Multiwavelet-based Operator Learning for

Differential Equations

Gaurav Gupta, Xiongye Xiao, Paul Bogdan

Ming Hsieh Department of Electrical and Computer Engineering
University of Southern California, Los Angeles, CA 90089
{ggaurav, xiongyex, pbogdan}@usc.edu

Abstract

The solution of a partial differential equation can be obtained by computing the

inverse operator map between the input and the solution space. Towards this end,
we introduce a multiwavelet-based neural operator learning scheme that com-
presses the associated operator’s kernel using fine-grained wavelets. By explicitly
embedding the inverse multiwavelet filters, we learn the projection of the kernel
onto fixed multiwavelet polynomial bases. The projected kernel is trained at mul-
tiple scales derived from using repeated computation of multiwavelet transform.
This allows learning the complex dependencies at various scales and results in
a resolution-independent scheme. Compare to the prior works, we exploit the
fundamental properties of the operator’s kernel which enable numerically efficient
representation. We perform experiments on the Korteweg-de Vries (KdV) equation,
Burgers’ equation, Darcy Flow, and Navier-Stokes equation. Compared with the
existing neural operator approaches, our model shows significantly higher accuracy
and achieves state-of-the-art in a range of datasets. For the time-varying equations,
the proposed method exhibits a (2X −10X) improvement (0.0018 (0.0033) relative
L2 error for Burgers’ (KdV) equation). By learning the mappings between function
spaces, the proposed method has the ability to find the solution of a high-resolution
input after learning from lower-resolution data.

1 Introduction
Many natural and human-built systems (e.g., aerospace, complex fluids, neuro-glia information
processing) exhibit complex dynamics characterized by partial differential equations (PDEs) [52, 60].
For example, the design of wings and airplanes robust to turbulence, requires to learn complex PDEs.
Along the same lines, complex fluids (gels, emulsions) are multiphasic materials characterized by a
macroscopic behavior [55] modeled by non-linear PDEs. Understanding their variations in viscosity
as a function of the shear rate is critical for many engineering projects. Moreover, modeling the
dynamics of continuous and discrete cyber and physical processes in complex cyber-physical systems
can be achieved through PDEs [68].
Recent efforts on learning PDEs (i.e., mappings between infinite-dimensional spaces of functions),
from trajectories of variables, focused on developing machine learning and in particular deep neural
networks (NNs) techniques. Towards this end, a stream of work aims at parameterizing the solution
map as deep NNs [2, 13, 33, 40, 71]. One issue, however, is that the NNs are tied to a specific
resolution during training, and therefore, may not generalize well to other resolutions, thus, requiring
retraining (and possible modifications of the model) for every set of discretizations. In parallel,
another stream of work focuses on constructing the PDE solution function as a NN architecture
[31, 42, 57, 65]. This approach, however, is designed to work with one instance of a PDE and,
therefore, upon changing the coefficients associated with the PDE, the model has to be re-trained.

35th Conference on Neural Information Processing Systems (NeurIPS 2021).

Additionally, the approach is not a complete data-dependent one, and hence, cannot be made oblivious
to the knowledge of the underlying PDE structure. Finally, the closest stream of work to the problem
we investigate is represented by the “Neural Operators" [14, 47, 48, 49, 56]. Being a complete
data-driven approach, the neural operators method aims at learning the operator map without having
knowledge of the underlying PDEs. The neural operators have also demonstrated the capability of
discretization-independence. Obtaining the data for learning the operator map could be prohibitively
expensive or time consuming (e.g., aircraft performance to different initial conditions). To be able to
better solve the problem of learning the PDE operators from scarce and noisy data, we would ideally
explore fundamental properties of the operators that have implications in data-efficient representation.
Our intuition is to transform the problem of learning a PDE to a domain where a compact representa-
tion of the operator exists. With a mild assumption regarding the smoothness of the operator’s kernel,
except finitely many singularities, the multiwavelets [5], with their vanishing moments property,
sparsify the kernel in their projection with respect to (w.r.t.) a measure. Therefore, learning an
operator kernel in the multiwavelet domain is feasible and data efficient. The wavelets have a rich
history in signal processing [24, 25], and are popular in audio, image compression [8, 61]. For
multiwavelets, the orthogonal polynomial (OP) w.r.t. a measure emerges as a natural basis for the
multiwavelet subspace, and an appropriate scale / shift provides a sequence of subspaces which
captures the locality at various resolutions. We generalize and exploit the multiwavelets concept to
work with arbitrary measures which opens-up new possibilities to design a series of models for the
operator learning from complex data streams.
We incorporate the multiwavelet filters derived using a variety of the OP basis into our operator
learning model, and show that the proposed architecture outperforms the existing neural operators.
Our main contributions are as follows: (i) Based on some fundamental properties of the integral
operator’s kernel, we develop a multiwavelet-based model which learns the operator map efficiently.
(ii) For the 1-D dataset of non-linear Korteweg-de Vries and Burgers equations, we observe an
order of magnitude improvement in the relative L2 error (Section 3.1, 3.3). (iii) We demonstrate
that the proposed model is in validation with the theoretical properties of the pseudo-differential
operator (Section 3.2). (iv) We show how the proposed multiwavelet-based model is robust towards
the fluctuation strength of the input signal (Section 3.1). (v) Next, we demonstrate the applicability
on higher dimensions of 2-D Darcy flow equation (Section 3.4), and finally show that the proposed
approach can learn at lower resolutions and generalize to higher resolutions. The code for reproducing
the experiments is available at: https://github.com/gaurav71531/mwt-operator.

2 Operator Learning using Multiwavelet Transform

We start by defining the problem of operator learning in Section 2.1. Section 2.2 defines the multi-
wavelet transform for the proposed operator learning problem and derives the necessary transformation
operations across different scales. Section 2.3 outlines the proposed operator learning model. Fi-
nally, Section 2.4 lists some of the useful properties of the operators which leads to an efficient
implementation of multiwavelet-based models.

