🔬 materials science

The Young-E(3) Tensor Product Decomposition for Rotation and Permutation Equivariant Cluster Expansions

This paper introduces the Young-E(3) Tensor Product (YE3T) decomposition, a generalization of cluster expansions that generates complete, orthonormal, and efficient bases for arbitrary rotation and permutation symmetries, thereby surpassing existing methods in speed and accuracy for machine-learned interatomic potentials.

Original authors: James M. Goff, Aidan P. Thompson

Published 2026-09-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: James M. Goff, Aidan P. Thompson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand the world of atoms, scientists often build digital twins: computer models that predict how atoms stick together, how they move, and how they react. These models are essential for designing new materials, from stronger alloys to better batteries. However, atoms are not static bricks; they are dynamic particles that obey strict rules of symmetry. If you rotate a molecule in space, its energy shouldn't change. If you swap two identical atoms, the physics remains the same. For decades, researchers have built models that respect these rules, but they have often done so by treating rotation and swapping as separate problems. This approach works, but it creates a lot of unnecessary clutter in the math, forcing computers to do extra work to filter out redundant information. It is like trying to organize a library by sorting books by color first, then by author, only to realize you have to re-sort the entire collection because the first step created duplicates.

A team of researchers at Sandia National Laboratories has developed a new way to build these atomic models that handles both rotation and swapping simultaneously. They call their method the Young-E(3) tensor product decomposition, or YE3T for short. Instead of building a massive, messy list of possibilities and then trying to clean it up, their method constructs the list perfectly from the start. By treating the symmetries of the physical world as a single, unified system, they have created a set of mathematical tools that are faster, more accurate, and capable of describing complex quantum behaviors that previous models could not easily capture. This advancement allows scientists to simulate materials with greater precision and less computing power, opening the door to more detailed studies of everything from simple molecules to the behavior of electrons in complex solids.

The core of the problem lies in how computers represent the interactions between atoms. In traditional methods, scientists often start by creating a huge, overcomplete set of features—essentially a giant list of every possible way atoms could interact. They then use heavy computational lifting to find the unique, independent items in that list, discarding the rest. This process is slow and inefficient, especially when the models try to look at groups of many atoms at once. The new approach by Goff and Thompson avoids this bottleneck entirely. They realized that the mathematical structures governing how atoms rotate and how they swap places are deeply connected. By using a specific mathematical technique known as Schur–Weyl decomposition, they can organize these interactions into neat, non-repeating blocks right from the beginning.

Imagine a set of building blocks where each block represents a specific type of atomic interaction. In the old way, you would dump all the blocks on the floor, mix them up, and then spend hours picking out the duplicates. The new method is like having a factory that only produces the exact unique blocks you need, arranged in a perfect order. The researchers demonstrated that their new basis sets, which are the mathematical building blocks of the model, contain the old, popular methods as a special case. This means their new system can do everything the old systems could do, but it does it without the wasted effort. In fact, for the specific case where the model only cares about swapping identical atoms in a symmetric way, their method reproduces the results of the best existing tools but runs two to five times faster.

The power of this new method becomes even more apparent when dealing with more complex physics, such as the behavior of electrons in molecules. Electrons have a property that makes them behave differently than simple atoms; they are "antisymmetric," meaning swapping two of them changes the sign of their mathematical description. Previous models struggled to handle this nuance without becoming incredibly slow or inaccurate. The researchers tested their new approach on the electron repulsion in molecules like carbonic acid and the carbonate ion. They found that by allowing the model to carry mixed symmetry information through its internal layers—rather than forcing it to be perfectly symmetric or antisymmetric at every single step—the model learned the correct physics much faster. In their tests, the new models reached a high level of accuracy with fewer features and significantly less computational time compared to models that were restricted to simpler, rigid symmetry rules.

Beyond speed and accuracy, the method offers a new level of flexibility. It allows scientists to tune the symmetry of their models to match the specific physical problem they are solving. Whether they are studying a simple crystal lattice or a complex electronic system, they can now choose the exact type of symmetry behavior they need without being forced into a one-size-fits-all approach. The researchers showed that this flexibility leads to better predictions. In tests on nickel and other materials, the new models continued to improve in accuracy as more features were added, whereas older models tended to hit a wall where adding more features did not help. This suggests that the new method captures the true physical relationships between atoms more effectively, avoiding the dead ends that plagued earlier approaches.

The implications of this work extend far beyond just making simulations run faster. By providing a rigorous, complete, and orthonormal set of building blocks, the researchers have laid a new foundation for machine learning in materials science. Their method removes the guesswork and redundancy that have slowed down progress in the field. It allows for the construction of models that are not only mathematically sound but also physically intuitive. The researchers have already made their software available to the community, integrating it into the widely used LAMMPS simulation package. This means that other scientists can immediately begin using these more efficient tools to explore new materials and understand the fundamental forces that hold our world together. The work represents a shift from brute-force computation to a more elegant, principled understanding of atomic symmetry, proving that sometimes the best way to solve a complex problem is to stop fighting the math and start working with it.

