspective \(f_0\) is a boundary condition and in fact we have sneakily introduced another one, which is that this has periodic boundary conditions on the spatial domain. In this one, \(C:= (0,1) \times (0,1].\) The internet is full of animations for the solution to this equation. Here is one:

So we could approximate that whole operator \(\mathcal{D}\). But often solving the whole thing for all time is not what we want to do, because we are running our algorithm online as data comes in, or because the whole solution will not fit in RAM, or something like that.

In this setting we might think about a different operator of interest. One that ends up being useful is the PDE propagator or forward operator \(\mathcal{P}_s[f(t, \cdot;\phi)]=f( t+s, \cdot)\) which produces a representation of the entire solution surface at some future moment, given current and boundary conditions, and some parameters \(\phi\). The forward propagator is what most methods in this page target.

Learning a PDE operator is naturally expressed by a layer which acts on functions, and which is indifferent to how those functions are discretized, and in particular the scale at which they are discretized. We learn to approximate \(\mathcal{P}_s[ f(t, cdot); \phi]\) from data by minimising prediction error over input/output pairs \(\left\{a_{j}, u_{j}\right\}_{j=1}^{N}\) where \(a_{j}\in \mathscr{A}\) and \(u_{j}=\mathcal{P}^{\dagger}_{1,\theta}(a_{j})+\epsilon_{j}\) where \(\epsilon{j}\) is a corrupting noise. Further, we observe these functions only at certain points \(\{x_{1},\dots,x_{n}\}\subset D.\) We approximate \(\mathcal{P}^{\dagger}_{1,\phi,\theta}\) by choosing good parameters \(\theta^{\dagger} \in \Theta\) from the parameter space in some parametric family of maps \(\{\mathcal{P}^{\dagger}_{1,\phi,\theta}\left(\cdot, \right);\theta\in\Theta\}\) so that \(\mathcal{P}^{\dagger}_{1,\theta,\phi^{\dagger}}\left(\cdot, \right)= \approx \mathcal{P}^{\dagger}_{1,\phi}\). Specifically we minimise some cost functional \(C: \mathscr{U} \times \mathscr{U} \rightarrow \mathbb{R}\) in the hope that we are solving \[ \min _{\theta \in \Theta} \mathbb{E}_{a \sim \mu}\left[C\left(\mathcal{P}_{1,\phi}(a), \mathcal{P}^{\dagger}_{1,\phi,\theta}(a)\right)\right]. \]

Figure 2: Training data for a neural operators: a series of timesteps generated by a fluid simulator.

2 Fourier neural operator

Zongyi Li blogs a neat trick: We use Fourier transforms to capture resolution-invariant and non-local behaviour in PDE forward-propagators. Essentially, we learn to approximate \(\mathcal{P}_1: \mathscr{A}\to\mathscr{A}.\) There is a bouquet of papers designed to leverage the trick (Li, Kovachki, Azizzadenesheli, Liu, Bhattacharya, et al. 2020b, 2020a; Li, Kovachki, Azizzadenesheli, Liu, Stuart, et al. 2020). See also Anima Anandkumar’s presentation, which phrases this in terms of Green’s functions and their relationship with Fourier transforms. See also nostalgebraist who has added some interesting commentary and also fixed up a typo in this page for me (Thanks!). Yannic Kilcher’s Fourier Neural Operator for Parametric Partial Differential Equations Explained is popular but I have not watched it. Code was at zongyi-li/fourier_neural_operator. Now generalized and professionalised at Neural Operators in PyTorch.

In most common application, Fourier Neural Operator is a method for approximating \(\mathcal{P}_{1,\phi}\).

Ok, so what is going on ?