2.1 Problem Setup

Given two functions a(x) and u(x) with x ∈ D, the operator is a map T such that T a = u. Formally,
let A and U be two Sobolev spaces Hs,p (s > 0, p ≥ 1), then the operator T is such that T : A → U.
The Sobolev spaces are particularly useful in the analysis of partial differential equations (PDEs),
and we restrict our attention to s > 0 and p = 2. Note that, for s = 0, the H0,p coincides with Lp ,
and, f ∈ H0,p does not necessarily have derivatives in Lp . We choose p = 2 in order to be able to
define projections with respect to (w.r.t.) measures µ in a Hilbert space structure.
We take the operator T as an integral operator with the kernel K : D × D → L2 such that
Z
T a(x) = K(x, y)a(y)dy. (1)
D

For the case of inhomogeneous linear PDEs, Lu = f , with f being the forcing function, L is the
differential operator, and the associated kernel is commonly termed as Green function. In our case,
we do not put the restriction of linearity on the operator. From eq. (1), it is apparent that learning the
complete kernel K(., .) would essentially solve the operator map problem, but it is not necessarily a

2
numerically feasible solution. Indeed, a better approach would be to exploit possible useful properties
(see Section 2.4) such that a compact representation of the kernel can be made. For an efficient
representation of the operator kernel, we need an appropriate subspace (or sequence of subspaces),
and projection tools to map to such spaces.
Norm with respect to measures: Projecting a given function onto a fixed basis would require a
measure dependent distance. For two functions f and g, we take the inner product w.r.t measure µ
R 1/2
as hf, giµ = f (x)g(x)dµ(x), and the associated norm as ||f ||µ = hf, f iµ . We now discuss the
next ingredient, which refers to the subspaces required to project the kernel.

2.2 Multiwavelet Transform

In this section, we briefly overview the concept of multiwavelets [4] and extend it to work with non-
uniform measures at each scale. The multiwavelet transform synergizes the advantages of orthogonal
polynomials (OPs) as well as the wavelets concepts, both of which have a rich history in the signal
processing. The properties of wavelet bases like (i) vanishing moments, and (ii) orthogonality
can effectively be used to create a system of coordinates in which a wide class of operators (see
Section 2.4) have a nice representation. Multiwavelets go few steps further, and provide a fine-grained
representation using OPs, but also act as a basis on a finite interval. For the rest of this section, we
restrict our attention to the interval [0, 1]; however, the transformation to any finite interval [a, b]
could be straightforwardly obtained by an appropriate shift and scale.
Multi Resolution Analysis: We begin by defining the space of piecewise polynomial functions,
S2n −1
for k ∈ N and n ∈ Z+ ∪ {0} as, Vkn = l=0 {f |deg(f ) < k for x ∈ (2−n l, 2−n (l + 1)) ∧
0, elsewhere}. Clearly, dim(Vnk ) = 2n k, and for subsequent n, each subspace is contained in
another as shown by the following relation:
V0k ⊂ V1k . . . ⊂ Vn−1
k
⊂ Vnk ⊂ . . . . (2)
Similarly, we define the sequence of measures µ0 , µ1 , . . . such that f ∈ Vnk is measurable w.r.t. µn
1/2 k
and the norm of f is taken as ||f || = hf, f iµn . Next, since Vn−1 ⊂ Vnk , we define the multiwavelet
k +
subspace as Wn for n ∈ Z ∪ {0}, such that
M
k
Vn+1 = Vnk Wnk , Vnk ⊥ Wnk . (3)

For a given OP basis for V0k as φ0 , φ1 , . . . , φk−1 w.r.t. measure µ0 , a basis of the subsequent spaces
Vnk , n > 1 can be obtained by shift and scale (hence the name, multi-scale) operations of the original
basis as follows:

φnjl (x) = 2n/2 φj (2n x − l), j = 0, 1, . . . , k − 1, l = 0, 1, . . . , 2n − 1, w.r.t. µn , (4)

where, µn is obtained as the collections of shift and scale of µ0 , accordingly.

Multiwavelets: For the multiwavelet subspace W0k , the orthonormal basis (of piecewise polynomials)
are taken as ψ0 , ψ1 , . . . , ψk−1 such that hψi , ψj iµ0 = 0 for i 6= j and 1, otherwise. From eq. (3),
Vnk ⊥ Wnk , and since Vnk spans the polynomials of degree at most k, therefore, we conclude that

Z1
xi ψj (x)dµ0 (x) = 0, ∀ 0 ≤ j, i < k. (vanishing moments) (5)
0

Similarly to eq. (4), a basis for multiwavelet subspace Wnk are obtained by shift and scale of ψi
n
as ψjl (x) = 2n/2 ψj (2n x − l) and ψjln n
are orthonormal w.r.t. measure µn , i.e. hψjl , ψjn0 l0 iµn = 1
0 0 k
if j = j , l = l , and 0 otherwise. Therefore, for a given OP basis for V0 (for example, Legendre,
Chebyshev polynomials), we only require to compute ψi , and a complete basis set at all the scales
can be obtained using scale/shift of φi , ψi .
Note: Since V1k = V0k
L k
W0 from eq. (3), therefore, for a given basis φi of V0k w.r.t. measure
n k
µ0 and φjl as a basis for V1 w.r.t. µ1 , a set of basis ψi can be obtained by applying Gram-Schmidt

3
(i) (ii) (iii)

coarse
reconstuct

decompose

fine

Figure 1: Multiwavelet representation of the Kernel. (i) Given kernel K(x, y) of an integral operator T ,
(ii) the bases with different measures (µ0 , µ1 ) at two different scales (coarse=0, fine=1) projects the kernel
into 3 components Ai , Bi , Ci . (iii) The decomposition yields a sparse structure, and the entries with absolute
magnitude values exceeding 1e−8 are shown in black. Given projections at any scale, the finer / coarser scale
projections can be obtained by reconstruction / decomposition using a fixed multiwavelet filters H (i) and
G(i) , i = 0, 1.

Orthogonalization using appropriate measures. We refer the reader to supplementary materials for
the detailed procedure.
Note: Since V0k and W0k lives in V1k , therefore, φi , ψi can be written as a linear combina-
tion of the basis of V1k . We term these linear coefficients as multiwavelet decomposition filters
(H (0) , H (1) , G(0) , G(1) ), since they are transforming a fine n = 1 to coarse scale n = 0. A uniform
measure (µ0 ) version is discussed in [4], and we extend it to any arbitrary measure by including the
correction terms Σ(0) and Σ(1) . We refer to supplementary materials for the complete details. The
capability of using the non-uniform measures enables us to apply the same approach to any OP basis
with finite domain, for example, Chebyshev, Gegenbauer, etc.
For a given f (x), the multiscale, multiwavelet coefficients at the scale n are defined as snl =
k×2n
[hf, φnil iµn ]k−1 n n k−1 n n
i=0 , dl = [hf, ψil iµn ]i=0 , respectively, w.r.t. measure µn with sl , dl ∈ R . The
decomposition / reconstruction across scales is written as

