nature computational science

Article https://doi.org/10.1038/s43588-024-00732-2

A scalable framework for learning the

geometry-dependent solution operators of
partial differential equations

Received: 11 July 2024 Minglang Yin 1,2, Nicolas Charon3, Ryan Brody1,4, Lu Lu 5
,
Natalia Trayanova 1,2,4,6 & Mauro Maggioni 2,6,7
Accepted: 30 October 2024

Published online: xx xx xxxx

Solving partial differential equations (PDEs) using numerical methods is a
Check for updates ubiquitous task in engineering and medicine. However, the computational
costs can be prohibitively high when many-query evaluations of PDE
solutions on multiple geometries are needed. Here we aim to address
this challenge by introducing Diffeomorphic Mapping Operator
Learning (DIMON), a generic artificial intelligence framework that learns
geometry-dependent solution operators of different types of PDE on a
variety of geometries. We present several examples to demonstrate the
performance, efficiency and scalability of the framework in learning both
static and time-dependent PDEs on parameterized and non-parameterized
domains; these include solving the Laplace equations, reaction–diffusion
equations and a system of multiscale PDEs that characterize the electrical
propagation on thousands of personalized heart digital twins. DIMON can
reduce the computational costs of solution approximations on multiple
geometries from hours to seconds with substantially less computational
resources.

Solving partial differential equations (PDEs) with numerical methods traced back to at least the 1990s3,4, and has received much attention in
has a pivotal role in a wide variety of disciplines in science and engineer- the past decade due to the advance in computational ability for train-
ing, with applications such as topology and design optimization1 and ing deep neural networks5–8. In particular, using artificial intelligence
clinical prognostication2. Although effective, numerical methods can to learn the PDE solution operator, termed neural operator, is of par-
be computationally intensive or even prohibitive due to performing ticular interest as it enables approximating PDE solutions with varying
multiple independent PDE approximations on families of geometries boundary and initial conditions, and over parametric families of PDEs
and with multiple initial and boundary conditions. Thus, developing (for example, modeling local material properties9,10).
techniques that aim at improving efficiency by exploiting relationships Despite its success, the computational bottleneck when per-
between solutions over different geometries and initial and boundary forming tasks in optimization and prognostication still exists due
conditions is a task of substantial importance. to the inability of neural operators to estimate geometry-dependent
Within the emerging field of scientific machine learning, many PDE solutions. Most current frameworks for neural operators are for-
artificial intelligence approaches have been proposed to approximate mulated on a domain with fixed boundaries, and variations in the
the PDE solutions. The idea of solving PDEs with neural networks can be shape of the learning domain require retraining the neural network.

1
Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, USA. 2Alliance for Cardiovascular Diagnostic and Treatment Innovation,
Johns Hopkins University, Baltimore, MD, USA. 3Department of Mathematics, University of Houston, Houston, TX, USA. 4School of Medicine, Johns
Hopkins University, Baltimore, MD, USA. 5Department of Statistics and Data Science, Yale University, New Haven, CT, USA. 6Department of Applied
Mathematics and Statistics, Johns Hopkins University, Baltimore, MD, USA. 7Department of Mathematics, Johns Hopkins University, Baltimore, MD, USA.
e-mail: [email protected]; [email protected]

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

Previous attempts to extend the learning ability to various geometries Set-up. Consider a family {Ω θ }θ∈ϴ of bounded, open domains with
adopted approaches based on transformer architecture11–13, neural Lipschitz boundaries in ℝd , with Θ a compact subset of ℝp . d is the
fields14, graph encoding15 and convolutional neural networks16, among dimension of boundaries and p is the dimension of parameter set.
others17. The main challenge of learning geometry-dependent opera- Assume that for each shape parameter θ ∈ Θ there is a 𝒞𝒞 2 embedding
tors emanates from the scalability of learning on three-dimensional φθ from the ‘template’ Ω0 to ℝd , which induces a 𝒞𝒞 2 diffeomorphism
(3D) geometries with complex shapes. Furthermore, the generic formu- (that is a bijective 𝒞𝒞 2 map with 𝒞𝒞 2 inverse) between Ω0 and Ωθ. We further
lation and the approximation theorem of neural operators on a family assume that the map ϴ → 𝒞𝒞 2 (Ω 0 , ℝd ), defined by θ ↦ φθ is continuous,
of Banach spaces over various domains have not been established yet. when 𝒞𝒞 2 is endowed with the usual 𝒞𝒞 2 norm that makes it a Banach
Here we aim to address the aforementioned computational space. Assume that for each θ ∈ Θ there is a solution operator
challenges by proposing Diffeomorphic Mapping Operator Learn- 𝒢𝒢θ ∶ X1 (Ω θ ) × … Xm (Ω θ ) → Y(Ω θ ) mapping functions vθ1 , … , vθm on Ωθ
ing (DIMON), an operator learning framework that enables scalable or ∂Ωθ, representing a set of initial and boundary conditions, to the
geometry-dependent operator learning on a wide family of diffeomor- solution u θ(*; v θ) on Ω θ of a fixed PDE, where vθ ∶= (vθ1 , … , vθm ) .
phic domains. DIMON combines neural operators with diffeomorphic X0, X1 …, Xm, Y are Banach spaces.
mappings between domains and shapes, formulating the learning We may represent this family {𝒢𝒢θ }θ∈ϴ of operators defined on
problem on a unified reference domain (‘template’) Ω0, but with respect functions on different domains by an operator ℱ0 dependent on the
to a suitably modified PDE operator. PDE solutions along with input shape parameter θ and on functions that are all defined on the reference
functions on a family of diffeomorphically equivalent shapes are each domain Ω0, by suitable compositions with the diffeomorphisms φθ and
mapped onto the reference domain Ω0 via a diffeomorphism, which φ−1
θ
, as follows:
essentially transports functions from a family of Banach spaces to
θ θ
a single space and provides a unified framework for approximation uθ (∗; vθ ) = 𝒢𝒢θ (v
⏟1⎵ ,⎵…
⏟⎵,⎵v⏟
m)
analysis. This map can also be interpreted as a change of variables in vθ
the PDE on the original shape, yielding a corresponding PDE on Ω0,
= vθ1 ∘⎵φ
ℱ0 (θ, ⏟⎵ ⎵⎵ , … , v⎵θm
θ ⎵⏟⎵ ⎵⎵∘⎵⏟
φθ ) ∘ φ−1
parameterized by the diffeomorphism and geometry. Then, a neural θ
(1)
operator is trained to learn the latent solution operator of this param- vθ ∘φθ

eterized and geometry-dependent PDE on Ω0, where the parameter = ℱ0 (θ, v0θ,1 , … , v0θ,m ) ∘ φ−1 = u0θ (∗; v0θ ) ∘ φ−1 ,
that encodes the shape of the original domain represents the diffeo- ⏟⎵⎵⏟⎵⎵⏟ θ θ
v0
morphic mapping. We propose to compress the mapping in a suitable θ

way to reduce both the dimension of the input space of the network
and the computational cost of training the neural network, hence where the second equality defines ℱ0 , the third equality
facilitating the network scalability on learning from PDE systems on 3D defines v0k ∶= vθk ∘ φθ , and the fourth equality shows that
shapes. The original operator is equivalent to the composition of the u0θ (∗; v0θ ) ∶= uθ (∗; vθ ) ∘ φθ. The ‘latent’ geometry-dependent operator
learned modified operator on the reference domain with the inverse of ℱ0 ∶ ϴ × X1 (Ω 0 )× ⋯ × Xm (Ω 0 ) → Y(Ω 0 ) maps θ ∈ Θ and a set of
the diffeomorphism from the original domain. Thereby, learning the functions v0k in some Banach Spaces (Xk (Ω 0 ), || ∗ ||Xk (Ω 0 ) ), for k = 1, …, m,
geometry-dependent, parameterized PDE operators on the reference of functions on Ω0 or ∂Ω0, to a target function u0θ,v0 in a Banach Space
domain yields an efficient way of solving PDEs on a family of domains. (Y(Ω 0 ), || ∗ ||Y(Ω 0 ) ) of functions over Ω0. As above, we used the vector
The idea of learning geometry-dependent operators by combining notation v0 to denote (v01 , … , v0m ). One can view ℱ0 as the operator that
diffeomorphism with neural operators, albeit quite natural, is theoreti- outputs the pullback of a solution of a PDE on Ωθ to the reference domain
cally sound and flexible. To demonstrate DIMON’s utility, we further Ω0 via the diffeomorphism φθ; this is analogous to the Lagrangian form
extend the universal approximation theorem18 and show that the latent of the solution operator in contrast with the Eulerian form 𝒢𝒢θ defined
operator on Ω0 can be approximated by adopting a multiple-input on the varying domain using the terminology in ref. 22.
operator (MIONet). In addition, DIMON is flexible with the choice θ θ
of diffeomorphic mapping algorithms, reference domains Ω0 and Learning problem. Given N obser vations {(Ω θi , v1 i … , vmi ,
N
θ θ θ
approaches for geometry parameterization19–21, and it poses no con- 𝒢𝒢θi (v1 i , … , vmi ))}
and a diffeomorphism φθ, where θi and are all vki
i=1
straints on choosing the architecture of neural operators. Moreover, independently sampled at random from a probability measure on Θ
we demonstrate the learning ability and efficiency of DIMON with three and on compact subsets of Xk, respectively, we aim at constructing an
examples that include learning diverse types of static and dynamic PDE estimator of ℱ0 defined in equation (1).
problems on both parameterized and non-parameterized domains. At We choose the estimator in the form of a neural operator ℱ̃ 0, which
the inference stage, DIMON reduces the wall time needed for predicting takes as inputs the shape parameter θ, the transported input functions
the solution of a PDE on an unseen geometry by a factor of up to 10,000 v0θ , and outputs an approximation to ℱ0 (θ, v0θ ). We visualize the frame-
with substantially less computational resources and high accuracy. In work in Fig. 1a and the high-level architecture of the neural operator in
particular, the last example approximates functionals of PDE solutions Fig. 1b; we discuss more details of the neural operator and its approxi-
(in fact, PDEs couple with a rather high-dimensional ordinary differen- mation ability in the following section.
tial equations system) for determining electrical wave propagation The framework discussed above extends immediately to at least
on thousands of personalized heart digital twins, demonstrating its the following settings:
scalability and utility in learning on realistic 3D shapes and dynamics
in precision medicine applications. • Time-dependent operators, by concatenating time t to coordinate
x (see example 2).
Results • Operators that involve vector- or tensor-valued functions, by
Method overview replacing the pullback operation on scalar functions with an
This section provides an overview of DIMON. appropriate action of diffeomorphisms on the vector and tensor
fields. For example, the action of φθ on a vector field v ∶ Ω 0 → ℝd
Problem formulation. In this section, we present the formulation of is the vector field defined for all xθ = φθ(x) ∈ Ωθ by: Jφθ ⋅ v(x), where
DIMON for learning operators on a family of diffeomorphic domains, Jφθ denotes the Jacobian matrix of the transformation. This action
along with the unfolding of how a pedagogical example problem is can be further extended to covariant/contravariant tensor fields
set up and solved. of any type.