For the neural Fourier operator in particular we assume that \(\mathcal{\mathcal{G}}_{\phi}\) has a particular iterative form, i.e. \(\mathcal{G}= Q \circ V_{T} \circ V_{T-1} \circ \ldots \circ V_{1} \circ P\). We introduce a new space \(\mathscr{V}=\mathscr{V}\left(D ; \mathbb{R}^{d_{v}}\right)\). \(P\) is a map \(\mathscr{A}\to\mathscr{V}\) and \(Q\) is a map \(\mathscr{V}\to\mathscr{U}\), and each \(V_{t}\) is a map \(\mathscr{V}\to\mathscr{V}\). Each of \(P\) and \(Q\) is ‘local’ in that they depend only upon pointwise evaluations of the function, e.g. for \(a\in\mathscr{A}\), \((Pa)(x)=p(a(x)\) for some \(p:\mathbb{R}^{d_{a}}\to\mathbb{R}^{d_{v}}\). Each \(v_{j}\) is function \(\mathbb{R}^{d_{v}}\). As a rule we are assuming \(d_{v}>d_{a}>d_{u}.\) \(V_{t}\) is not local. In fact, we define \[ (V_{t}v)(x):=\sigma\left(W v(x)+\left(\mathcal{K}(a ; \phi) v\right)(x)\right), \quad \forall x \in D \] where \(\mathcal{K}(\cdot;\phi): \mathscr{A} \rightarrow \mathcal{L}\left(\mathscr{V}, \mathscr{V}\right)\). This map is parameterized by \(\phi \in \Theta_{\mathcal{K}}\). \(W: \mathbb{R}^{d_{v}} \rightarrow \mathbb{R}^{d_{v}}\) is a linear transformation, and \(\sigma: \mathbb{R} \rightarrow \mathbb{R}\) is a local, component-wise, non-linear activation function. \(\mathcal{K}(a ; \phi)\) is a kernel integral transformation, by which is meant \[ \left(\mathcal{K}(a ; \phi) v\right)(x):=\int_{D} \kappa_{\phi}(x, a(x), y, a(y)) v(y) \mathrm{d} y, \quad \forall x \in D \] where \(\kappa_{\phi}: \mathbb{R}^{2\left(d+d_{a}\right)} \rightarrow \mathbb{R}^{d_{v} \times d_{v}}\) is some mapping parameterized by \(\phi \in \Theta_{\mathcal{K}}\).

Anyway, it looks like this:

Figure 3: Li’s Fourier neural operator layer.

Because this is neural net land we can be a certain type of sloppy and we don’t worry overmuch about the bandwidth of the signals we approximate; We keep some small number of low harmonics from the Fourier transform and use a fairly arbitrary non-Fourier maps too, and conduct the whole thing in a “high” dimensional space with vague relationship to the original problem without worrying too much about what it means.

2.1 Justification in terms of Fourier transform

The basic concept is heuristic and the main trick is doing lots of accounting of dimensions, and being ok with an arbitrary-dimensional Fourier transform \(\mathcal{F}:(D \rightarrow \mathbb{R}^{d}) \to (D \rightarrow \mathbb{C}^{d})\), which looks like this:

\[\begin{aligned} (\mathcal{F} f)_{j}(k) &=\int_{D} f_{j}(x) e^{-2 i \pi\langle x, k\rangle} \mathrm{d} x \\ \left(\mathcal{F}^{-1} f\right)_{j}(x) &=\int_{D} f_{j}(k) e^{2 i \pi\langle x, k\rangle} \mathrm{d} k \end{aligned}\] for each dimension \(j=1, \ldots, d\).

We notionally use the property that convolutions become multiplications under Fourier transform, which motivates the use of convolutions to construct our operators. Good.

2.2 Justification in terms of Green’s functions

Now what if \(\mathcal{D}\) is linear? Then we can write solutions in terms of convolutions of Greens functions \[ \begin{aligned} -\nabla \cdot(a(x) \nabla u(x)) & =f(x), & & x \in D \\ u(x) & =0, & & x \in \partial D \end{aligned} \] Inverse of differential operator can be written in form of kernel \[ u(x)=\int_D G_a(x, y) f(y) d y \] Where \(\mathrm{G}\) is the Green’s function (+ a correction term for non-zero boundaries)

To my mind this is very similar to the Fourier transform justification, but in a different language, because we can convolve functions by their Fourier transforms; see above.