snl = H (0) sn+1

2l + H (1) sn+1
2l+1 , (6) sn+1
2l = Σ(0) (H (0) T snl + G(0) T dnl ), (8)
dnl = n+1
G(0) s2l + H (1) sn+1
2l+1 . (7) sn+1
2l+1 =Σ (1)
(H (1) T snl +G (1) T
dnl ). (9)
The wavelet (and also multiwavelet) transformation can be straightforwardly extended to multiple
dimensions using tensor product of the bases. For our purpose, a function f ∈ Rd has multiscale,
n
multiwavelet coefficients snl , dnl ∈ Rk×...×k×2 which are also recursively obtained by replacing the
filters in eq. (6)-(7) with their Kronecker product, specifically, H (0) with H (0) ⊗ H (0) ⊗ . . . H (0) ,
where ⊗ is the Kronecker product repeated d times. For eq. (8)-(9) H (0) Σ(0) (and similarly others)
are replaced with their d-times Kronecker product.
Non-Standard Form: The multiwavelet representation of the operator kernel K(x, y) can be ob-
tained by an appropriate tensor product of the multiscale and multiwavelet basis. One issue, however,
in this approach, is that the basis at various scales are coupled because of the tensor product. To
untangle the basis at various scales, we use a trick as proposed in [11] called the non-standard wavelet
representation. The extra mathematical price paid for the non-standard representation, actually serves
as a ground for reducing the proposed model complexity (see Section 2.3), thus, providing data
efficiency. For the operator under consideration T with integral kernel K(x, y), let us denote Tn
as the projection of T on Vnk , which essentially is obtained by projecting the kernel K onto basis

4
 +
+

decomposition reconstruction
Figure 2: MWT model architecture. (Left) Decomposition cell using 4 neural networks (NNs) A, B and
C, and T (for the coarsest scale L) performs multiwavelet decomposition from scale n + 1 to n. (Right)
Reconstruction module using pre-defined filters H (i) , G(i) performs inverse multiwavelet transform from scale
n − 1 to n.

φnjl w.r.t. measure µn . If Pn is the projection operator such that Pn f = j,l hf, φnjl iµn φnjl , then
P
Tn = Pn T Pn . Using telescopic sum, Tn is expanded as
Xn
Tn = (Qi T Qi + Qi T Pi−1 + Pi−1 T Qi ) + PL T PL , (10)
i=L+1

where, Qi = Pi − Pi−1 and L is the coarsest scale under consideration (L ≥ 0). From eq. (3), it is
apparent that Qi is the multiwavelet operator. Next, we denote Ai = Qi T Qi , Bi = Qi T Pi−1 , Ci =
Pi−1 T Qi , and T̄ = PL T PL . In Figure 1, we show the non-standard multiwavelet transform for a
given kernel K(x, y). The transformation has a sparse banded structure due to smoothness property
of the kernel (see Section 2.4). For the operator T such that T a = u, the map under multiwavelet
domain is written as
Udnl = An dnl + Bn snl , Uŝnl = Cn dnl , UsLl = T̄ sL
l , (11)
where, (Usnl , Udnl )/(snl , dnl ) are the multiscale, multiwavelet coefficients of u/a, respectively, and
L is the coarsest scale under consideration. With these mathematical concepts, we now proceed to
define our multiwavelet-based operator learning model in the Section 2.3.

2.3 Multiwavelet-based Model

Based on the discussion in Section 2.2, we propose a multiwavelet-based model (MWT) as shown in
Figure 2. For a given input/output as a/u, the goal of the MWT model is to map the multiwavelet-
transform of the input (sN N
l ) to output (Us l ) at the finest scale N . The model consists of two parts:
(i) Decomposition (dec), and (ii) Reconstruction (rec). The dec acts as a recurrent network, and at
each iteration the input is sn+1 . Using (6)-(7), the input is used to obtain multiscale and multiwavelet
coefficients at a coarser level sn and dn , respectively. Next, to compute the multiscale/multiwavelet
coefficients of the output u, we approximate the non-standard kernel decomposition from (11)
using four neural networks (NNs) A, B, C and T̄ such that Udnl ≈ AθA (dnl ) + BθB (snl ), Uŝnl ≈
CθC (dnl ), ∀ 0 ≤ n < L, and UsLl ≈ T̄θT̄ (sL l ). This is a ladder-down approach, and the dec part
performs the decimation of signal (factor 1/2), running for a maximum of L cycles, L < log2 (M ) for
a given input sequence of size M . Finally, the rec module collects the constituent terms Usnl , Uŝnl , Udnl
(obtained using the dec module) and performs a ladder-up operation to compute the multiscale
coefficients of the output at a finer scale n + 1 using (8)-(9). The iterations continue until the finest
scale N is obtained for the output.
At each iteration, the filters in dec module downsample the input, but compared to popular techniques
(e.g., maxpool), the input is only transformed to a coarser multiscale/multiwavelet space. By virtue
of its design, since the non-standard wavelet representation does not have inter-scale interactions, it
basically allows us to reuse the same kernel NNs A, B, C at different scales. A follow-up advantage
of this approach is that the model is resolution independent, since the recurrent structure of dec is
input invariant, and for a different input size M , only the number of iterations would possibly change
for a maximum of log2 M . The reuse of A, B, C by re-training at various scales also enable us to

5
learn an expressive model with fewer parameters (θA , θB , θC , θT̄ ). We see in Section 3, that even a
single-layered CNN for A, B, C is sufficient for learning the operator.
The dec / rec module uses the filter matrices which are fixed beforehand, therefore, this part does not
require any training. The model does not work for any arbitrary choice of fixed matrices H, G. We
show in Section 3.4 that for randomly selected matrices, the model does not learn, which validates
that careful construction of filter matrices is necessary.

2.4 Operators Properties

This section outlines definition of the integral kernels that are typically useful in an efficient com-
pression of the operators through multiwavelets. We then discuss a fundamental property of the
pseudo-differential operator.
Definition 1 ([54]). Calderón-Zygmund Operator. The integral operators that have kernel K(x, y)
which is smooth away from the diagonal, and satisfy the following.
1
|K(x, y)| ≤ ,
|x − y|
(12)
C0
|∂xM K(x, y)| + |∂yM K(x, y)| ≤ .
|x − y|M +1

The smooth functions with decaying derivatives are gold to the multiwavelet transform. Note that,
smoothness implies Taylor series expansion, and the multiwavelet transform with sufficiently large k
zeroes out the initial k terms of the expansion due to vanishing moments property (5). This is how
multiwavelet sparsifies the kernel (see Figure 1 where K(x, y) is smooth). Although, the definition
of Calderón-Zygmund is simple (singularities only at the diagonal), but the multiwavelets are capable
to compresses the kernel as long as the number of singularities are finite.
The next property, from [19], points out that with input/output being single-dimensional functions,
for any pseudo-differential operator (with smooth coefficients), the singularity at the diagonal is also
well-characterized.
Property 1. Smoothness of Pseudo-Differential Operator. For the integral kernel K(x,y) of a
pseudo-differential operator, K(x, y) ∈ C ∞ ∀x 6= y, and for x = y, K(x, y) ∈ C T −1 , where T + 1
is the highest derivative order in the given pseudo-differential equation.