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

a vθ1 vθN (vθ1) (v θN)

b
θ1 θN
DIMON
Ωθ1 ΩθN

... θ
... θi

~
vθ0i 0 (θi, v 0θi )(x)
φ–1
θ φθ

Ω0 Ω0 x

... 0
... Geo. branch
Func. branches
Trunk
v0θ1 v0θN 0 (θ1, v0θ1) 0 (θN, v0θN )

Fig. 1 | Learning geometry-dependent solution operator of PDEs on a family (indicated by red marks and blue dots). There exists a latent operator ℱ0 such
of diffeomorphic domains with DIMON. a, A partial differential operator 𝒢𝒢θ is that 𝒢𝒢θ (vθ ) = ℱ0 (θ, v0θ ) ∘ φ−1
θ
. b, The operator ℱ0 is approximated by a neural
posed on Ω θi together with functions vθi (i = 1, … , N) as input. A map φθ, operator ℱ̃ 0: a shape parameter θ (typically estimated from data) and the
parameterized by the shape parameter θ, transports v0θ posed on reference transported function (or Func.) v0θ are the inputs of branch networks; the
domain Ω0 onto Ωθ and identifies point correspondences across all Ωθ and Ω0 coordinates of the reference domain x are the inputs of the trunk network.

Universal approximation theorem on a diffeomorphic family of simply connected domain for each θ, Ω0 may be chosen to be the open
domains. In this section, we extend the applicability of the universal unit disk {x ∈ ℝd ∶ ||x||2 < 1}, and φθ smooth (in fact, holomorphic if
approximation theorem (UAT) presented in ref. 18 on learning paramet- ℝ2 is identified with ℂ) bijection of Ωθ onto Ω0 exists by the Riemann
ric domains for multi-input operators under the DIMON framework. mapping theorem.
In short, the UAT in ref. 18 guarantees the approximation ability of a For each θ ∈ Θ, we are interested in obtaining the solution of the
multi-branch DeepONet for nonlinear and continuous operators with PDE
the assumption of fixed domains. Transporting functions defined on
Δu = 0 in Ω θ ,
various domains via φθ onto a reference domain with a fixed boundary { (2)
naturally aligns well with the assumption of UAT, hence enabling learn- g = vθ1 on ∂Ω θ .
ing on various domains. To differentiate the mapping φ and the shape 1
of Ωθ, we define a shape parameter θ ∈ Θ ⊂ X0 in a Banach space X0 and for vθ1 a function on ∂Ωθ with vθ1 ∈ H 2 (∂Ω θ ) , where Hs is the standard
introduce the branch for encoding geometric information, referred to Sobolev space of order s, and with the boundary conditions intended
as Geo. branch. Therefore, the adapted UAT can be written as: in the trace sense. In this setting, it is known that there exists a unique
weak solution uθ (∗; vθ1 ) ∈ H1 (Ω θ ) to this boundary problem, compare,
Theorem 1. (Adapted from ref. 18.) Assume that X0, X1 …, Xm, Y are for instance, in ref. 25 or ref. 26. We are interested in learning the solu-
Banach spaces, Θ ⊂ X0, Ki ⊂ Xi (i = 1, …, m) are compact sets, and Xi tion operator 𝒢𝒢θ (vθ1 ) ∶= uθ (∗; vθ1 ) for this Laplace equation over the
(i = 0, …, m) has a Schauder basis. If ℱ0 ∶ ϴ × K1 × ⋯ × Km → Y is a various domains Ωθ with boundary conditions vθ1 .
continuous operator, then for any ϵ > 0, there exists a neural network ℱ0̃ The operator ℱ0 to approximate maps a shape parameter θ and
with the form given in equation (8), such that boundary function vθ1 to the (weak) solution of equation (2) pulled back
on the template domain Ω0. Formally, denoting u0θ (∗; v0θ ) = ℱ0 (θ, v0θ ),
‖ℱ − ℱ ̃ ‖
‖ 0 0‖ < ϵ, where v0θ = vθ1 ∘ φθ , u0θ (∗; v0θ ) can be equivalently viewed as the solution
C(ϴ×K1 ×⋯×Km ,Y )
of a uniformly elliptic Dirichlet problem on Ω0, namely
∗
where the norm C is the sup norm, Ki are compact sets as shown in the φ∗θ Δ (φ−1
θ
) u0θ (∗; v0θ ) = 0 in Ω 0
theorem and ϵ is a positive number. { , (3)
g = v0θ on ∂Ω 0
Despite its generality, the use of Theorem 1 for the approximation
of solution operators of PDEs typically requires extra analysis, in par-
ticular to ensure that the assumptions on ℱ0 are satisfied, namely, the in which φ∗θ denotes the operator mapping an arbitrary function h on
well posedness and continuity with respect to the domain geometry Ωθ to the function h ∘ φθ on Ω0. Note that the specific coefficients of
∗
as well as initial and boundary conditions: this can be non-trivial for this linear second-order elliptic operator Lθ = φ∗θ Δ (φ−1 θ
) could be
many PDEs. While it is out of the scope of the present paper to discuss written explicitly based on the first and second-order derivatives of
this in a systematic fashion, we detail below and in Supplementary φθ, and are in particular continuous functions on Ω0, compare in Sup-
Information the particular case of the Laplace equation introduced plementary Section 1 for more details.
above, under suitable regularity assumptions. The key issue to apply Theorem 1 in this case is to determine ade-
Although we chose a particular MIONet as the baseline neural quate functional spaces for the solution and boundary function under
operator model in this work, DIMON is agnostic to the choice of which u0θ (⋅; v0θ ) is first well defined but also continuous with respect to
neural operators. This flexibility allows one to seamlessly integrate v0θ and θ. The existence and uniqueness of weak solutions in H1(Ω0) for
1
between previous endeavors in neural operators9,10,12,23,24 and DIMON smooth domains and v0θ ∈ H 2 (∂Ω 0 ) follow from classical results on
in a plug-and-play fashion. elliptic PDEs (Ch. 8 in ref. 25), and have been extended to Lipschitz
regular domains in ref. 26. Well-developed regularity estimates for this
Illustrative example. Let us consider the problem of solving the class of PDEs also guarantee the continuity of the solution with respect
Laplace equation on a family of bounded open domains {Ω θ }θ∈ϴ of ℝ2 to the boundary function v0θ . However, to our knowledge, regularity
with Lipschitz regular boundaries, where Θ is a compact subset of ℝp. with respect to domain deformations (and thus θ) is much less well
We assume that Ωθ is, for each θ ∈ Θ, diffeomorphic to a fixed domain understood in such a general setting. On the basis of the continuity of
Ω 0 ⊂ ℝ2 , which we call the reference domain. For example, if Ωθ is a the mapping θ ↦ φθ from Θ to 𝒞𝒞 2 (Ω 0 , ℝd ) and by adapting the ideas