We immediately throw out the dependence on \(a\) in the kernel definition replacing it with \[\kappa_{\phi}(x, a(x), y, a(y)) := \kappa_{R}(x-y)\] so that the integral operator becomes a convolution. This convolution can be calculated cheaply in Fourier space, which suggests we may as well define and calculate it also in Fourier space. Accordingly, the real work happens when they define the Fourier integral operator \[ \left(\mathcal{K}(\phi) v\right)(x)=\mathcal{F}^{-1}\left(R \cdot\left(\mathcal{F} v\right)\right)(x) \quad \forall x \in D \] where $R $ is the Fourier transform of a periodic function \(\kappa: D \rightarrow \mathbb{R}^{d_{v} \times d_{v}}\) parameterized by \(\phi \in \Theta_{\mathcal{K}}\). Checking our units, we have that \(\left(\mathcal{F} v\right):D \to \mathbb{C}^{d_{v}}\) and \(R (k): D \to \mathbb{C}^{d_{v} \times d_{v}}\). In practice, since we can work with a Fourier series rather than a continuous transform, we will choose \(k\in\{0,1,2,\dots,k_{\text{max}}\}\) and then $R $ can be represented by a tensor \(\mathbb{C}^{k_{\text{max}}\times d_{v}\times d_{v}}.\) 1 Not quite sure who is right. Caveat emptor. We can use a different \(W\) and \(R\) for each iteration if we want, say \(\{W_t,R_t\}_{1\leq t \leq T}\). So, the parameters of each of these, plus those of the maps \(P,Q\) comprise the parameters of the whole process.

Anyway, every step in this construction is differentiable in those parameters, and some of the steps can even be found rapidly using FFTs and so on, so we are done with the setup, and have an operator that can be learned from data.

2.3 Good stuff

Is this method any good? Yes. In the happy but very specific world where the PDE solutions are not too far from periodic, this model converges fast to an excellent approximation of the classical numerical solution, at a very low cost (Takamoto et al. 2022). Surprsingly (to me) it is also a really good estimator of the jacobian of the solution (MacKinlay et al. 2021). This is not magic — we still need to train the models— but even the cost of that is remarkably low, especially if we train over a fixed data set.

The question is: how do we handle things outside that very specific world?

2.4 Bad stuff

Quibble: They use the term ‘resolution-invariant’ loosely at a few points in the paper, and it takes some work to understand the actual defensible claim. They do not actually prove that things are resolution invariant per se. Rather, it is not even clear what that would specifically mean in this context — no attempt to prove Nyquist conditions or other sampling-theoretic properties over the spatial domain. And, AFAICT in the time domain their method is resolution-dependent in every sense. Step size is fixed.

What is clear and what I think they mean is that there is an obvious interpretation of the solution as a continuous operator, in the sense that it can be evaluated at arbitrary (spatial) points for the same computational cost as evaluating it at the training points. Thus there is a sense in which it does not depend upon the resolution of the training set, in that we don’t any resampling of the solution to evaluate the functions our operator produces at unobserved coordinates.

What does this get us? You can, in a certain sense, treat many network structures as discrete approximations to PDE operators with a certain resolution (at least, deep Resnets with ReLU activations have such an interpretation, presumably others) and then use resampling methods to evaluate them at a different resolution, which is a more laborious process that potentially gets a similar result — see the notebook on deep learning as dynamical system for examples of doing that.

Next quibble: Why Fourier transforms, instead of a different basis? Their rationale is that integral operators can be understood as continuous convolutions in the linear case, and therefore a bit like convolutions more generally. Heuristically, stacked continuous convolutions and non-linearities might well-approximate the operations of solving a nonlinear PDE. So convolutions are what we want, and Fourier transforms turn convolutions into multiplications so we can use them. It might sounds like we should use actual convolutional neural networks to solve the PDE, but that would impose a given resolution on the solution which is not what we want. For me, a better rationale is that this Fourier transform puts us in a “nice” basis for PDEs because Fourier transforms have well-understood interpretations which encode derivatives, translations, interpolations and other useful operations, which is why we use them in classic PDE solvers. The convolution thing is more “deep net-like” though.