Technical Summary: The Young-E(3) Tensor Product Decomposition for Rotation and Permutation Equivariant Cluster Expansions

Problem Statement
Current machine-learned interatomic potentials (MLIPs) and electronic-structure models, such as the Atomic Cluster Expansion (ACE) and equivariant message-passing networks (e.g., NequIP, MACE, GRACE), typically enforce rotational equivariance ( $SO(3)$ ) and permutation invariance ( $S_N$ ) as separate or sequential steps. Standard approaches often construct an overcomplete basis of rotation products and subsequently symmetrize them with respect to permutations (or average over them early). This "post-hoc" reduction leads to several issues:

Redundancy: The tensor product space contains linearly dependent representations and redundant coupling paths, particularly when tensor product factors have repeated content.
Computational Cost: Constructing and reducing these overcomplete bases becomes computationally intractable for moderate to high tensor-product ranks (e.g., $N > 9$ ).
Information Loss: Early averaging over permutations discards non-trivial permutation information, preventing the modeling of systems requiring specific permutation characters (e.g., antisymmetric fermionic wavefunctions or specific electronic operator symmetries) beyond simple invariance.

Existing methods often rely on Singular Value Decomposition (SVD) or generalized angular-momentum recurrence relations to reduce these overcomplete sets, which is inefficient and lacks a rigorous, simultaneous treatment of joint symmetries.

Methodology: Young–E(3) (YE3T) Decomposition
The authors propose a generalization of cluster expansions that treats rotation and permutation symmetries simultaneously using representation theory. The core methodology involves:

Joint Carrier Definition: The method defines a carrier space for the joint $S_N \times SO(3)$ representations. Instead of treating permutations as a global averaging step, they are viewed as actions permuting tensor factors (or interaction functions) within a specific channel ordering.
Schur–Weyl Decomposition: The authors apply Schur–Weyl decomposition within "blocks" of the tensor product space where channel content is repeated. By utilizing Young subgroup stabilizers ( $G_\nu$ ) and canonical ordering schemes (specifically frequency-based ordering), the method organizes repeated channel content.
Blockwise Reduction: Rather than decomposing the full tensor product space at once, the decomposition is performed blockwise. Within each block of size $k_b$ , the general linear group decomposition is restricted to $SO(3)$ . This yields irreducible representations of the symmetric group $S_{k_b}$ (labeled by Young diagrams $\kappa_b$ ) coupled with angular momentum representations.
Recoupling: The block-decomposed states are then recoupled to form the final parent states. Because rotation and permutation actions commute, the coupling of block rotation carriers and block permutation carriers can be treated simultaneously.
Basis Construction: This process yields a complete, orthonormal basis for arbitrary tensor product ranks and arbitrary angular/radial content. The basis is indexed by the parent Young symmetry $\lambda$ , total angular momentum $L$ , and multiplicity indices.

Key Contributions

YE3T Basis: The derivation of a complete, orthonormal basis for $S_N \times SO(3)$ that is irreducible by construction. This eliminates the need for overcomplete constructions and subsequent numerical reduction.
Generalization of ACE: The paper demonstrates that the standard ACE basis (which is permutation-invariant) corresponds to the specific case where the permutation character is restricted to the fully symmetric carrier ( $\lambda = (N)$ ). Thus, ACE is a subset of the more general YE3T framework.
Non-Trivial Permutation Characters: Unlike ACE, YE3T allows for the systematic inclusion of non-trivial permutation characters (e.g., antisymmetric or mixed symmetry). This enables the construction of equivariant models for fermionic systems and electronic operators where specific permutation symmetries are required.
Analytical Multiplicity Counting: The authors provide an analytical formula (Eq. 10) to calculate the exact number of independent basis functions (multiplicities) for high-rank tensor products without approximation, confirming the fundamental correctness of the approach against existing Lie-algebra kernel methods (GE-PI).

Results

Validation: The YE3T decomposition reproduces the basis sizes and multiplicities of the fully symmetric sector ( $\lambda = (N)$ ) exactly as calculated by the GE-PI method (Barthelemy et al.), confirming the theoretical consistency of the approach.
Performance (Linear Models): In linear MLIP benchmarks (using LAMMPS), YE3T models (both symmetric and mixed) demonstrated a significant speedup (2–5 $\times$ ) over existing PACE (Product ACE) evaluators. The "YE3T-symmetric" models, which are mathematically equivalent to ACE but constructed via irreducible decomposition, achieved this speedup by avoiding overcompleteness.
Scalability: YE3T exhibits superior scaling with the number of descriptors. Adding features to YE3T models incurs negligible computational cost compared to the linear or exponential cost increases seen in traditional ACE constructions.
Accuracy: YE3T models with mixed permutation intermediates ("YE3T-mix") achieved higher accuracy with fewer descriptors compared to standard linear models. In fermionic operator learning (Electron Repulsion Integrals for $H_2CO_3$ and $CO_3^{2-}$ ), models allowing mixed permutation symmetry in hidden layers converged faster and achieved lower validation RMSE than models restricted to pair-antisymmetric symmetry throughout.
Pareto Front: The results define a new Pareto front for MLIPs, offering simultaneous improvements in speed and accuracy.

Significance
The paper claims that YE3T represents a fundamental shift in how physical symmetries are handled in machine-learned potentials. By treating rotation and permutation symmetries simultaneously through Schur–Weyl decomposition, the method:

Resolves the issue of linearly dependent descriptors and overcompleteness that plagues current ACE-based methods.
Provides a rigorous framework for extending equivariant modeling beyond permutation-invariant outputs to include specific permutation characters, which is crucial for electronic structure and fermionic problems.
Delivers immediate practical benefits, including substantial computational speedups (up to 1000 $\times$ in specific coupling benchmarks relative to cuEquivariance) and the ability to construct higher-rank features that were previously intractable.

The authors position YE3T not just as an optimization of ACE, but as a more natural and complete description of particle interaction symmetries, enabling a new class of rotation- and permutation-equivariant models for both atomistic and electronic simulations.

More like this