The property 1 implies that, for the class of pseudo-differential operator, and any set of basis with
the initial J vanishing moments, the projection of kernel onto such bases will have the diagonal
dominating the non-diagonal entries, exponentially, if J > T − 1 [19]. For the case of multiwavelet
basis with k OPs, J = k (from eq. (5)). Therefore, k > T − 1 sparsifies the kernel projection onto
multiwavelets, for a fixed number of bits precision . We see the implication of the Property 1 on our
proposed model in the Section 3.2.

3 Empirical Evaluation
In this section, we evaluate the multiwavelet-based model (MWT) on several PDE datasets. We
show that the proposed MWT model not only exhibits orders of magnitude higher accuracy when
compared against the state-of-the-art (Sota) approaches but also works consistently well under
different input conditions without parameter tuning. From a numerical perspective, we take the data
as point-wise evaluations of the input and output functions. Specifically, we have the dataset (ai , ui )
with ai = a(xi ), ui = u(xi ) for x1 , x2 , . . . , xN ∈ D, where xi are M -point discretization of the
domain D. Unless stated otherwise, the training set is of size 1000 while test is of size 200.
Model architectures: Unless otherwise stated, the NNs A, B and C in the proposed model (Figure 2)
are chosen as a single-layered CNNs following a linear layer, while T̄ is taken as single k × k linear
layer. We choose k = 4 in all our experiments, and the OP basis as Legendre (Leg), Chebyshev (Chb)
with uniform, non-uniform measure µ0 , respectively. The model in Figure 2 is treated as single layer,
and for 1D equations, we cascade 2 multiwavelet layers, while for 2D dataset, we use a total 4 layers
with ReLU non-linearity.

6
 Networks s = 64 s = 128 s = 256 s = 512 s = 1024
MWT Leg 0.00338 0.00375 0.00418 0.00393 0.00389
MWT Chb 0.00715 0.00712 0.00604 0.00769 0.00675
FNO 0.0125 0.0124 0.0125 0.0122 0.0126
MGNO 0.1296 0.1515 0.1355 0.1345 0.1363
LNO 0.0429 0.0557 0.0414 0.0425 0.0447
GNO 0.0789 0.0760 0.0695 0.0699 0.0721
Table 1: Korteweg-de Vries (KdV) equation benchmarks for different input resolution s. Top: Our methods.
Bottom: previous works of Neural operator.

1 MWT Leg
1 FNO
true
0

u(x, 1)
u0(x)

1 1

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
x x
Figure 3: The output of the KdV equation. (Left) An input u0 (x) with λ = 0.02. (Right) The predicted
output of the MWT Leg model learning the high fluctuations.

From a mathematical viewpoint, the dec and rec modules in Figure 2 transform only the multiscale
and multiwavelet coefficients. However, the input and output to the model are point-wise function
samples, i.e., (ai , ui ). A remedy around this is to take the data sequence, and construct hypothetical
PN PN
functions fa = i=1 ai φnji and fu = i=1 ui φnji . Clearly, fa , fu lives in Vnk with n = log2 N .
(n)
Now the model can be used with s(n) = ai and Us = ui . Note that fa , fu are not explicitly used,
but only a matter of convention.
Benchmark models: We compare our MWT model using two different OP basis (Leg, Chb) with
the most recent successful neural operators. Specifically, we consider the graph neural operator
(GNO) [48], the multipole graph neural operator (MGNO) [49], the LNO which makes a low-rank
(r) representation of the operator kernel K(x, y) (also similar to unstacked DeepONet [50]), and
the Fourier neural operator (FNO ) [47]. We experiment on three competent datasets setup by the
work of FNO (Burgers’ equation (1-D), Darcy Flow (2-D), and Navier-Stokes equation (time-varying
2-D)). In addition, we also experiment with Korteweg-de Vries equation (1-D). For the 1-D cases, a
modified FNO with careful parameter selection and removal of Batch-normalization layers results in
a better performance compared with the original FNO, and we use it in our experiments. The MWT
model demonstrates the highest accuracy in all the experiments. The MWT model also shows the
ability to learn the function mapping through lower-resolution data, and able to generalize to higher
resolutions.
All the models (including ours) are trained for a total of 500 epochs using Adam optimizer with an
initial learning rate (LR) of 0.001. The LR decays after every 100 epochs with a factor of γ = 0.5.
The loss function is taken as relative L2 error [47]. All of the experiments are performed on a single
Nvidia V100 32 GB GPU, and the results are averaged over a total of 3 seeds.

3.1 Korteweg-de Vries (KdV) Equation

The Korteweg-de Vries (KdV) equation was first proposed by Boussinesq [16] and rediscovered by
Korteweg and de Vries [23]. KdV is a 1-D non-linear PDE commonly used to describe the non-linear
shallow water waves. For a given field u(x, t), the dynamics takes the following form:
∂u ∂u ∂ 3 u
= −0.5u − , x ∈ (0, 1), t ∈ (0, 1]
∂t ∂x ∂x3 (13)
u0 (x) = u(x, t = 0)

7
The task for the neural operator is to learn the mapping of the initial condition u0 (x) to the solutions
u(x, t = 1). We generate the initial condition in Gaussian random fields according to u0 ∼
N (0, 74 (−∆ + 72 I)−2.5 ) with periodic boundary conditions. The equation is numerically solved
using chebfun package [27] with a resolution 210 , and datasets with lower resolutions are obtained by
sub-sampling the highest resolution data set.
Varying resolution: The experimental results of the KdV equation for different input resolutions
s are shown in Table1. We see that, compared to any of the benchmarks, our proposed MWT Leg
exhibits the lowest relative error and is lowest nearly by an order of magnitude. Even in the case of the
resolution of 64, the relative error is low, which means that a sparse data set with a coarse resolution
of 64 is sufficient for the neural operator to learn the function mapping between infinite-dimensional
spaces.
Varying fluctuations: We now vary the smoothness
Leg:10 FNO:10
of the input function u0 (x, 0) by controlling the pa- Leg:12 FNO:12
1 Leg:14 FNO:14
rameter λ, where low values of λ imply more fre- 10 Leg:16 FNO:16
quent fluctuations and λ → 0 reaches the Brownian

L2 error
motion limit [30]. To isolate the importance of incor-
porating the multiwavelet transformation, we use the
10 2
same convolution operation as in FNO, i.e., Fourier
transform-based convolution with different modes
km (only single-layer) for A, B, C. We see in Fig-
ure 4 that MWT model consistently outperforms the 0.05 0.04 0.03 0.02
recent baselines for all the values of λ. A sample in-
put/output from test set is shown in the Figure 3. The
Figure 4: Comparing MWT by varying the degree
FNO model with higher values of km has better per- of fluctuations λ in the input with resolution s =
formance due to more Fourier bases for representing 1024. For each convolution, we fix the number of
the high-frequency signal, while MWT does better Fourier bases as km . For FNO, the width is 64.
even with low modes in its A, B, C CNNs, highlight-
ing the importance of using wavelet-based filters in the signal processing.