Also Fourier transforms are notionally fast, which is another popular deep net rationale. Implementation detail: the authors do not make any particular effort to be maximally fast by using dilation to truncate the Fourier series. Instead they calculate the whole Fourier transform then throw some bits out. Presumably because they were already fast enough without bothering with being clever.

OTOH Fourier series also encode some troubling assumptions. Essentially these amount to an assumption that the basis functions are periodic, which in turn amounts to assuming that the domain of functions is toroidal, or, I suppose, a box with uniform boundary conditions. This is pretty common, but also annoying. Li, Kovachki, Azizzadenesheli, Liu, Bhattacharya, et al. (2020b) argue that does not matter because the edge effects can be more-or-less ignored and that things will still work OK because of the “local linear projection” part, which is … fine I guess? Maybe true. It still grates. Certainly the boundary conditions situation will just be messy AFAICT for any neural operator so I am at ease with the idea they would like to simply ignore it for now.

Would alternative bases fix that boundary problem, or just be more of an annoying PITA? Would there be anything to gain from learning a basis for the expansion? Certainly you would lose speed.

Also, this method uses a lot of the right buzzwords to sound like a kernel trick approach, and one can’t help but feel there might be a logical Gaussian process regression formulation with nice Bayesian interpretation. But I can’t see an obvious way of making that fly; the operator passes though a lot of nonlinear functions and thus the pushforward measure will get gnarly, not to mention the correlation kernel. I suppose it could possibly be approximated as ‘just’ another deep Gaussian process with some tactical assumptions, perhaps infinite-width asymptotics. Laplace approximation? Or a variational approximation? But what is a plausible variational family?

Update: See Kovachki, Lanthaler, and Mishra (2021) for some analysis of the performance of this method. TBD: read that mammoth paper.

2.5 Is FNO secretly a DeepONet?

Kinda sorta. See Kovachki, Lanthaler, and Mishra (2021), Lu et al. (2022).

2.6 Avoiding inverse transform

Poli et al. (2022) transforms once (how DO they recover the original domain? not obvious from skimming) rather than transforming inverse transforming. This is notionally faster.

2.7 Boundary conditions

Are hard to handle in FNOs. Many papers mention possible solutions. Liu, Yeo, and Lu (2020) just assumes that there are some initial conditions and boundaries can be “handled” by hoping that simulating a toroidal domain then cutting out a segment of interest. Possibly alternative transforms would help.

3 U-net


4 Active learning

Seems logical. But how much of that is there? Gajjar, Hegde, and Musco (2022).

5 Alternative basis transforms in spectral operators

The Fourier transform is generalised to the Laplace transform in Q. Cao, Goswami, and Karniadakis (2023).

Alternative transforms are important in the DeepONet lineages.

Learnable transforms are discussed in Meuris, Qadeer, and Stinis (2023); Witman et al. (2022).

6 DeepONet

From the people who brought us the PINN, comes Lu, Jin, and Karniadakis (2020). The setup is related to the PINNs, but AFAICT differs in a few ways in that

  1. we don’t (necessarily?) use the derivative information at the sensor locations
  2. we learn an operator mapping initial/latent conditions to output functions
  3. we decompose the input function space into a basis and them sample from the bases in order to span (in some sense) the input space at training time

The authors argue they have found a good topology for a network that does this and efficiently solves PDEs

A DeepONet consists of two sub-networks, one for encoding the input function at a fixed number of sensors \(x_i, i = 1, \dots, m\) (branch net), and another for encoding the locations for the output functions (trunk net).

This addresses some problems with generalisation that make the PINN setup seem unsatisfactory; in particular we can change the inputs, or project arbitrary inputs forward.

The boundary conditions and input points appear to stay fixed though, and inference of the unknowns is still vexed.



7 Recursive estimation

See recursive identification for generic theory of learning under the distribution shift induced by a moving parameter vector.