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

presented in ref. 22, one can recover the solution’s continuity with domains, as well as map not only functions but also vector fields or
respect to θ albeit under the stronger assumption that Ω0 has a 𝒞𝒞 2 tensor fields from one domain to another. Another interesting feature
boundary; see Supplementary Section 1 for additional details. With of LDDMM is the fact that it is connected to a Riemannian metric on
such assumptions, the conclusion of Theorem 1 holds and it is thus the diffeomorphism group47. Thus, any optimal deformation path in
possible to approximate the solution operator up to any precision using that model can be interpreted as a geodesic and, as such, be repre-
the operator learning framework developed in this paper. sented compactly via some initial ‘momentum’ field defined on Ω0. The
Riemannian setting further enables the extension of many statistical
Algorithms and practical considerations. In practice, the domain techniques to the space of deformations, including methods such as
Ωθ may be available only in discretized form, and numerical methods principal component analysis (PCA) for dimensionality reduction, as
are used to approximate the solution of a PDE on the domain of Ωθ. we discuss next.
For all the examples presented in this work, we adopt a finite element In the perspective of learning the solution operator over a family
method (FEM) with a piece-wise linear basis to generate a training set of diffeomorphic domains, we also seek a representation of the defor-
of PDE solutions, consistent with the assumption uθ ∈ H1(Ωθ) on the mations φθ that can be effectively used as input for the operator
weak solution. One could adopt numerical methods with higher-order learning neural network. The estimator of the map φθ generally leads
polynomials, such as spectral element methods27,28 or hp-FEM29, to to high-dimensional objects, either a large set of point correspond-
approximate smoother PDE solutions with higher accuracy; note that ences between the mesh of Ω0 and Ωθ or as a set of momentum vectors
in this case, the differentiability of φθ needs to be sufficiently high to attached to vertices of Ω0 in the case of the LDDMM framework
preserve the smoothness solution after composition. described above. To obtain an informative representation encoded
by lower-dimensional shape parameter θ ,̃ we apply dimensionality
Choosing the reference domain Ω0. In general, there is no canonical reduction (herein PCA; see ‘Principal component analysis’ in Methods
choice for Ω0, or for the diffeomorphisms between Ω0 and Ωθ. Often for additional details). A first approach is to apply standard PCA to
the domains are unparameterized, given, for instance, meshes with no the set of displacement vectors {x − φθ (x)}x∈lndmrk(Ω 0 ) , where
explicit point correspondences between them. For such unparameter- lndmrk(Ω0) is a fixed set of landmark points in Ω0. Alternatively, when
ized domains, a practical choice is then to set a reference domain that deformations are estimated from the LDDMM model, we instead
has a simple shape and is diffeomorphically equivalent to all the Ωθs. leverage the geodesic PCA method48 on either the momentum or the
For instance, we choose a hollow semi-sphere as Ω0 as a simple shape deformation representation. In either case, projection onto the first
visually close to the left ventricle in example 3. In general, one may p′ principal components leads to our estimated low-dimensional
select one particular Ω θi , with θi ∈ Θ, as reference, or estimate a tem- representation θ ̃ of the shape parameter θ for each Ωθ, and gives a
plate shape, constrained to lie in {Ωθ}, from a given dataset of sample more compact representation of φθ. Notice that we adopt only
domains using one of the many possible approaches developed for reduced-order PCA coefficients to represent the shape of Ωθ for
this purpose including, among others, refs. 30–34. Although one can network training, but still rely on the originally (non-reduced) esti-
measure the similarity and distance between domains using a metric35, mated mappings φθ to transport the estimated solution on Ω0 back
that is, Chamfer distance36,37, Earth Movers’ distance38,39 or Riemannian to Ωθ. It is also noteworthy that a low-dimensional representation of
metrics on shape spaces21,40,41, this is not a prerequisite for successfully the mapping facilitates the network scalability as representing the
adopting this learning framework. This flexibility in choosing the refer- diffeomorphism and geometry inefficiently can lead to a substantial
ence domain facilitates a wide set of applications. expense of computational resources and lead to drastic reductions
in performance. We compare an efficient and an inefficient approach
Estimating and encoding domain geometry with shape parameter. to geometry encoding, or parameterization, in example 3 and Sup-
Once a reference domain Ω0 is chosen, the diffeomorphisms φθ that plementary Section 3.
map Ω0 to the domains Ωθ need to be constructed. In the case of a family
of parameterized shapes, as in the examples of ‘Example 1 solving the Example 1 solving the Laplace equation
Laplace equation’ below, the φθ can be constructed naturally by com- In this section, we present the network prediction for the pedagogical
posing parameterizations, which also yields point correspondences, example of the Laplace equation discussed above.
see ‘Domain generation’ in Methods.
For unparameterized data, however, such as the meshes in the PDE problem. The PDE problem on a domain Ωθ is
example of ‘Example 2 learning reaction–diffusion dynamics’ and
‘Example 3 predicting electrical wave propagation’, point correspond- Δu = 0, in Ω θ ,
{
ences are not a priori available and the mesh structure itself may not u|∂Ω θ = v ∼ GRF (0, kl ),
even be consistent from one data sample to another. In such cases, the
maps φθ are not known and need to be approximated for each domain.
A variety of models can be used for this purpose relying, for example, where for the boundary conditions we use the Gaussian random field
on maps that are area-preserving42, quasi-conformal43,44 or quasi- (GRF) with the covariance kernel:
isometric45, and which can be estimated, for instance, via the versatile
2
functional maps framework of refs. 19,46. Yet, in some applications, kl (x1 , x2 ) = exp(−‖x1 − x2 ‖ /2l2 ) , (4)
one may need a larger class of possible deformations, especially if larger
distortions occur between different domains. For that reason, in the Data preparation. The process of data preparation is illustrated in
present work, we choose to rather rely on the diffeomorphic registra- Fig. 2a. Following the steps detailed in ‘Domain generation’ in Meth-
tion setting introduced in ref. 21, known as Large Deformation Diffeo- ods, we first generate 3,500 samples of the shape parameters θ and
morphic Metric Mapping (LDDMM). This model has several notable create the corresponding domains Ω θi , which are deformations of
advantages when it comes to the applications we consider here. First, the unit square, which itself is set as the reference domain Ω0. We then
it generates transformations as flows of smooth vector fields that result impose a boundary condition sampled from a GRF with a length scale
in smooth diffeomorphisms of the whole ambient space ℝd , which is l = 0.5 and solve the governing equation using a finite element solver.
consistent with the theoretical setting developed in the previous sec- At the end of data preparation, we map the PDE solution from Ω θi to
tion. From these estimated deformations, one can easily recover Ω 0 for network training. The simulations are generated using
approximate point correspondences between Ω 0 and the other FEniCS49.

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

a Pointwise Mapped solution b

Meshing and deformation and PDE on reference ~
θ Geo. branch
correspondence parameterization solution domain

~ ~
1.0 1.0
ν BC branch 0
~
Case 1 θ1 0.
8

8
0.
5
0.

0.5

0.1

1
0.
1.0 1.0 x Trunk
…

d
0. 0.3
3
~ 0.
6 0.6
0
Case N θN
0.9 1
0.9
1.2
1.2
0 0.01 0.02 0.03 0.04
–3
0. Mean absolute error (εabs), (n = 197), εabs = 4.3 × 10
–3
1. Relative L2 error (εL2), (n = 197), εL2 = 8.3 × 10
c Testing case 1 Testing case 2 Testing case 3
e εL2
uNN
0.1 0.3 –0.20 100
Test error
Train error
–0.2 0 –0.85
10–1

–0.5 –0.3 –1.50

|uFEM – uNN| 10–2
0.02 0.02 0.02

0.01 0.01 0.01 10–3

0 0 0 64 128 256 512 1,024 3,300

Number of training cases

Fig. 2 | DIMON for learning solution operator of the Laplace equation on We adopt a fully connected neural network as the underlying structure in each
parameterized two-dimensional domains. a, Data generation on N parametric branch. c, Network predictions (uNN) and their absolute errors with numerical
two-dimensional domains. Landmark points are matched across two- solutions (∣uFEM − uNN∣) on three testing domains. The color bars present the
′
dimensional domains with different meshing. Shape parameter θ ̃ ∈ ℝp is magnitude of solutions and prediction errors. d, Distributions of the mean
calculated by PCA on the deformation field at the landmark points. Here we absolute error εabs, and of the relative L2 error (εL2) on the testing dataset with the
adopt the first 15 principal components as θ .̃ The PDE is solved using finite corresponding means (vertical black lines). The testing dataset contains n = 197
elements and mapped onto the reference domain. b, Neural operator structure. ̄ ) with standard deviation for the training (red) and
cases. e, The average of εL2 ( εL2
Shape parameter θ ̃ and boundary condition v ̃ are fed into the geometry (Geo.) testing dataset (blue) is plotted against the number of training cases. Data are
and boundary condition (BC) branch, respectively. Coordinates in the reference presented as mean ± s.d.
domain x are the input of the trunk net. ⨂ represents pointwise multiplication.