3.2 Theoretical Properties Validation

We test the ability of the proposed MWT model to capture the theoretical properties of the pseudo-
differential operator in this Section. Towards that, we consider the Euler-Bernoulli equation [62] that
models the vertical displacement of a finite length beam over time. A Fourier transform version of
the beam equation with the constraint of both ends being clamped is as follows
∂4u ∂u
4
− ω 2 u = f (x), x=1 =0
∂x ∂x x=0 (14)
u(0) = u(1) = 0,
where u(x) is the Fourier transform of the time-varying beam displacement, ω is the frequency,
f (x) is the applied force. The Euler-Bernoulli is a pseudo-differential equation with the maximum
derivative order T + 1 = 4. We take the task of learning the map from f to u. In Figure 5, we see that
for k ≥ 3, the models relative error across epochs is similar, however, they are different for k < 3,
which is in accordance with the Property 1. For k < 3, the multiwavelets will not be able to annihilate
the diagonal of the kernel which is C T −1 , hence, sparsification cannot occur, and the model learns
slow.

3.3 Burgers’ Equation

The 1-D Burgers’ equation is a non-linear PDE occurring in various areas of applied mathematics.
For a given field u(x, t) and diffusion coefficient v, the 1-D Burgers’ equation reads:
∂u ∂u ∂2u
= −u + v 2 , x ∈ (0, 2π), t ∈ (0, 1]
∂t ∂x ∂x (15)
u0 (x) = u(x, t = 0).
The task for the neural operator is to learn the mapping of initial condition u(x, t = 0) to the solutions
at t = 1 u(x, t = 1). To compare with many advanced neural operators under the same conditions,

8
 1
10
0
10
Relative error 1

Relative error
2
3
1 4 2 MWT Leg
10 5
6 10 MWT Chb
FNO
MGNO
LNO
GNO
2
10
1 100 200 300 400 500 10 3
epochs 256 2048 4096 8192
Resolution (s)
Figure 5: Relative L2 error vs epochs for MWT Leg
with different number of OP basis k = 1, . . . , 6. Figure 6: Burgers’ Equation validation at various input
resolution s. Our methods: MWT Leg, Chb.

Networks s = 32 s = 64 s = 128 s = 256 s = 512

MWT Leg 0.0152 0.00899 0.00747 0.00722 0.00654
MWT Chb 0.0174 0.0108 0.00872 0.00892 0.00891
MWT Rnd 0.2435 0.2434 0.2434 0.2431 0.2432
FNO 0.0177 0.0121 0.0111 0.0107 0.0106
MGNO 0.0501 0.0519 0.0547 0.0542 -
LNO 0.0524 0.0457 0.0453 0.0428 -
Table 2: Benchmarks on Darcy Flow equation at various input resolution s. Top: Our methods. MWT Rnd
instantiate random entries of the filter matrices in (6)-(9). Bottom: prior works on Neural operator.

we use the Burgers’ data and the results that have been published in [47] and [49]. The initial
condition is sampled as Gaussian random fields where u0 ∼ N (0, 54 (−∆ + 52 I)−2 ) with periodic
boundary conditions. ∆ is the Laplacian, meaning the initial conditions are sampled by sampling
its first several coefficients from a Gaussian distribution. In the Burgers’ equation, v is set to 0.1.
The equation is solved with resolution 213 , and the data with lower resolutions are obtained by
sub-sampling the highest resolution data set.
The results of the experiments on Burgers’ equation for different resolutions are shown in Figure 6.
Compared to any of the benchmarks, our MWT Leg obtains the lowest relative error, which is an
order of magnitude lower than the state-of-the-art. It’s worth noting that even in the case of low
resolution, MWT Leg still maintains a very low error rate, which shows its potential for learning the
function mapping through low-resolution data, that is, the ability to map between infinite-dimensional
spaces by learning a limited finite-dimensional spaces mapping.

3.4 Darcy Flow

Darcy flow formulated by Darcy[22] is one of the basic relationships of hydrogeology, describing the
flow of a fluid through a porous medium. We experiment on the steady-state of the 2-d Darcy flow
equation on the unit box, where it takes the following form:
∇ · (a(x)∇u(x)) = f (x), x ∈ (0, 1)2
(16)
u(x) = 0, x ∈ ∂(0, 1)2
We set the experiments to learn the operator mapping the coefficient a(x) to the solution u(x). The
coefficients are generated according to a ∼ N (0, (−∆ + 32 I)−2 ), where ∆ is the Laplacian with
zero Neumann boundary conditions. The threshold of a(x) is set to achieve ellipticity. The solutions
u(x) are obtained by using a 2nd-order finite difference scheme on a 512 × 512 grid. Data sets of
lower resolution are sub-sampled from the original data set.
The results of the experiments on Darcy Flow for different resolutions are shown in Table2. MWT
Leg again obtains the lowest relative error compared to other neural operators at various resolutions.

9
We also perform an additional experiment, in which the multiwavelet filters H (i) , G(i) , i = 0, 1 are
replaced with random values (properly normalized). We see in Table 2, that MWT Rnd does not learn
the operator map, in fact, its performance is worse than all the models. This signifies the importance
of the careful choice of the filter matrices.

3.5 Additional Experiments

Full results for these experiments are provided in the supplementary materials.
Navier Stokes Equation: The Navier-Stokes (NS) are 2d time-varying PDEs modeling the viscous,
incompressible fluids. The proposed MWT model does a 2d multiwavelet transform for the velocity
u, while uses a single-layered 3d convolution for A, B and C to learn dependencies across space-time.
We have observed that the proposed MWT Leg is in par with the Sota on the NS equations.
Prediction at high resolution: We show that MWT model trained at lower resolutions for various
datasets (for example, training with s = 256 for Burgers) can predict the output at finer resolutions
s = 2048, with relative error of 0.0226, thus eliminating the need for expensive sampling. The
training and testing with s = 2048 yields a relative error of 0.00189.
Train/evaluation with different sampling rules: We study the operator learning behavior when the
training and evaluation datasets are obtained using the random function from different generating
rules. The training is done with squared exponential kernel but evaluation is done on different
generating rule [30] with controllable parameter λ.

4 Conclusion
We address the problem of data-driven learning of the operator that maps between two function
spaces. Motivated from the fundamental properties of the integral kernel, we found that multiwavelets
constitute a natural basis to represent the kernel sparsely. After generalizing the multiwavelets to
work with arbitrary measures, we proposed a series of models to learn the integral operator. This
work opens up new research directions and possibilities toward designing efficient Neural operators
utilizing properties of the kernels, and the suitable basis. We anticipate that the study of this problem
will solve many engineering and biological problems such as aircraft wing design, complex fluids
dynamics, metamaterials design, cyber-physical systems, neuron-neuron interactions that are modeled
by complex PDEs.

