---
title: "Lin, Dahiya, Cersonsky propose generalized approach to incorporating geometry and directionality into coarse-grained machine-learned potentials University of Wisconsin–Madison; enables anisotropic coarse-grained simulations"
sdDatePublished: "2026-09-03T19:04:00Z"
source: "https://arxiv.org/pdf/2609.01911"
topics:
  - name: "scientific research"
    identifier: "medtop:20000735"
  - name: "scientific innovation"
    identifier: "medtop:20000736"
  - name: "materials science"
    identifier: "medtop:20000764"
  - name: "artificial intelligence"
    identifier: "medtop:20001298"
  - name: "chemistry"
    identifier: "medtop:20000725"
locations:
  - "Madison"
---


Lin, Dahiya, Cersonsky propose generalized approach to incorporating geometry and directionality into coarse-grained machine-learned potentials University of Wisconsin–Madison; enables anisotropic coarse-grained simulations

A Generalized Approach for Incorporating Geometry and Directionality into Coarse-Grained Machine-Learned Potentials

A Generalized Approach for Incorporating Geometry and Directionality into
Coarse-Grained Machine-Learned Potentials
Arthur Y. Lin,1 Tejas Dahiya,2 and Rose K. Cersonsky1, 3, 4, a)
1)Department of Chemical and Biological Engineering, University of Wisconsin - Madison, Madison, WI,
USA
2)Department of Computer Science, University of Wisconsin - Madison, Madison, WI,
USA
3)Department of Materials Science and Engineering, University of Wisconsin - Madison, Madison, WI,
USA
4)Data Science Institute, University of Wisconsin - Madison, Madison, WI,
USA
I.
INTRODUCTION
Machine-learned interatomic potentials (MLIPs) have
dramatically expanded the scope of molecular simulation,
enabling the prediction of atomic energies and forces with
near ab initio accuracy at costs compatible with large-
scale molecular dynamics.1–3 As these models continue to
mature, the principal obstacle to computational discovery
is increasingly shifting from the accuracy of atomistic
interactions to the challenge of accessing the larger length
and time scales associated with collective behavior, self-
assembly, and phase transitions. Coarse-graining (CG), in
which groups of atoms are mapped onto effective particles
with reduced degrees of freedom, remains one of the most
powerful approaches for bridging this gap.4–8
For any MLIP, its capabilities are determined jointly
by the quality of the underlying data, the choice of repre-
sentation X, and the architecture f(X), where X varies
from simple forms (e.g., distance matrices or graphs)
to more complex descriptors (e.g., Smooth Overlap of
Atomic Positions [SOAP]9 or Behler-Parinello symme-
try functions10). A representation is complete if distinct
data objects map to distinct data points; in the atomistic
MLIP community there has been considerable discussion
on what constitutes a complete descriptor, as this has
direct implications to the ceiling of predictive accuracy
in subsequent models.11–14. Simply put, while expressive
architectures can approximate complex functions f(X),
they cannot recover information that is not encoded in
X, and degeneracy in data representation can explicitly
limit model performance.
However, unlike atomistic machine-learned potentials,
the variables describing a coarse-grained system are
not fixed, but instead chosen.
Historically, much of
coarse-graining has therefore been built around isotropic
particles.4 This choice is attractive for both concep-
tual and computational reasons.
Interactions become
functions of intermolecular separation alone, simulation
methodologies are well-established, and many successful
coarse-grained models have been developed within this
framework.15 Such approaches have proven remarkably
a)Electronic mail: rose.cersonsky@wisc.edu
successful, particularly in biomolecular systems where
coarse-graining has enabled simulations inaccessible at
atomistic resolution.15 Yet there are also many examples
where isotropic descriptions struggle to reproduce experi-
mentally observed behavior, particularly in systems whose
organization is governed by packing, local structure, or
directional interactions.16,17
One possible explanation of these difficulties is that
geometry itself carries critical information. Consider two
molecular configurations possessing similar intermolecular
separations but different relative orientations. Depending
on the system, these configurations may exhibit substan-
tially different interaction energies. When orientation
is removed from the coarse-grained representation, how-
ever, both are represented by the same set of variables
R. The resultant coarse-grained configuration is there-
fore associated not with a single underlying interaction
energy, but with a distribution of possible interaction en-
ergies. As this distribution broadens, the construction of a
coarse-grained potential becomes increasingly difficult, as
distinct molecular environments become indistinguishable
within the representation.
This observation naturally suggests retaining orienta-
tional and geometric information within coarse-grained
models. However, doing so raises a separate question: how
should such information be represented? While atomistic
machine-learned potentials have undergone rapid develop-
ment over the past decade, the overwhelming majority of
descriptor frameworks and architectures were developed
for collections of isotropic atoms. Extending these ap-
proaches to anisotropic particles requires representations
that remain sensitive to particle geometry while preserv-
ing the rotational and translational symmetries of the
underlying problem.
Several recent approaches have begun to address this
challenge through anisotropic extensions to machine-
learning architectures.18? ,19 Of particular interest are
many-body density expansion methods, which have
proven highly successful in atomistic machine learning due
to their systematic improvability and clear connection to
local structure.9,20,21 AniSOAP extended the SOAP formal-
ism to non-spherical particles through anisotropic density
expansions, enabling the explicit incorporation of particle
shape and orientation into a many-body descriptor,22,23,
and similar efforts expanded the Chebyshev polynomial
arXiv:2609.01911v1 [physics.chem-ph] 1 Sep 2026