Construction of φθ and θ .̃ We obtain a set of point correspondences shows network predictions uNN and the absolute pointwise errors
between 11 × 11 uniformly distributed landmarks on Ω0 to their images ∣uFEM(x) − uNN(x)∣ on three representative domains in the testing cases,
in Ω θi , using the map φθi , which is considered known in this example indicating that network predictions are in good agreement with the
(see ‘Domain generation’ in Methods). We visualize the deformation numerical solutions. The maximum value of the absolute pointwise
via vector fields in the second column of Fig. 2a, where the length error is less than 0.02. We summarize the statistics of the mean absolute
of the arrows corresponds to the magnitude of the deformation. pointwise error εabs and relative L2 error εL2 ≔ ∣∣uFEM − uNN∣∣2/∣∣uFEM∣∣2 in
Next, with the identified point correspondence and deformation field, the testing cases by presenting their distributions in Fig. 2d. Both
each domain Ω θi can be parameterized using truncating principal distributions are left-skewed. The average εabs, denoted by εabs, and the
component coefficients in PCA, yielding the low-dimensional shape mean relative L2 error εL2 are 4.3 × 10−3 and 8.3 × 10−3, respectively.
parameter θ.̃ Although the diffeomorphisms in this example are explic- Overall, these distributions show that both measures of error are well
itly known, we intentionally use PCA to parameterize the shape to distributed around and below their mean values, and only in a few
demonstrate the generalization of our framework when the diffeomor- extreme cases they are close to their maximum value. The behavior of
phisms are unknown. A comparison of network performance using θ train and test error as a function of the number of training cases is
and θ ̃ can be found in Supplementary Section 4. ̄ of the testing dataset decreases to the same
represented in Fig. 2e. εL2
order of training error with an increasing number of training cases,
Network training and performance. Figure 2b illustrates the network from O(100) to O(102). On the basis of the above results, we conclude
architecture for approximating the latent operator on the reference that our framework enables effective learning of the solution operator
domain for this example. As the new operator is a function of shape for the Laplace equation on a diffeomorphic family of domains.
parameters and boundary conditions, the network has two branch Recall that, in the case of the Laplace equation considered here,
networks for inputting the truncated shape parameter θ ̃ and boundary we proved that sufficient conditions for Theorem 1 to hold included
values represented by a set of discrete points, denoted as v .̃ The trunk that the domains have 𝒞𝒞 2 boundaries (see the discussion after
network takes the coordinates x on the reference shape as input. We Theorem 1, and in Supplementary Section 1). We have elected here to
train the network on n = 3,303 training cases (see Supplementary Sec- consider domains with corners, to demonstrate that at least in some
tion 6 for details) and use the remaining n = 197 cases for testing. cases our approach appears to be robust to some violations of such
Network predictions in Ω0 are then mapped back to the original domain regularity assumptions sufficient, but perhaps not necessary, for
via φ θ for further visualization and error analysis. Figure 2c Theorem 1. Note that in this case, the deformations ‘preserve’ corners,

Nature Computational Science

 Article https://doi.org/10.1038/s43588-024-00732-2

a Pointwise deformation Mapped solution on b

Meshing and parameterization PDE solution reference domain ~
θ Geo. branch

~ ~
Case 1 ν IC branch
0

x, t Trunk
…

d Relative L2 error (εL2) distribution (n = 297)

Case N

0
c 1
0.7 0.050
Testing case 1

0 0.025 3
0 0.1 0.2 0.3
–0.7 0
~
Axis φ θ εL2
0.7
Testing case 2

0.050
0 Affine Parameter 3.2 × 10–2
0 0.025
1 LDDMM PCA (momentum) 4.3 × 10–2

–0.7 0 2 LDDMM Parameter 5.5 × 10–2

Testing case 3

0 0.050 3 LDDMM PCA (deformation) 7.5 × 10–2

–0.35 0.025

t –0.70 t 0

Fig. 3 | DIMON for learning solution operator of the reaction–diffusion network as the underlying structure in each branch. c, Network predictions with
equation on parametric two-dimensional domains. a, Data preparation on N true shape parameters of the affine transformation at t = 0, 1 s and 2 s and their
parametric domains. Each domain is parameterized by either the deformation or absolute errors on three testing domains. The color bars present the magnitude
the momentum field at the landmark points with p′ principal component of solutions and prediction errors. d, Distributions of the relative L2 error (εL2)
coefficients. A finite element method is adopted to numerically solve the obtained from four geometry parameterization approaches on the testing
reaction–diffusion equation with various initial conditions (IC) on Ω θi , set as an dataset with the corresponding means. The four approaches are affine with true
annulus, and map the solution to Ω0 at different time snapshots. b, Neural shape parameters, LDDMM with PCA on momentum and deformation, and
operator structure. Truncated shape parameter θ ̃ and initial condition v ̃ are fed LDDMM with true shape parameters. The corresponding means (vertical black
into the geometry (Geo.) and initial condition (IC) branch, respectively. lines) are shown in the table. The number of testing cases is n = 297. Affine, affine
Coordinates in the reference domain x are the input of the trunk net. polar transformation; Parameter, true shape parameter.
⨂ represents pointwise multiplication. We adopt a fully connected neural

in a well-conditioned way, and do not create new corners, which may θ following the steps in ‘Domain generation’ in Methods. Next, we
explain the good generalization power of our approach, notwithstand- simulate the system dynamics from t = 0 to t = 2 with a finite element
ing the violation of regularity assumptions on the boundary of the method and map the numerical solution at different snapshots to the
domains. reference domain.

Example 2 learning reaction–diffusion dynamics Construction of φθ and θ .̃ To demonstrate the compatibility of our
In this section, we consider a PDE that includes time and is nonlinear. framework with generic methods of domain mapping, we adopt two
approaches for estimating point correspondences between Ω θi and
PDE problem. We now consider a PDE that includes time and is non- Ω0 (Fig. 3a): an affine polar transformation (referred to as ‘affine’ for
linear: a reaction–diffusion system in the form the rest of this section) and LDDMM, with details provided in details in
‘Large deformation diffeomorphism metric mapping’ in Methods. The
∂u
= Δu + R(u) , with R(u) = u(1 − u2 ), in Ω θ former approach converts the Cartesian coordinates of Ω θi into polar
∂t
∂u |
coordinates with the radius normalized to [0.5, 1.0]. These affine trans-
u|t=0 = v, = 0. formations are less general than LDDMM, as they are tailored to the
∂n |∂Ω θ
particular shapes of Ω0 and Ω θi . With the identified point correspond-
where the Ωθs are elongated elliptical domains with a central hole, and ence, we proceed to calculate the truncated shape parameters θ ̃ using
Ω0 is a circular annulus. The reaction term R(u) is adopted from the PCA based on (1) deformation from LDDMM and (2) momentum from
Newell–Whitehead–Segel model50 to describe post-critical Rayleigh– LDDMM.
Bénard convection. We impose a Neumann boundary condition and
impose initial conditions by sampling from a GRF. Network training and performance. Network training in this work
follows the steps presented in Supplementary Algorithm 1. Figure 3b
Data preparation. We visualize the data generation and learning shows the architecture of the network. The network contains a branch
process in Fig. 3a: we generate N = 6,000 sets of shape parameters for encoding the truncated shape parameter θ ̃ and one for the initial

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

a Training data preparation (n = 1,006) c Network structure

~
θ Geo. branch

xpace ~
Pacing location

x Trunk
Cardiac magnetic Meshing and fiber ATs RTs
resonance generation

b Pacing locations and LV segments d UVC

Anterior

0
3
Septum

Lateral
4 2
6
5
1
Inferior
Pacing points on mesh and AHA bull’s eye Apicobasal ρ Rotational φ Transmural δ
LV segments (0–6) x := [ρ, δ, φ] ε [0, 1] × [0, 1] × [0, 2π]

Fig. 4 | Predicting electrical signal propagation on patient-specific LVs. located on the endocardium with reduced American Heart Association (AHA)
a, Data preparation for the network training. The cohort assembled consists LV segments from zones 0 to 6. c, Network structure. ⨂ represents pointwise
of 1,006 CMR scans of patients from Johns Hopkins Hospital, based on which multiplication. We adopt a fully connected neural network as the underlying
LV meshes and fiber fields are generated. Next, we simulate a paced beat for structure in each branch. d, UVC calculation for registering points across
600 ms at an endocardial site until the LV is fully repolarized and record the geometries.
ATs and the RTs for network training. b, Pacing locations (marked in white balls)