Acknowledgement
We are thankful to the anonymous reviewers for providing their valuable feedback which improved
the manuscript. We would also like to thank Radu Balan for his valuable feedback. We grate-
fully acknowledge the support by the National Science Foundation Career award under Grant No.
CPS/CNS-1453860, the NSF award under Grant CCF-1837131, MCB-1936775, CNS-1932620, the
U.S. Army Research Office (ARO) under Grant No. W911NF-17-1-0076, the Okawa Foundation
award, and the Defense Advanced Research Projects Agency (DARPA) Young Faculty Award and
DARPA Director Award under Grant No. N66001-17-1-4044, an Intel faculty award and a Northrop
Grumman grant. A part of this work used the Extreme Science and Engineering Discovery Environ-
ment (XSEDE), which is supported by National Science Foundation grant number ACI-1548562.
The views, opinions, and/or findings contained in this article are those of the authors and should
not be interpreted as representing the official views or policies, either expressed or implied by the
Defense Advanced Research Projects Agency, the Army Research Office, the Department of Defense
or the National Science Foundation.

10
References
[1] M. Abramowitz and I.A. Stegun. Handbook of Mathematical Functions: With Formulas,
Graphs, and Mathematical Tables. Applied mathematics series. Dover Publications, 1965.
ISBN 9780486612720.
[2] Jonas Adler and Ozan Öktem. Solving ill-posed inverse problems using iterative deep neural
networks. Inverse Problems, 33(12), Nov 2017. ISSN 1361-6420. doi: 10.1088/1361-6420/
aa9581. URL http://dx.doi.org/10.1088/1361-6420/aa9581.
[3] B. Alpert, G. Beylkin, R. Coifman, and V. Rokhlin. Wavelet-like bases for the fast solution of
second-kind integral equations. SIAM Journal on Scientific Computing, 14(1):159–184, 1993.
doi: 10.1137/0914010.
[4] B. Alpert, G. Beylkin, D. Gines, and L. Vozovoi. Adaptive solution of partial differential
equations in multiwavelet bases. Journal of Computational Physics, 182(1):149–190, 2002.
ISSN 0021-9991.
[5] Bradley K. Alpert. A class of bases in L2 for the sparse representation of integral operators.
SIAM Journal on Mathematical Analysis, 24(1):246–262, 1993. doi: 10.1137/0524016.
[6] Bradley K. Alpert and Vladimir Rokhlin. A fast algorithm for the evaluation of legendre
expansions. SIAM Journal on Scientific and Statistical Computing, 12(1):158–179, 1991. doi:
10.1137/0912009.
[7] Kevin Amaratunga and John Williams. Wavelet based green’s function approach to 2d pdes.
Engineering Computations, 10, 07 2001. doi: 10.1108/eb023913.
[8] Radu Balan, Bernhard G. Bodmann, Peter G. Casazza, and Dan Edidin. Painless reconstruction
from magnitudes of frame coefficients. Journal of Fourier Analysis and Applications, 16:
488–501, 2009.
[9] J.J. Benedetto. Wavelets: Mathematics and Applications. Studies in Advanced Mathematics.
Taylor & Francis, 1993. ISBN 9780849382710.
[10] G. Beylkin. On the representation of operators in bases of compactly supported wavelets. SIAM
Journal on Numerical Analysis, 29(6):1716–1740, 1992. doi: 10.1137/0729097.
[11] G. Beylkin, R. Coifman, and V. Rokhlin. Fast wavelet transforms and numerical algorithms
i. Communications on Pure and Applied Mathematics, 44(2):141–183, 1991. doi: https:
//doi.org/10.1002/cpa.3160440202.
[12] Gregory Beylkłn and James M. Keiser. An adaptive pseudo-wavelet approach for solving
nonlinear partial differential equations. In Multiscale Wavelet Methods for Partial Differential
Equations, volume 6 of Wavelet Analysis and Its Applications, pages 137–197. Academic Press,
1997.
[13] Saakaar Bhatnagar, Yaser Afshar, Shaowu Pan, Karthik Duraisamy, and Shailendra Kaushik.
Prediction of aerodynamic flow fields using convolutional neural networks. Computational
Mechanics, 64(2):525–545, Jun 2019. ISSN 1432-0924. doi: 10.1007/s00466-019-01740-0.
[14] Kaushik Bhattacharya, Bamdad Hosseini, Nikola B. Kovachki, and Andrew M. Stuart. Model
reduction and neural networks for parametric pdes, 2020.
[15] Nicolas Boulle, Christopher J. Earls, and Alex Townsend. Data-driven discovery of physical
laws with human-understandable deep learning, 2021.
[16] Joseph Boussinesq. Essai sur la théorie des eaux courantes. Impr. nationale, 1877.
[17] R. Y. Chang and M. L. Wang. Shifted legendre direct method for variational problems. Journal
of Optimization Theory and Applications, 39(2):299–307, Feb 1983. ISSN 1573-2878. doi:
10.1007/BF00934535.
[18] Zhengdao Chen, Jianyu Zhang, Martin Arjovsky, and Léon Bottou. Symplectic recurrent
neural networks. In International Conference on Learning Representations, 2020. URL
https://openreview.net/forum?id=BkgYPREtPr.
[19] Kenneth C. Chou and Gary S. Guthart. Representation of green’s function integral operators
using wavelet transforms. Journal of Vibration and Control, 6(1):19–48, 2000. doi: 10.1177/
107754630000600102.