2
suite for nanoscale particles.24–26.
In this work, we demonstrate the strategies and subse-
quent trade-offs for incorporating geometric information
into MLIPs. First, we show how incorporating geometric
tensors, both through AniSOAP and as node features in
MACE, compare to their isotropic analogs as a function
of molecular symmetry and objective weighting. Then, we
demonstrate how to successfully incorporate geometric in-
formation into MACE-style potentials for arbitrary molec-
ular symmetries, in ways that extend to coarse-graining
at any resolution, even that of rigid composite bodies
typically found in the nanoscale community. We do so for
single-bead coarse-grained systems with minimal mapping
entropy in order to demonstrate the limitations of poten-
tial energy fitting for coarse-graining, but will note that
analogous schemes can be applied to learn the potential
of mean force. Our coinciding packages, AniSOAP23 and
MACE-CG are also included open-source, to allow for
adoption of this pipeline for new coarse-graining tasks.
II.
METHODS
A.
AniSOAP Descriptors
Many-body density expansions provide a natural frame-
work for constructing machine-learning representations
of local environments. In conventional SOAP-based ap-
proaches, each particle is represented by an isotropic
Gaussian density centered at its position. AniSOAP22,23
extends this formalism to anisotropic particles by replac-
ing isotropic densities with ellipsoidal density fields that
explicitly depend on particle geometry and orientation.
Each coarse-grained particle is described by a position
ri, an orientation matrix Ri, and geometric parameters
corresponding to its principal semiaxes (ai, bi, ci). The
particle geometry is encoded through the matrix
Σi = Ri


a2
i
0
0
0
b2
i
0
0
0 c2
i

RT
i .
(1)
The local density surrounding particle i is then written as
ρi(r) =
X
j
exp

−1
2(r −rij)T Σ−1
j (r −rij)