conditions f. x and t are the coordinates in Ω0 and time. We train a parameters and the underlying ordinary differential equations in
network with 5,703 training cases and examine its performance with refs. 51,52. We add the cellular model downloaded from the CellML
297 testing cases. Figure 3c shows the predicted solutions and the repository to our data repository. More details can be found in ‘Data
absolute errors for three representative testing cases at t = 0, 1 s and availability’. It is important to note that the system is essentially ani-
2 s. θ ̃ are adopted as the true shape parameters. Qualitatively, the sotropic and multiscale.
network predictions closely match the finite element solutions with
low absolute errors. Figure 3d illustrates the relative L2 error distribu- Data preparation. Figure 4a provides an overview of training data gen-
tion calculated over the 297 testing cases from various point corre- eration. We collected 1,006 anonymized cardiac magnetic resonance
spondence algorithms, including (0) affine with true parameter and (CMR) scans from patients at Johns Hopkins Hospital. However, our
(1)–(3) LDDMM with PCA on momentum, true parameter and PCA on goal in this work is not to predict the full solution of this multiscale
deformation, where the true parameters refer to the shape parameters system at all times; we focus on two clinically relevant quantities,
in ‘Domain generation’ in Methods. All of the distributions show left the distributions of activation times (ATs) and repolarization times
skewness with means (black line) ranging from 3.2 × 10−2 to 7.5 × 10−2. (RTs) of a propagating electrical signal, defined as the time when local
Mean errors from affine transformation are smaller than those of myocardium tissue is depolarized (activated) and repolarized. For
LDDMM due to the more uniform distribution of landmarks in Ω θi and each LV, we apply an external electrical stimulation (pacing) at seven
no matching error between Ω0 and Ω θi . A comparison of affine trans- myocardium points (Fig. 4b) in each segment and record the local ATs
formation and LDDMM is presented in Supplementary Section 5. Nev- and RTs for each elicited wave (visualized in Fig. 4a) for network train-
ertheless, the above results indicate that DIMON can accurately learn ing and testing. We provide more details of data generation in ‘Cohort
the reaction–diffusion dynamics on the diffeomorphic family of and simulations details in example 3’ in Methods.
domains with a flexible choice of point-matching algorithms.
Construction of φθ and θ .̃ We choose a hollow semi-sphere with an
Example 3 predicting electrical wave propagation inner radius of 0.5 and an outer radius of 1 as Ω0 for LVs as they are dif-
In this section, we demonstrate the capabilities and scalability of our feomorphically equivalent. To calculate point correspondence
framework for predicting electrical signal propagation on the left between these non-rigid and unparameterized LV meshes with Ω0, we
ventricles (LVs) reconstructed from patient-specific imaging scans. adopt a registration method by matching the universal ventricular
coordinates (UVCs)53,54. As depicted in Fig. 4d, UVC converts Cartesian
PDE problem. Human ventricular electrophysiology is described by coordinates of the LV into 3 new coordinates in fixed ranges: apicobasal
the following mono-domain reaction–diffusion equation: (ρ ∈ [0, 1]), rotational (ϕ ∈ [0, 2π]) and transmural (δ ∈ [0, 1]). Here, ρ
and δ represent a normalized ‘distance’ to the basal and epicardial
Cm ut = −(Iion + Istim ) + ∇D∇u, (5) surfaces, while ϕ represents a rotational angle with respect to the right
ventricle center. We calculate the UVCs for each LV in the cohort and
where Cm is the membrane capacitance, u is the transmembrane poten- establish their correspondence to Ω0 by matching their UVCs. Once
tial, D is a conductivity tensor controlled by the local fiber orientation, the point correspondence is attained, we perform PCA on the coordi-
Istim is the transmembrane current from external stimulation, and Iion is nate difference at landmark points to derive θ ̃, a low-dimensional
the sum of all ionic current. Iion is calculated based on a system of ordi- representation of the shape parameter θ. Supplementary Section 3
nary differential equations with more than 100 equations. Therefore, presents network performance with LDDMM as the registration
we refer readers to find a comprehensive description of Istim, Iion, model method.

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

Testing cohort ATs (ms) RTs (ms)

Neural network prediction Absolute error Neural network prediction Absolute error
300 10 580 10

500
200
5 5
400
100

25 0 300 0

350 15 610 20

10 500
200
10
5 400
100

0 0 300 0

250 20 550 20

200 500

10 10
100 400

0 0 300 0
Pacing location

Relative L2 error distribution (n = 106) Mean absolute error distribution (n = 106) Tenfold cross-validation
Pacing Pacing

Absolute error (ms)

Segment Segment 0.06 7.0

Relative L2 error
ID ATs RTs ID ATs RTs
0 0 0 0 6.5
1 1 1 1 0.04
2 2 2 2 6.0
3 3 3 3
4 4 4 4 5.5
0.02
5 5 5 5
6 6 6 6 5.0
0 0.05 0.10 0.15 0 0.05 0.10 0.15 0 5 10 15 20 ms 0 20 40 60 ms

s
s

s
s
AT
RT

AT

RT
Fig. 5 | Network prediction of ATs and RTs from the testing cohort. Top: LV prediction errors. Bottom left: distributions of the relative L2 error of ATs and RTs
shape with their predicted ATs and RTs and absolute errors. The red stars are on the testing dataset with the corresponding means (vertical black lines). Pacing
the pacing locations. The color bar is selected to be consistent with the CARTO segment ID ranges from 0 to 6. Bottom right: bar plots of the mean relative L2
clinical mapping system, with red denoting early activation and progressing to errors with standard deviation from tenfold cross-validation. Data are presented
purple for late activation. The color bars present the magnitude of solutions and as mean ± s.d.

Notably, we assume that the fiber field depends solely on the shape error in the bottom panel of Fig. 5. Each row represents a different
of the LV53, thereby encoding the LV shape with θ ̃ simultaneously leads myocardial segment as defined in Fig. 4b. The error distributions are
to encoding the fiber field. To illustrate how the fiber fields influence roughly centered around their respective means (black lines) with left
electrical propagation, we present three exemplary cases with ATs on skewness. However, for apical regions (segment 6), both errors are
the reference domain in Supplementary Section 7, each with the same slightly larger than for the other regions, which could be attributed to
pacing location. the complex fiber architecture created by rule-based methods at the
apex55. More specifically, we found that for the three cases with the
Network training and performance. We study the neural network per- largest error, the reconstructed LV shape unfaithfully reflects the true
formance in predicting ATs and RTs on patient-specific LVs. The network shape of the LV due to the collective artifacts from data acquisition
architecture is presented in Fig. 4b. We input shape parameters θ ̃ along and imaging segmentation. We discuss this further in Supplementary
̂ into two separate branches. The trunk network
with pacing locations xpace Section 8.
takes UVCs (x) of each training point as input. We present the network Furthermore, we perform 10-fold cross-validation to quantify the
prediction in the top panel of Fig. 5. After random shuffling, we split the generalization ability of the network, for example, adopting sets 2–10
training dataset (n = 1,006) into 10 subsets with 100 geometries in sets for training and set 1 for testing. The bottom-right panel shows the rela-
1–9 and 106 geometries in set 10. First, we train a network using the geom- tive and absolute errors from 10-fold cross-validation along with their
etries and the ATs and RTs in sets 1–9 and set 10 for testing. We visualize means (0.018 and 0.045) and standard deviations. In general, relative
predicted ATs and RTs along with their absolute errors for three repre- errors of ATs from ten iterations are larger than the RTs errors, although
sentative LVs from the testing cohort. We observe that ATs and RTs on the magnitude of errors in ATs is smaller than the RTs.
each LV with various pacing locations are faithfully predicted by our Finally, we investigate the scalability and efficiency of DIMON.
framework, with the maximum absolute error of ATs and RTs being around Figure 6 presents an estimation of memory usage when adopting
20 ms (purple areas in the ‘Absolute error’ columns in Fig. 5). This error an efficient and inefficient approach of geometry parameterization
is relatively small compared with the maximum values of ATs and RTs. with different spatial resolutions in the training data. We can see that
To quantitatively investigate how well the network predicts for the an inefficient parameterization approach (red regime), that is, vox-
testing cases, we present the distribution of relative L2 and absolute elizing the LV and the ATs and RTs17,56, represents the geometry in a

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

60 Discussion
Not-scalable regime Our framework provides an approach that enables fast prediction of
Inefficient parametrization PDE solutions on multiple domains and facilitates many downstream
50
applications using artificial intelligence. One concrete example
arises from cardiology. For instance, in cardiac electrophysiology, an
40
Memory usage (GB)

CT resolution approach to examine the arrhythmia inducibility of the substrate is to

MRI resolution
perform simulations based on the patient’s heart shape acquired from
30 high-quality imaging scans. Although this is effective2, the excessively
high computational cost poses a substantial challenge to its vast appli-
20
cation and implementation in clinics. With our framework, one can
Scalable regime train a neural operator that predicts electrical propagation on various
Efficient parametrization
hearts without the need to perform expensive numerical simulations.
10
Training cost example 3 Our framework is also promising for topology optimization, where
shapes are optimized to meet certain design requirements, and PDEs
0 need to be solved repeatedly on a large number of domains during
such optimization procedure. DIMON allows for the computation of
0