11
[20] N.M. Chuong. Pseudodifferential Operators and Wavelets over Real and p-adic Fields. Springer
International Publishing, 2018. ISBN 9783319774732.
[21] M. Cotronei, L.B. Montefusco, and L. Puccio. Multiwavelet analysis and signal processing.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 45(8):
970–987, 1998.
[22] Henry Darcy. Les fontaines publiques de la ville de Dijon: exposition et application... Victor
Dalmont, 1856.
[23] Olivier Darrigol. Worlds of flow: A history of hydrodynamics from the Bernoullis to Prandtl.
Oxford University Press, 2005.
[24] Ingrid Daubechies. Orthonormal bases of compactly supported wavelets. Communications
on Pure and Applied Mathematics, 41(7):909–996, 1988. doi: https://doi.org/10.1002/cpa.
3160410705.
[25] Ingrid Daubechies. Ten Lectures on Wavelets. Society for Industrial and Applied Mathematics,
1992. doi: 10.1137/1.9781611970104.
[26] D. Deng, Y. Meyer, and Y. Han. Harmonic Analysis on Spaces of Homogeneous Type. Lecture
Notes in Mathematics. Springer Berlin Heidelberg, 2008. ISBN 9783540887447.
[27] T. A Driscoll, N. Hale, and L. N. Trefethen. Chebfun Guide. Pafnuty Publications, 2014.
[28] Yuwei Fan, Jordi Feliu-Faba, Lin Lin, Lexing Ying, and Leonardo Zepeda-Nunez. A multiscale
neural network based on hierarchical nested bases, 2019.
[29] Jordi Feliu-Fabà, Yuwei Fan, and Lexing Ying. Meta-learning pseudo-differential operators
with deep neural networks. Journal of Computational Physics, 408:109309, May 2020. ISSN
0021-9991. doi: 10.1016/j.jcp.2020.109309.
[30] Silviu-Ioan Filip, Aurya Javeed, and Lloyd Trefethen. Smooth random functions, random odes,
and gaussian processes. SIAM Review, 61:185–205, 01 2019. doi: 10.1137/17M1161853.
[31] Daniel Greenfeld, Meirav Galun, Ron Kimmel, Irad Yavneh, and Ronen Basri. Learning to
optimize multigrid pde solvers, 2019.
[32] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory
with optimal polynomial projections. In Advances in Neural Information Processing Systems,
volume 33, pages 1474–1487, 2020.
[33] Xiaoxiao Guo, Wei Li, and Francesco Iorio. Convolutional neural networks for steady flow ap-
proximation. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’16, page 481–490. Association for Computing Machinery,
2016.
[34] Mustafa Hajij, Ghada Zamzmi, Matthew Dawson, and Greg Muller. Algebraically-informed
deep networks (aidn): A deep learning approach to represent algebraic structures, 2021.
[35] M.H. Heydari, M.R. Hooshmandasl, and F. Mohammadi. Legendre wavelets method for solving
fractional partial differential equations with dirichlet boundary conditions. Applied Mathematics
and Computation, 234:267–276, 2014. ISSN 0096-3003. doi: https://doi.org/10.1016/j.amc.
2014.02.047.
[36] Chyi Hwang and Yen-Ping Shih. Solution of integral equations via laguerre polynomials.
Computers & Electrical Engineering, 9(3):123–129, 1982. ISSN 0045-7906. doi: https:
//doi.org/10.1016/0045-7906(82)90018-0.
[37] Ameya D. Jagtap, Yeonjong Shin, Kenji Kawaguchi, and George Em Karniadakis. Deep
kronecker neural networks: A general framework for neural networks with adaptive activation
functions, 2021.
[38] M. Tavassoli Kajani, A. Hadi Vencheh, and M. Ghasemi. The chebyshev wavelets opera-
tional matrix of integration and product operation matrix. International Journal of Computer
Mathematics, 86(7):1118–1125, 2009. doi: 10.1080/00207160701736236.
[39] F. Khellat and S.A. Yousefi. The linear legendre mother wavelets operational matrix of inte-
gration and its application. Journal of the Franklin Institute, 343(2):181–190, 2006. ISSN
0016-0032. doi: https://doi.org/10.1016/j.jfranklin.2005.11.002.

12
[40] Yuehaw Khoo, Jianfeng Lu, and Lexing Ying. Solving parametric pde problems with artificial
neural networks. European Journal of Applied Mathematics, 32(3):421–435, Jul 2020. ISSN
1469-4425.
[41] Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential
equations for irregular time series, 2020.
[42] Dmitrii Kochkov, Jamie A. Smith, Ayya Alieva, Qing Wang, Michael P. Brenner, and Stephan
Hoyer. Machine learning–accelerated computational fluid dynamics. Proceedings of the Na-
tional Academy of Sciences, 118(21), 2021. ISSN 0027-8424. doi: 10.1073/pnas.2101784118.
[43] Vasil Kolev, Todor Cooklev, and Fritz Keinert. Design of a simple orthogonal multiwavelet filter
by matrix spectral factorization. Circuits, Systems, and Signal Processing, 39(4):2006–2041,
Aug 2019.
[44] Samuel Lanthaler, Siddhartha Mishra, and George Em Karniadakis. Error estimates for deep-
onets: A deep learning framework in infinite dimensions, 2021.
[45] Yuanlu LI. Solving a nonlinear fractional differential equation using chebyshev wavelets.
Communications in Nonlinear Science and Numerical Simulation, 15(9):2284–2292, 2010.
ISSN 1007-5704. doi: https://doi.org/10.1016/j.cnsns.2009.09.020.
[46] Yuanlu Li and Weiwei Zhao. Haar wavelet operational matrix of fractional order integration
and its applications in solving the fractional order differential equations. Applied Mathematics
and Computation, 216(8):2276–2285, 2010. ISSN 0096-3003. doi: https://doi.org/10.1016/j.
amc.2010.03.063.
[47] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya,
Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen-
tial equations, 2020.
[48] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya,
Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial
differential equations, 2020.
[49] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew Stuart, Kaushik
Bhattacharya, and Anima Anandkumar. Multipole graph neural operator for parametric partial
differential equations. In Advances in Neural Information Processing Systems, volume 33,
pages 6755–6766, 2020.
[50] Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for
identifying differential equations based on the universal approximation theorem of operators,
2020.
[51] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning
nonlinear operators via deeponet based on the universal approximation theorem of operators.
Nature Machine Intelligence, 3(3):218–229, Mar 2021. ISSN 2522-5839.
[52] Ryan McKeown, Rodolfo Ostilla-Mónico, Alain Pumir, Michael P. Brenner, and Shmuel M.
Rubinstein. Turbulence generation through an iterative cascade of the elliptical instability.
Science Advances, 6(9), 2020. doi: 10.1126/sciadv.aaz2717.
[53] Y. Meyer and D.H. Salinger. Wavelets and Operators: Volume 1. Cambridge Studies in
Advanced Mathematics. Cambridge University Press, 1992. ISBN 9780521458696.
[54] Y. Meyer, R. Coifman, and D. Salinger. Wavelets: Calderón-Zygmund and Multilinear Opera-
tors. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 1997. ISBN
9780521420013.
[55] Guillaume Ovarlez, Anh Vu Nguyen Le, Wilbert J. Smit, Abdoulaye Fall, Romain Mari,
Guillaume Chatté, and Annie Colin. Density waves in shear-thickening suspensions. Science
Advances, 6(16), 2020. doi: 10.1126/sciadv.aay5589.
[56] Ravi G. Patel, Nathaniel A. Trask, Mitchell A. Wood, and Eric C. Cyr. A physics-informed
operator regression framework for extracting data-driven continuum models. Computer Methods
in Applied Mechanics and Engineering, 373:113500, 2021. ISSN 0045-7825. doi: https:
//doi.org/10.1016/j.cma.2020.113500.
[57] M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep
learning framework for solving forward and inverse problems involving nonlinear partial