8 Conservation-lawsey-architectures

Lanthaler and Stuart (2023):

Neural operator architectures employ neural networks to approximate operators mapping between Banach spaces of functions; they may be used to accelerate model evaluations via emulation, or to discover models from data. Consequently, the methodology has received increasing attention over recent years, giving rise to the rapidly growing field of operator learning. The first contribution of this paper is to prove that for general classes of operators which are characterized only by their \(C^r\)- or Lipschitz-regularity, operator learning suffers from a curse of dimensionality, defined precisely here in terms of representations of the infinite-dimensional input and output function spaces. The result is applicable to a wide variety of existing neural operators, including PCA-Net, DeepONet and the FNO. The second contribution of the paper is to prove that the general curse of dimensionality can be overcome for solution operators defined by the Hamilton-Jacobi equation; this is achieved by leveraging additional structure in the underlying solution operator, going beyond regularity. To this end, a novel neural operator architecture is introduced, termed HJ-Net, which explicitly takes into account characteristic information of the underlying Hamiltonian system. Error and complexity estimates are derived for HJ-Net which show that this architecture can provably beat the curse of dimensionality related to the infinite-dimensional input and output function spaces.

Ruhe et al. (2023),Brandstetter et al. (2022):

Partial differential equations (PDEs) see widespread use in sciences and engineering to describe simulation of physical processes as scalar and vector fields interacting and coevolving over time. Due to the computationally expensive nature of their standard solution methods, neural PDE surrogates have become an active research topic to accelerate these simulations. However, current methods do not explicitly take into account the relationship between different fields and their internal components, which are often correlated. Viewing the time evolution of such correlated fields through the lens of multivector fields allows us to overcome these limitations. Multivector fields consist of scalar, vector, as well as higher-order components, such as bivectors and trivectors. Their algebraic properties, such as multiplication, addition and other arithmetic operations can be described by Clifford algebras. To our knowledge, this paper presents the first usage of such multivector representations together with Clifford convolutions and Clifford Fourier transforms in the context of deep learning. The resulting Clifford neural layers are universally applicable and will find direct use in the areas of fluid dynamics, weather forecasting, and the modeling of physical systems in general. We empirically evaluate the benefit of Clifford neural layers by replacing convolution and Fourier operations in common neural PDE surrogates by their Clifford counterparts on 2D Navier-Stokes and weather modeling tasks, as well as 3D Maxwell equations. For similar parameter count, Clifford neural layers consistently improve generalization capabilities of the tested neural PDE surrogates.

9 Incoming

Ilze Amanda Auzina’s tutorial DNN Tutorial 2 Part II: Physics inspired Machine Learning for Phillip Lippe’s Deep Learning course at the University of Amsterdam.