00

00
50

00

,0
2,

7,
5,

the derivative of the cost functional with respect to the parameters

10
Number of training samples
that describe the domain shapes57 via automatic differentiation. This,
Fig. 6 | Scalability of DIMON. Memory usage estimation given training in turn, simplifies the calculation of admissible deformations in shape
data at the spatial resolution of CT (dashed) and MRI (dashed and dot) using optimization. This framework conveniently allows imposing bound-
inefficient (red regime) and efficient (blue regime) approaches of geometry
ary conditions, or initial conditions, or local material properties on
parameterization. The black dot indicates the memory usage of DIMON. Memory
different geometries.
usage for the inefficient approach is estimated based on data voxelization.
DIMON also has several limitations. We consider only mapping
scalar functions in this work, whereas mapping vector or tensor fields
higher dimension and leads to a sharp increase in memory usage with is yet to be addressed given that it holds particular importance in rep-
the increase of training samples. To incorporate pathophysiological resenting stress and velocity in fluid and solid mechanics. Incorporat-
features reflected in computed tomography (CT) and magnetic reso- ing methods that transport vector and tensor fields across shapes or
nance imaging (MRI) images, one should train the network with data manifolds21,58,59 can potentially address this issue and will be of inter-
at MRI- and CT-level resolution. This can further exacerbate the need est in the future. Also, although PCA is indeed a general approach for
for computational resources, making the network training even more parameterizing shapes irrespective of their topology, other local
expensive. However, reducing the dimensionality with an efficient methods for shape parameterization, such as non-uniform rational
geometry parameterization approach (such as PCA) can alleviate this B-spline (NURBS), can be adopted due to their explicit form of splines
situation and make the training scalable even with higher resolutions. or polynomials60,61. In addition, conducting a rigorous analysis of
The black line represents an estimation of memory usage with more the network’s convergence rate is beyond the scope of this study,
training data and the dot stands for the training cost in this case. In making it difficult to estimate a priori how large of a training set is
conclusion, our framework demonstrates that a neural operator can needed to achieve a given accuracy of the solutions. This framework
efficiently predict patient-specific electric propagation on realistic is designed with the intention to learn operators on a family of dif-
heart geometries with satisfactory accuracy, substantial speedup and feomorphic domains, and learning operators on non-diffeomorphic
scalability in computation. In addition, we tabulate the computational or topologically different domains is beyond the scope of DIMON.
cost in Extended Table 1 between DIMON and openCARP, a traditional This is a challenging problem, not only because of the difficulty of
finite elements solver and used widely by the cardiac computational modeling deformations combined with topological changes but
modeling community. Our neural operator substantially accelerates also because PDE solutions tend to heavily depend on the topology
the overall speed in testing with a short training time: it requires only of the domain.
9 min for training and less than 1 s on a single graphics processing unit
to predict the ATs and RTs for a new patient, whereas it takes our solver Methods
in openCARP 3 h on 48 central processing unit nodes to simulate ATs Large deformation diffeomorphism metric mapping
and RTs in a single beat, not accounting for the additional time spent We give a brief summary of the LDDMM framework that is used spe-
pre-processing and meshing. cifically to estimate diffeomorphic mappings between the domains of
We also perform a comparison between DIMON and Geometry- example 2, but that could more generally apply to a variety of different
Aware Fourier Neural Operator (Geo-FNO)17, a representative attempt situations and geometric objects. The interested reader can further
of learning operators on multiple geometries15–17,56, for predicting refer to refs. 21,47 for additional details on the model.
ATs and RTs on LVs. The results demonstrate that DIMON has bet- The central principle of LDDMM is to introduce diffeomorphic
ter efficiency, accuracy and scalability for learning on families of 3D transformations of ℝd which are obtained as the flows of time-dependent
complex shapes. Details regarding this comparison are presented in vector fields. More precisely, given a time-dependent integrable vector
Supplementary Section 2. field t ∈ [0, 1] ↦ v(t, ∗) ∈ C10 (ℝd , ℝd ) (the space of C1 vector fields on ℝd
This example demonstrates that our framework enables real-time that vanish at infinity as well as their first derivatives). Then the flow
t
human cardiac electrophysiology predictions. Model validation on map φv of v, defined by φv (t, x) = x + ∫0 v(s, φv (s, x))ds , can be shown to
1
such a large-scale realistic cohort is not only very rare given the huge be a C diffeomorphism of ℝ , where s is a standard notation for a dummy
d

efforts on data collection and high computational demands but also variable. The group of all such flows taken at the terminal time t = 1 can
lends credibility to the model applicability and scalability in engineer- be further equipped with a metric as follows. One first restricts to a given
ing applications given its high prediction accuracy. In addition, the Hilbert space V of vector fields in C10 (ℝd , ℝd )(for example a Sobolev space
learning ability of DIMON reaches beyond solution operators of PDEs. of sufficient regularity or reproducing kernel Hilbert space with ade-
The ATs and RTs are the times when the PDE solution, transmembrane quate kernel function) and define the energy of the flow map φv(1, ⋅ ) by
1 2
potential, exceeds or decreases below certain thresholds, and therefore ∫0 ∥ v(t, ∗)∥V dt . This provides a quantitative measure of deformation
functionals of solutions of the PDE, rather than solutions themselves. cost and, given two shapes Ω and Ω′ in ℝd , an optimal transformation

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

mapping Ω to Ω′ can be found by introducing the registration of each function and a trunk network that takes coordinates x ∈ Ω0 as
problem: input. Finally, the output function value at point x is approximated by
1
2
inf {∫ ∥ v(t, ∗)∥V dt | φv (Ω) = Ω ′ } (6) ℱ̃ 0 (θ, vm )(x)
v∈L2 ([0,1],V) 0

(8)
= 𝒮𝒮 (g ̃ 0
g0̃ (φ1 ̃ (φm ̃ )
In practice, it is common to further relax the boundary constraint ⏟⎵0 (φq0 (θ))
⎵⏟⎵ ⎵⏟ ⊙ ⏟⎵ q1 (v
⎵⏟⎵ ⎵⏟1 )) ⊙⋯⊙ ⏟⎵
gm qm (v
⎵⎵⏟⎵ m )) ⊙ f(x)
⎵⎵⏟ ⏟
trunk
φv (Ω) = Ω ′ based on a similarity measure between the deformed branch0 branch1 branchm

domain φv(Ω) and Ω′. This is typically done to ensure the robustness
of the method to noise and inconsistencies between the two shapes. where gĩ (called branch net i) and f ̃ (called trunk net) are neural net-
Optimality conditions for the above problem can be derived from works, ⊙ is the Hadamard product, and 𝒮𝒮 is the summation of all the
the Prontryagin maximum principle of optimal control (compare Ch. 10 components of a vector18. q0, q1..., qm are positive integers. Equation (8)
in ref. 47) from which it can be shown that a minimizing vector field for can be written in a simplified form as:
equation (6) can be parameterized by an initial costate (or momentum)
k m
variable. Furthermore, when Ω is a discrete shape (that is represented j geo
ℱ̃ 0 (θ, v1,…,m )(x) = ∑ ti ∏ bi bi + b.
by a finite set of vertices), this initial momentum can be identified with i=1 j=1
a distribution of vectors over the vertices of Ω. This allows us to reduce
geo
the registration problem to a minimization over the finite-dimensional bi is the output of the Geo. branch given θ, ti is the trunk network at
initial momentum variable, which can be tackled for instance with a ith neuron, and bi j is the output of ith neuron from jth branch network9,18.
shooting approach. For the applications presented in this paper, we In this work, we adopt the mean squared error as the loss function.
make use specifically of the implementation of diffeomorphic registra-
tion from the MATLAB library FshapesTk. Cohort and simulations details in example 3
This cohort includes patients with hypertrophic cardiomyopathy
Principal component analysis (HCM) who were treated at the HCM Center of Excellence at the Johns
Consider p landmark points on Ω0 with known point correspondences Hopkins Hospital from 2005 to 2021. This study complies with all
on each Ω θi given N domains, A ∈ ℝp×N denotes the Euclidean difference relevant ethical regulations and was approved by the Johns Hopkins
between the target domains and the reference domain in each direc- Medicine institutional review boards (approval number IRB00205270).
tion. We define Ā = A − μ , where Ā is normalized to zero mean. We Our dataset consists of de-identified late gadolinium-enhanced CMR
perform the singular value decomposition Ā = U ΣVT , where Ā ∈ ℝN×p, (LGE-CMR) scans from over 850 patients. The LGE-CMR scans have slice
U ∈ ℝN×p, Σ ∈ ℝp×p, VT ∈ ℝp×p and T is the standard notation for matrix thicknesses ranging from 6 mm to 10 mm and spatial resolutions vary-
transpose. We truncate U and Σ to first p′ rank and obtain the reduced ing from 1.17 mm to 2.56 mm. Notice that some patients have multiple
′ ′ ′ ′ ′
matrices as Ur ∈ ℝN×p , Σr ∈ ℝp ×p and VTr ∈ ℝp ×p . We define the coeffi- scans over time, which sums up to 1,006 scans.
N×p′
cient matrix D = UrΣr, where each row in D ∈ ℝ is a reduced repre- To simulate electrical propagation, we first segment the myocar-
sentation of the deformation from Ωθ to Ω0, referred to as reduced dium and the blood pool from the raw images. We then reconstruct a
shape parameter θ .̃ mesh and generate fibers for each LV based on the segmented images.
The average length of mesh size is 400 μm, leading to an average num-
Domain generation ber of vertices on the order of 4 million2,62. To account for the direc-
We generate the computational domain in example 1 following the tionality of the electrical signal propagated along the fibers, we adopt
equations below. the validated rule-based method proposed in ref. 53 to generate the
x = X + a1 cos(a2 π + a3 ) sin(a2 π + a3 ),
myocardial fibers in each patient’s LV.
⎧ Next, we divide the myocardium surface into seven segments,
⎪ 2
y = Y + b1 sin (b2 π + b3 ), which is a simplified version of the standardized myocardium segmen-
⎨
⎪ tation from the American Heart Association63,64. Within each segment
⎩ a1 , b1 , a3 , b3 ∼ U[−0.5, 0.5], a2 , b2 ∼ U[0, 0.75],
highlighted with a different color in Fig. 4, we simulate a paced beat
where X, Y ∈ [0, 1] × [0, 1] are the coordinates in the reference domain, for 600 ms at a randomly picked endocardial site until the LV is fully
and x and y are their corresponding coordinates on newly generated repolarized. Specifically, ATs and RTs are chosen as the time when
shape. ai and bj are shape parameters, all chosen independently as fol- the transmembrane potential u reaches 0 mV from the resting state
lows: a1, a3 and b1, b3 from the uniform distribution in [−0.5, 0.5] and a2, (−85 mV) and the time when u decreases below −70 mV, recovering
b2 from the uniform distribution on [0, 0.75]. from depolarization. More details regarding the solver and simulation
In example 2, we define the reference domain as an annulus with procedure can be found in refs. 2,65.
inner and outer radii at 0.5 and 1.0, respectively. The parametric fam-
ily of domains is rotated ellipsis with a void. We sample the shape Reporting summary
parameters following: Further information on research design is available in the Nature
Portfolio Reporting Summary linked to this article.
c1 ≈ U[1.5, 2.0], c2 ≈ U[1.0, 1.5], c3 ≈ U[0, π], c4 ≈ U[0.2, 0.8], (7)
Data availability
Here, c1 and c2 stand for the length of the semi-major and minor axes, c3 Source data for Figs. 2, 3 and 5 are available with this paper. Data for
is the rotational angle, and c4 is the radius of the inner circle. examples 1 and 2 are are generated using FEniCS software (version
2019.1.0) and are available on Zenodo at https://doi.org/10.5281/
Neural operator zenodo.13958884 (ref. 66). The data in example 3 are not included in
In this section, we briefly summarize the architecture of MIONet18. Notice this link as the heart geometries are patient-specific clinical data. The
that our framework is inclusive of the architecture of the neural operator data in example 3 can be provided by request to the corresponding
we use to approximate ℱ0. We denote ℱ̃ 0 as an approximation of the authors and potentially after an IRB for sharing data is approved. We
mapping from the m + 1 input functions [θ, v1, …, vm] to the output func- also include the Ten Tusscher downloaded from https://models.cellml.
tion u. We remove the subscript that denotes the domain for clarity. We org/e/80d in our repository as references. Source data are provided
then define m + 1 branch networks that input a discrete representation with this paper.