13
 differential equations. Journal of Computational Physics, 378:686–707, 2019. ISSN 0021-
9991.
[58] C.E. Rasmussen and C.K.I. Williams. Gaussian Processes for Machine Learning. Adaptative
computation and machine learning series. University Press Group Limited, 2006.
[59] Mohsen Razzaghi and Samira Yousefi. The legendre wavelets operational matrix of integration.
International Journal of Systems Science, 32:495–502, 04 2001. doi: 10.1080/00207720120227.
[60] M. F. Shlesinger, B. J. West, and J. Klafter. Lévy dynamics of enhanced diffusion: Application
to turbulence. Phys. Rev. Lett., 58:1100–1103, Mar 1987. doi: 10.1103/PhysRevLett.58.1100.
[61] B. W. Silverman, J. C. Vassilicos, and Nick Kingsbury. Image processing with complex wavelets.
Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical
and Engineering Sciences, 357(1760):2543–2560, 1999. doi: 10.1098/rsta.1999.0447.
[62] S. Timoshenko. History of Strength of Materials: With a Brief Account of the History of Theory
of Elasticity and Theory of Structures. Dover Civil and Mechanical Engineering Series. Dover
Publications, 1983. ISBN 9780486611877.
[63] J. Towns, T. Cockerill, M. Dahan, I. Foster, K. Gaither, A. Grimshaw, V. Hazlewood, S. Lathrop,
D. Lifka, G. D. Peterson, R. Roskies, J. R. Scott, and N. Wilkins-Diehr. Xsede: Accelerating
scientific discovery. Computing in Science & Engineering, 16(5):62–74, Sept.-Oct. 2014. ISSN
1521-9615. doi: 10.1109/MCSE.2014.80. URL doi.ieeecomputersociety.org/10.1109/
MCSE.2014.80.
[64] Lifeng Wang, Yunpeng Ma, and Zhijun Meng. Haar wavelet method for solving fractional
partial differential equations numerically. Applied Mathematics and Computation, 227:66–76,
2014. ISSN 0096-3003. doi: https://doi.org/10.1016/j.amc.2013.11.004.
[65] Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric
partial differential equations with physics-informed deeponets, 2021.
[66] Yanxin Wang and Qibin Fan. The second kind chebyshev wavelet method for solving fractional
differential equations. Applied Mathematics and Computation, 218(17):8592–8601, 2012. ISSN
0096-3003. doi: https://doi.org/10.1016/j.amc.2012.02.022.
[67] Shiying Xiong, Xingzhe He, Yunjin Tong, and Bo Zhu. Neural vortex method: from finite
lagrangian particles to infinite dimensional eulerian dynamics, 2020.
[68] Yuankun Xue and Paul Bogdan. Constructing compact causal mathematical models for complex
dynamics. In 2017 ACM/IEEE 8th International Conference on Cyber-Physical Systems
(ICCPS), pages 97–108, April 2017.
[69] Li Zhu and Yanxin Wang. Solving fractional partial differential equations by using the second
chebyshev wavelet operational matrix method. Nonlinear Dynamics, 89(3):1915–1925, Aug
2017. ISSN 1573-269X. doi: 10.1007/s11071-017-3561-7.
[70] Li Zhu and Yanxin Wang. Solving fractional partial differential equations by using the second
chebyshev wavelet operational matrix method. Nonlinear Dynamics, 89(3):1915–1925, Aug
2017. ISSN 1573-269X. doi: 10.1007/s11071-017-3561-7.
[71] Yinhao Zhu and Nicholas Zabaras. Bayesian deep convolutional encoder–decoder networks
for surrogate modeling and uncertainty quantification. Journal of Computational Physics, 366:
415–447, Aug 2018. ISSN 0021-9991.

14
Checklist
1. For all authors...
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s
contributions and scope? [Yes] For example, see Table 1 2, Figure 6 for benchmarks on
the datasets. Also, see Figure 3 for robustness plot, and Figure 5 for theoretical insights
for pseudo-differential operators.
(b) Did you describe the limitations of your work? [Yes] We discuss in details the possible
numerical issues that can occur in estimating the filter matrices H (0) , H (1) , G(0) , G(1)
for large values of k. The issue is not related to the mathematics involved but due to
the nature of floating-point precision. We discuss this in details in the supplementary
materials.
(c) Did you discuss any potential negative societal impacts of your work? [N/A]
(d) Have you read the ethics review guidelines and ensured that your paper conforms to
them? [Yes]
2. If you are including theoretical results...
(a) Did you state the full set of assumptions of all theoretical results? [Yes] Please refer to
Section 2.2, where we list all the necessary derived results, while we have referred the
reader (at appropriate places) to the supplementary materials for the complete details.
(b) Did you include complete proofs of all theoretical results? [Yes] We provide detailed
derivations of all the measure dependent filter matrices H (0) , H (1) , G(0) , G(1) and also
the correction terms Σ(0) , Σ(1) in the supplementary materials.
3. If you ran experiments...
(a) Did you include the code, data, and instructions needed to reproduce the main exper-
imental results (either in the supplemental material or as a URL)? [Yes] The code is
uploaded with the supplementary materials.
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they
were chosen)? [Yes] Please refer to Section 3 where we list all of the details regarding
training and model architectures.
(c) Did you report error bars (e.g., with respect to the random seed after running experi-
ments multiple times)? [Yes] All the results included in the paper are averaged over a
total of 3 seeds. We have also mentioned the same in Section 3.
(d) Did you include the total amount of compute and the type of resources used (e.g.,
type of GPUs, internal cluster, or cloud provider)? [Yes] All of the experiments were
performed on a single Nvidia V100 32 GB GPU, please refer to Section 3.
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
(a) If your work uses existing assets, did you cite the creators? [Yes] A part of the datasets
are taken from the FNO work [47], while some are generated using the scripts provided
by the same authors. We have properly cited the work in Section 3 Benchmark models.
(b) Did you mention the license of the assets? [N/A] A part of the code and datasets ([47])
used by us are openly available with no license restriction, to the best of our knowledge.
(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]

(d) Did you discuss whether and how consent was obtained from people whose data you’re
using/curating? [Yes] The code and the dataset is openly available.
(e) Did you discuss whether the data you are using/curating contains personally identifiable
information or offensive content? [N/A] All the datasets are synthetically generated.
5. If you used crowdsourcing or conducted research with human subjects...
(a) Did you include the full text of instructions given to participants and screenshots, if
applicable? [N/A]
(b) Did you describe any potential participant risks, with links to Institutional Review
Board (IRB) approvals, if applicable? [N/A]
(c) Did you include the estimated hourly wage paid to participants and the total amount
spent on participant compensation? [N/A]

15