,
(2)
where rij = rj −ri denotes the position of neighboring
particle j relative to particle i.
Following the SOAP
formalism, the density is expanded in a basis of radial
functions and spherical harmonics,
ρi(r) =
X
nlm
c(i)
nlmgn(r)Ylm(ˆr),
(3)
producing a set of rotationally equivariant coefficients
c(i)
nlm. Rotationally invariant features are subsequently ob-
tained through contractions of these coefficients. Through-
out this work, we employ the AniSOAP power spectrum,
p(i)
nn′l =
X
m
c(i)
nlmc(i)∗
n′lm,
(4)
which forms the descriptor vector xi used for regression.
By construction, the descriptor is invariant to global
rotations and translations while remaining sensitive to
particle geometry and local orientational environments.
We can consider linear energy models acting on the
AniSOAP power spectrum written as
E =
X
i
xT
i w
(5)
where xi denotes the AniSOAP descriptor associated with
particle i and w is a learned coefficient vector.
B.
Incorporating Rigid-Body Orientation into MACE
As a complementary approach to AniSOAP, we extended
the MACE equivariant message-passing framework to de-
scribe coarse-grained particles with explicit rigid-body
orientation. The central objective was to preserve the
existing MACE treatment of translational geometry while
augmenting each intermolecular edge with orientational
information. In this construction, the molecular symme-
try is incorporated into the representation of the body
orientation before that representation enters the MACE
interaction network.
1.
Standard MACE edge geometry
For a pair of coarse-grained particles i and j, conven-
tional MACE begins from the relative displacement
rij = rj −ri,
(6)
which is decomposed into the scalar separation
rij = ∥rij∥
(7)
and the unit edge direction
ˆrij = rij
rij
.
(8)
The distance is expanded in a radial basis, while the
angular dependence is represented by spherical harmonics,
Yℓm(ˆrij).
(9)
Under a global rotation S ∈SO(3), these functions trans-
form according to the ℓth irreducible representation,
Yℓ(Sˆrij) = D(ℓ)(S)Yℓ(ˆrij),
(10)

3
FIG. 1: Workflow for incorporating molecular geometry and directionality into MACE-style potentials.
which allows the edge geometry to be coupled to node
features through the Clebsch–Gordan tensor products
used throughout MACE.
For an atomistic system, the combination of particle
identity and relative positions is sufficient to define the
local geometry. This is no longer true when an entire
molecule is mapped to a single coarse-grained site. Two
molecular configurations can possess identical center-of-
mass positions, and therefore identical rij and Yℓm(ˆrij),
while differing substantially in molecular orientation and
interaction energy. A single-site rigid-body model there-
fore requires an additional orientational degree of freedom.
2.
Rigid-body state and orientation
Each coarse-grained molecule is represented by
(ri, Ri, si) ,
(11)
where ri is the molecular center-of-mass position, Ri ∈
SO(3) is a rotation matrix mapping a fixed molecular
body frame to the simulation frame (we will call this
the “simulation-frame” rotation), and si denotes molec-
ular species. In practice, orientations are stored as unit
quaternions and converted to rotation matrices before
constructing equivariant features.
Writing the columns of Ri as the simulation-frame body
axes,
Ri =


|
|
|
ai bi ci
|
|
|

,
(12)
a global rotation S acts as
ri →Sri,
Ri →SRi.
(13)
The body axes therefore transform as ordinary equivariant
vectors and can, in principle, be coupled directly to the
spherical-harmonic edge features using the same tensor-
product machinery already present in MACE.
A full-frame rigid-body representation can consequently
be constructed from the equivariant product
Y (ˆrij) ⊗Ri ⊗Rj,
(14)
where the three columns of each rotation matrix are
treated as ℓ= 1 features. Clebsch–Gordan tensor prod-
ucts decompose this product into irreducible SO(3) chan-
nels, which are subsequently projected into a chosen set
of edge irreducible representations.
Although Eq. 14 preserves complete body-frame infor-
mation, it contains a redundancy whenever the molecule
possesses a nontrivial rotational point-group symmetry.
3.
Molecular symmetry and the orientation quotient
Let G ⊂SO(3) denote the proper rotational symmetry
group of a molecule. For any g ∈G, the two body-frame
orientations
Ri
and
Rig
(15)
describe the same physical molecular configuration. The
physical orientational degree of freedom therefore belongs
to the quotient space
[Ri] ∈SO(3)/G.
(16)
An orientational feature map ΦG(R) should accordingly
satisfy two distinct transformation properties. First, a
rotation of the entire physical system should act equivari-
ant