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

Code availability 20. Vercauteren, T., Pennec, X., Perchant, A. & Ayache, N.
The source code for DIMON is available on GitHub at https://github. Diffeomorphic demons: efficient non-parametric image
com/MinglangYin/DIMON and on Zenodo at https://doi.org/10.5281/ registration. NeuroImage 45, S61–S72 (2009).
zenodo.14009828 (ref. 67). The LDDMM pipeline is available at https:// 21. Beg, M. F., Miller, M. I., Trouvé, A. & Younes, L. Computing
plmlab.math.cnrs.fr/benjamin.charlier/fshapesTk. large deformation metric mappings via geodesic flows of
diffeomorphisms. Int. J. Comput. Vision 61, 139–157 (2005).
References 22. Henry, D. Perturbation of the Boundary in Boundary-Value Problems
1. Sokolowski, J. & Zolésio, J.-P. Introduction to Shape Optimization of Partial Differential Equations No. 318 (Cambridge Univ. Press,
(Springer, 1992). 2005).
2. Prakosa, A. et al. Personalized virtual-heart technology for guiding 23. Kissas, G. et al. Learning operators with coupled attention.
the ablation of infarct-related ventricular tachycardia. Nat. J. Mach. Learn. Res. 23, 9636–96981 (2022).
Biomed. Eng. 2, 732–740 (2018). 24. Cao, Q., Goswami, S. & Karniadakis, G. E. Laplace neural operator
3. Lagaris, I. E., Likas, A. & Fotiadis, D. I. Artificial neural networks for solving differential equations. Nat. Mach. Intell. 6, 631–640
for solving ordinary and partial differential equations. IEEE Trans. (2024).
Neural Netw. 9, 987–1000 (1998). 25. Gilbarg, D., Trudinger, N. S., Gilbarg, D. & Trudinger, N. S. Elliptic
4. Chen, T. & Chen, H. Universal approximation to nonlinear Partial Differential Equations of Second Order Vol. 224, No. 2
operators by neural networks with arbitrary activation functions (Springer, 1977).
and its application to dynamical systems. IEEE Trans. Neural Netw. 26. Jerison, D. & Kenig, C. E. The inhomogeneous Dirichlet problem in
6, 911–917 (1995). Lipschitz domains. J. Funct. Anal. 130, 161–219 (1995).
5. Karniadakis, G. E. et al. Physics-informed machine learning. Nat. 27. Karniadakis, G. & Sherwin, S. J. Spectral/hp Element Methods for
Rev. Phys. 3, 422–440 (2021). Computational Fluid Dynamics (Oxford Univ. Press, 2005).
6. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed 28. Patera, A. T. A spectral element method for fluid dynamics:
neural networks: a deep learning framework for solving forward laminar flow in a channel expansion. J. Comput. Phys. 54,
and inverse problems involving nonlinear partial differential 468–488 (1984).
equations. J. Comput. Phys. 378, 686–707 (2019). 29. Babuška, I. & Suri, M. The p-and hp versions of the finite element
7. Yu, B. The Deep Ritz Method: a deep learning-based numerical method, an overview. Comput. Methods Appl. Mech. 80, 5–26
algorithm for solving variational problems. Commun. Math. Stat. (1990).
6, 1–12 (2018). 30. Avants, B. & Gee, J. C. Geodesic estimation for large deformation
8. Cai, S., Mao, Z., Wang, Z., Yin, M. & Karniadakis, G. E. anatomical shape averaging and interpolation. Neuroimage 23,
Physics-informed neural networks (PINNs) for fluid mechanics: a S139–S150 (2004).
review. Acta Mech. Sin. 37, 1727–1738 (2021). 31. Joshi, S., Davis, B., Jomier, M. & Gerig, G. Unbiased diffeomorphic
9. Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning atlas construction for computational anatomy. NeuroImage 23,
nonlinear operators via DeepONet based on the universal S151–S160 (2004).
approximation theorem of operators. Nat. Mach. Intell. 3, 218–229 32. Ma, J., Miller, M. I. & Younes, L. A Bayesian generative model for
(2021). surface template estimation. Int. J. Biomed. Imaging 2010, 974957
10. Li, Z., Huang, D. Z., Liu, B. & Anandkumar, A. Fourier neural (2010).
operator for parametric partial differential equations. J. Mach. 33. Cury, C., Glaunès, J., Chupin, M. & Colliot, O. Analysis of
Learn Res. 24, 18593–18618 (2024). anatomical variability using diffeomorphic iterative centroid in
11. Wu, H., Luo, H., Wang, H., Wang, J. & Long, M. Transolver: a patients with Alzheimer’s disease. Comput. Methods Biomech.
fast transformer solver for PDEs on general geometries. In Biomed. Eng. Imaging Vis. 5, 350–358 (2017).
International Conference on Machine Learning (Association for 34. Hartman, E., Sukurdeep, Y., Klassen, E., Charon, N. & Bauer, M.
Computing Machinery, 2024). Elastic shape analysis of surfaces with second-order sobolev
12. Hao, Z. et al. GNOT: a general neural operator transformer metrics: a comprehensive numerical framework. Inte. J. Comput.
for operator learning. In International Conference on Machine Vision 131, 1183–1209 (2023).
Learning 12556–12569 (PMLR, 2023). 35. Miller, M. I., Trouvé, A. & Younes, L. On the metrics and Euler–
13. Hang, Z., Ma, Y., Wu, H., Wang, H. & Long, M. Unisolver: Lagrange equations of computational anatomy. Annu. Rev.
PDE-conditional transformers are universal PDE solvers. Preprint Biomed. Eng. 4, 375–405 (2002).
at https://arxiv.org/abs/2405.17527 (2024). 36. Fan, H., Su, H. & Guibas, L. J. A point set generation network
14. Serrano, L. et al. Operator learning with neural fields: tackling for 3D object reconstruction from a single image. In Proc. IEEE
PDEs on general geometries. Adv. Neural Inf. Process. Syst. 36, Conference on Computer Vision and Pattern Recognition 605–613
70581–70611 (2023). (IEEE, 2017).
15. Li, Z. et al. Geometry-informed neural operator for large-scale 3D 37. Park, J. J., Florence, P., Straub, J., Newcombe, R. & Lovegrove,
PDEs. In Proc. 37th Conference on Neural Information Processing S. DeepSDF: learning continuous signed distance functions for
Systems 35836–35854 (Curran Associates, 2024). shape representation. In Proc. IEEE/CVF Conference on Computer
16. He, J. et al. Novel DeepONet architecture to predict stresses in Vision and Pattern Recognition 165–174 (IEEE, 2019).
elastoplastic structures with variable complex geometries and 38. Rock, J. et al. Completing 3D object shape from one depth
loads. Comput. Methods Appl. Mech. Eng. 415, 116277 (2023). image. In Proc. IEEE Conference on Computer Vision and Pattern
17. Li, Z., Zhengyu Huang, D., Liu, B. & Anandkumar, A. Fourier Recognition 2484–2493 (IEEE, 2015).
neural operator with learned deformations for PDEs on general 39. Rubner, Y., Tomasi, C. & Guibas, L. J. The Earth Mover’s distance
geometries. J. Mach. Learn. Res. 24, 18593–18618 (2023). as a metric for image retrieval. Int. J. Comput. Vision 40, 99–121
18. Jin, P., Meng, S. & Lu, L. MIONet: learning multiple-input operators (2000).
via tensor product. SIAM J. Sci. Comput. 44.6, A3490–A3514 (2022). 40. Grenander, U. & Miller, M. I. Computational anatomy: an emerging
19. Ovsjanikov, M., Ben-Chen, M., Solomon, J., Butscher, A. & Guibas, discipline. Q. Appl. Math. 56, 617–694 (1998).
L. Functional maps: a flexible representation of maps between 41. Bauer, M., Harms, P. & Michor, P. W. Sobolev metrics on shape
shapes. ACM Trans. Graph. 31, 1–11 (2012). space of surfaces. J. Geom. Mech. 3, 389–438 (2010).