10 References

Bastek, Sun, and Kochmann. 2024. Physics-Informed Diffusion Models.”
Blechschmidt, and Ernst. 2021. Three Ways to Solve Partial Differential Equations with Neural Networks — A Review.” GAMM-Mitteilungen.
Bodnar, Bruinsma, Lucic, et al. 2024. Aurora: A Foundation Model of the Atmosphere.”
Brandstetter, Berg, Welling, et al. 2022. Clifford Neural Layers for PDE Modeling.” In.
Brandstetter, Worrall, and Welling. 2022. Message Passing Neural PDE Solvers.” In International Conference on Learning Representations.
Cao, Shuhao. 2021. Choose a Transformer: Fourier or Galerkin.” In Advances in Neural Information Processing Systems.
Cao, Qianying, Goswami, and Karniadakis. 2023. LNO: Laplace Neural Operator for Solving Differential Equations.”
Faroughi, Pawar, Fernandes, et al. 2023. Physics-Guided, Physics-Informed, and Physics-Encoded Neural Networks in Scientific Computing.”
Frerix, Kochkov, Smith, et al. 2021. Variational Data Assimilation with a Learned Inverse Observation Operator.” In.
Gajjar, Hegde, and Musco. 2022. Provable Active Learning of Neural Networks for Parametric PDEs.” In.
Ghattas, and Willcox. 2021. Learning Physics-Based Models from Data: Perspectives from Inverse Problems and Model Reduction.” Acta Numerica.
Gilpin. 2023. Model Scale Versus Domain Knowledge in Statistical Forecasting of Chaotic Systems.” Physical Review Research.
Goswami, Bora, Yu, et al. 2022. Physics-Informed Deep Neural Operator Networks.”
Grohs, and Herrmann. 2022. Deep Neural Network Approximation for High-Dimensional Elliptic PDEs with Boundary Conditions.” IMA Journal of Numerical Analysis.
Guibas, Mardani, Li, et al. 2021. Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers.”
Haitsiukevich, Poyraz, Marttinen, et al. 2024. Diffusion Models as Probabilistic Neural Operators for Recovering Unobserved States of Dynamical Systems.”
Holzschuh, Vegetti, and Thuerey. 2022. “Score Matching via Differentiable Physics.”
Jin, Zhang, and Espinosa. 2023. Recent Advances and Applications of Machine Learning in Experimental Solid Mechanics: A Review.”
Kadri, Duflos, Preux, et al. 2016. Operator-Valued Kernels for Learning from Functional Response Data.” The Journal of Machine Learning Research.
Karlbauer, Praditia, Otte, et al. 2022. Composing Partial Differential Equations with Physics-Aware Neural Networks.”
Kidger. 2022. On Neural Differential Equations.”
Kochkov, Sanchez-Gonzalez, Smith, et al. 2020. “Learning Latent FIeld Dynamics of PDEs.” In Machine Learning and the Physical Sciences Workshop at the 33rd Conference on Neural Information Processing Systems (NeurIPS).
Kossaifi, Kovachki, Azizzadenesheli, et al. 2023. Multi-Grid Tensorized Fourier Neural Operator for High-Resolution PDEs.”
Kovachki, Lanthaler, and Mishra. 2021. On Universal Approximation and Error Bounds for Fourier Neural Operators.” arXiv:2107.07562 [Cs, Math].
Kovachki, Li, Liu, et al. 2021. Neural Operator: Learning Maps Between Function Spaces.” In arXiv:2108.08481 [Cs, Math].
———, et al. 2023. Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs.” Journal of Machine Learning Research.
Lam, Sanchez-Gonzalez, Willson, et al. 2023a. GraphCast: Learning Skillful Medium-Range Global Weather Forecasting.”
———, et al. 2023b. Learning Skillful Medium-Range Global Weather Forecasting.” Science.
Lanthaler, and Stuart. 2023. The Curse of Dimensionality in Operator Learning.”
Li, Huang, Liu, et al. 2022. Fourier Neural Operator with Learned Deformations for PDEs on General Geometries.”
Li, Kovachki, Azizzadenesheli, Liu, Stuart, et al. 2020. Multipole Graph Neural Operator for Parametric Partial Differential Equations.” In Advances in Neural Information Processing Systems.
Li, Kovachki, Azizzadenesheli, Liu, Bhattacharya, et al. 2020a. Neural Operator: Graph Kernel Network for Partial Differential Equations.” In.
———, et al. 2020b. Fourier Neural Operator for Parametric Partial Differential Equations.” arXiv:2010.08895 [Cs, Math].
Liu, Yeo, and Lu. 2020. Statistical Modeling for Spatio-Temporal Data From Stochastic Convection-Diffusion Processes.” Journal of the American Statistical Association.