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

42. Desbrun, M., Meyer, M. & Alliez, P. Intrinsic parameterizations 63. American Heart Association Writing Group on Myocardial
of surface meshes. Comput. Graph. Forum 21, 209–218 Segmentation and Registration for Cardiac Imaging. Standardized
(2002). myocardial segmentation and nomenclature for tomographic
43. Praun, E. & Hoppe, H. Spherical parametrization and remeshing. imaging of the heart: a statement for healthcare professionals
ACM Trans. Graph. 22, 340–349 (2003). from the Cardiac Imaging Committee of the Council on Clinical
44. Choi, G. P. T. & Lui, L. M. in Handbook of Mathematical Models Cardiology of the American Heart Association. Circulation 105,
and Algorithms in Computer Vision and Imaging: Mathematical 539–542 (2002).
Imaging and Vision (eds Chen, K. et al.) 1–41 (Springer, 2022). 64. Deng, D., Prakosa, A., Shade, J., Nikolov, P. & Trayanova, N. A.
45. Sun, X. & Hancock, E. R. Quasi-isometric parameterization for Sensitivity of ablation targets prediction to electrophysiological
texture mapping. Pattern Recognit. 41, 1732–1743 (2008). parameter variability in image-based computational models of
46. Ovsjanikov, M. et al. Computing and processing correspondences ventricular tachycardia in post-infarction patients. Front. Physiol.
with functional maps. In Special Interest Group on Computer 10, 628 (2019).
Graphics and Interactive Techniques Conference 1–60 65. Plank, G. et al. The openCARP simulation environment for cardiac
(Association for Computing Machinery, 2017). electrophysiology. Comput. Methods Programs Biomed. 208,
47. Younes, L. Shapes and Diffeomorphisms Vol. 171 (Springer, 2010). 106223 (2021).
48. Fletcher, P. T., Lu, C., Pizer, S. M. & Joshi, S. Principal geodesic 66. Yin, M. et al. Data repository, a scalable framework for learning
analysis for the study of nonlinear statistics of shape. IEEE Trans. the geometry-dependent solution operators of partial differential
Med. Imaging 23, 995–1005 (2004). equations. Zenodo https://doi.org/10.5281/zenodo.13958884
49. Scroggs, M. W., Dokken, J. S., Richardson, C. N. & Wells, G. N. (2024).
Construction of arbitrary order finite element degree-of-freedom 67. Yin, M. et al. Code repository, a scalable framework for learning
maps on polygonal and polyhedral cell meshes. ACM Trans. Math. the geometry-dependent solution operators of partial differential
Softw. 48, 1–23 (2022). equations. Zenodo https://doi.org/10.5281/zenodo.14009828
50. Newell, A. C. & Whitehead, J. A. Finite bandwidth, finite amplitude (2024).
convection. J. Fluid Mech. 38, 279–303 (1969).
51. Ten Tusscher, K. H., Noble, D., Noble, P. J. & Panfilov, A. V. A model Acknowledgements
for human ventricular tissue. Am. J. Physiol. Heart Circ. Physiol. This work is supported by NIH grants R01HL166759 and R01HL174440
286, H1573–H1589 (2004). (to N.T.), a grant from the Leducq Foundations (to N.T.), a Heart Rhythm
52. Ten Tusscher, K. H. & Panfilov, A. V. Alternans and spiral breakup Society Fellowship (to M.Y.), the US Department of Energy under
in a human ventricular tissue model. Am. J. Physiol. Heart Circ. grants DE-SC0025592 and DE-SC0025593 (to L.L.), and NSF grants
Physiol. 291, H1088–H1100 (2006). DMS-2347833 (to L.L.), DMS-1945224 (to N.C.) and DMS-2436738 (to
53. Bayer, J. D., Blake, R. C., Plank, G. & Trayanova, N. A. A novel M.M. and N.T.). M.M. is also supported by AFOSR awards FA9550-20-
rule-based algorithm for assigning myocardial fiber orientation to 1-0288, FA9550-21-1-0317 and FA9550-23-1-0445. M.Y. thanks Z. Li, A.
computational heart models. Ann. Biomed. Eng. 40, 2243–2254 Prakosa and S. Loeffler for multiple discussions during the preparation
(2012). of this paper.
54. Bayer, J. et al. Universal ventricular coordinates: a generic
framework for describing position within the heart and Author contributions
transferring data. Med. Image Anal. 45, 83–93 (2018). M.Y., N.C., M.M. and N.T. designed the research. M.Y., N.C.
55. Doste, R. et al. A rule-based method to model myocardial and R.B. performed the research. M.Y., N.C., L.L. and M.M.
fiber orientation in cardiac biventricular geometries with analyzed the data. M.Y., N.C., L.L. and M.M. contributed to the
outflow tracts. Int. J. Numer. Methods Biomed. Eng. 35, e3185 theoretical results. M.Y., N.C., R.B., L.L., M.M. and N.T. wrote
(2019). the paper.
56. Liu, N., Jafarzadeh, S. & Yu, Y. Domain agnostic Fourier neural
operators. In Proc. 37th International Conference on Neural Competing interests
Information Processing System 47438–47450 (Curran Associates, The authors declare no competing interests.
2024).
57. Shukla, K. et al. Deep neural operators as accurate surrogates for Additional information
shape optimization. Eng Appl. Artif. Intel. 129, 107615 (2024). Extended data is available for this paper at https://doi.org/10.1038/
58. Donati, N., Corman, E., Melzi, S. & Ovsjanikov, M. Complex s43588-024-00732-2.
functional maps: a conformal link between tangent bundles.
Comput. Graph. Forum 41, 317–334 (2022). Supplementary information The online version contains
59. Azencot, O., Ben-Chen, M., Chazal, F. Ovsjanikov, M. An operator supplementary material available at https://doi.org/10.1038/s43588-
approach to tangent vector field processing. Comput. Graph. 024-00732-2.
Forum 32, 73–82 (2013).
60. Bucelli, M., Salvador, M. & Quarteroni, A. Multipatch isogeometric Correspondence and requests for materials should be addressed to
analysis for electrophysiology: simulation in a human heart. Natalia Trayanova or Mauro Maggioni.
Comput. Methods Appl. Mech. Eng. 376, 113666 (2021).
61. Willems, R., Janssens, K. L. P. M., Bovendeerd, P. H. M., Verhoosel, Peer review information Nature Computational Science thanks Daniel
C. V. & van der Sluis, O. An isogeometric analysis framework for Zhengyu Huang and the other, anonymous, reviewer(s) for their
ventricular cardiac mechanics. Comput. Mech. 73, 465–506 contribution to the peer review of this work. Primary Handling Editor:
(2024). Fernando Chirigati, in collaboration with the Nature Computational
62. Boyle, P. M., Ochs, A. R., Ali, R. L., Paliwal, N. & Trayanova, N. A. Science team.
Characterizing the arrhythmogenic substrate in personalized
models of atrial fibrillation: sensitivity to mesh resolution and Reprints and permissions information is available at
pacing protocol in AF models. EP Europace 23, i3–i11 (2021). www.nature.com/reprints.

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

Publisher’s note Springer Nature remains neutral with regard to adapted material derived from this article or parts of it. The images
jurisdictional claims in published maps and institutional affiliations. or other third party material in this article are included in the article’s
Creative Commons licence, unless indicated otherwise in a credit
Open Access This article is licensed under a Creative Commons line to the material. If material is not included in the article’s Creative
Attribution-NonCommercial-NoDerivatives 4.0 International License, Commons licence and your intended use is not permitted by statutory
which permits any non-commercial use, sharing, distribution and regulation or exceeds the permitted use, you will need to obtain
reproduction in any medium or format, as long as you give appropriate permission directly from the copyright holder. To view a copy of this
credit to the original author(s) and the source, provide a link to the licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Creative Commons licence, and indicate if you modified the licensed
material. You do not have permission under this licence to share © The Author(s) 2024

Nature Computational Science

Article https://doi.org/10.1038/s43588-024-00732-2

Extended Data Table 1 | Computational costs of DIMON and openCARP in Example 3

Computational costs of DIMON in training and testing are included.

Nature Computational Science