Li, Zheng, Kovachki, et al. 2021. Physics-Informed Neural Operator for Learning Partial Differential Equations.”
Long, Lu, Ma, et al. 2018. PDE-Net: Learning PDEs from Data.” In Proceedings of the 35th International Conference on Machine Learning.
Lu. 2020. “Theory, Algorithms, and Software for Physics-Informed Deep Learning.”
Lu, Jin, and Karniadakis. 2020. DeepONet: Learning Nonlinear Operators for Identifying Differential Equations Based on the Universal Approximation Theorem of Operators.” arXiv:1910.03193 [Cs, Stat].
Lu, Meng, Cai, et al. 2022. A Comprehensive and Fair Comparison of Two Neural Operators (with Practical Extensions) Based on FAIR Data.” Computer Methods in Applied Mechanics and Engineering.
Lu, Meng, Mao, et al. 2021. DeepXDE: A Deep Learning Library for Solving Differential Equations.” SIAM Review.
Luo, O’Leary-Roseberry, Chen, et al. 2023. Efficient PDE-Constrained Optimization Under High-Dimensional Uncertainty Using Derivative-Informed Neural Operators.”
MacKinlay, Pagendam, Kuhnert, et al. 2021. Model Inversion for Spatio-Temporal Processes Using the Fourier Neural Operator.” In.
Magnani, Krämer, Eschenhagen, et al. 2022. Approximate Bayesian Neural Operators: Uncertainty Quantification for Parametric PDEs.”
Meuris, Qadeer, and Stinis. 2023. Machine-Learning-Based Spectral Methods for Partial Differential Equations.” Scientific Reports.
Nüsken, and Richter. 2021. Solving High-Dimensional Hamilton–Jacobi–Bellman PDEs Using Neural Networks: Perspectives from the Theory of Controlled Diffusions and Measures on Path Space.” Partial Differential Equations and Applications.
Opschoor, Petersen, and Schwab. 2020. Deep ReLU Networks and High-Order Finite Element Methods.” Analysis and Applications.
Pathak, Subramanian, Harrington, et al. 2022. Fourcastnet: A Global Data-Driven High-Resolution Weather Model Using Adaptive Fourier Neural Operators.”
Pestourie, Mroueh, Rackauckas, et al. 2022. Physics-Enhanced Deep Surrogates for PDEs.”
Poli, Massaroli, Berto, et al. 2022. Transform Once: Efficient Operator Learning in Frequency Domain.” In Advances in Neural Information Processing Systems.
Psaros, Meng, Zou, et al. 2023. Uncertainty Quantification in Scientific Machine Learning: Methods, Metrics, and Comparisons.” Journal of Computational Physics.
Ronneberger, Fischer, and Brox. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation.” Edited by Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in Computer Science.
Ruhe, Gupta, de Keninck, et al. 2023. Geometric Clifford Algebra Networks.” In arXiv Preprint arXiv:2302.06594.
Saha, and Balamurugan. 2020. Learning with Operator-Valued Kernels in Reproducing Kernel Krein Spaces.” In Advances in Neural Information Processing Systems.
Takamoto, Praditia, Leiteritz, et al. 2022. PDEBench: An Extensive Benchmark for Scientific Machine Learning.” In.
Tran, Mathews, Xie, et al. 2022. Factorized Fourier Neural Operators.”
Wang, Kashinath, Mustafa, et al. 2020. Towards Physics-Informed Deep Learning for Turbulent Flow Prediction.” In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. KDD ’20.
Wen, Li, Azizzadenesheli, et al. 2022. U-FNO—An Enhanced Fourier Neural Operator-Based Deep-Learning Model for Multiphase Flow.” Advances in Water Resources.
Witman, New, Alkendry, et al. 2022. Neural Basis Functions for Accelerating Solutions to High Mach Euler Equations.”
Xu, and Darve. 2020. ADCME: Learning Spatially-Varying Physical Fields Using Deep Neural Networks.” In arXiv:2011.11955 [Cs, Math].
Yin, Yuan, Kirchmeyer, Franceschi, et al. 2023. Continuous PDE Dynamics Forecasting with Implicit Neural Representations.”
Yin, Minglang, Zhang, Yu, et al. 2022. Interfacing Finite Elements with Deep Neural Operators for Fast Multiscale Modeling of Mechanics Problems.” Computer Methods in Applied Mechanics and Engineering, A Special Issue in Honor of the Lifetime Achievements of J. Tinsley Oden,.


  1. NB — my calculations occasionally came out differing from the versions the authors gave in the paper with regards to the dimensionality of the spaces.↩︎