Learning local equivariant representations for quantum operators
Zhanghao Zhouyin, Zixi Gan, MingKang Liu, Shishir Kumar, Pandey, Linfeng Zhang, Qiangqiang Gu
TL;DR
This paper introduces SLEM, a novel deep learning model that predicts multiple quantum operators with high accuracy and efficiency by leveraging strict locality and equivariance, enabling scalable quantum simulations.
Contribution
The paper presents SLEM, a new local equivariant deep learning model that improves accuracy and scalability in predicting quantum operators for large systems.
Findings
Achieves state-of-the-art accuracy in quantum operator prediction.
Significantly improves computational efficiency and scalability.
Effective across diverse 2D and 3D materials.
Abstract
Predicting quantum operator matrices such as Hamiltonian, overlap, and density matrices in the density functional theory (DFT) framework is crucial for material science. Current methods often focus on individual operators and struggle with efficiency and scalability for large systems. Here we introduce a novel deep learning model, SLEM (strictly localized equivariant message-passing) for predicting multiple quantum operators, that achieves state-of-the-art accuracy while dramatically improving computational efficiency. SLEM's key innovation is its strict locality-based design for equivariant representations of quantum tensors while preserving physical symmetries. This enables complex many-body dependency without expanding the effective receptive field, leading to superior data efficiency and transferability. Using an innovative SO(2) convolution and invariant overlap parameterization,…
Peer Reviews
Decision·ICLR 2025 Spotlight
- The two key contributions are novel, they are clearly explained in the paper and based on my understanding they are technically sound. - The strictly local structure is significant and likely to be widely adopted in the future given the nice properties, not only more parallelizable, but also leads to lower errors.
I have a major concern on the way the evaluation is carried out - The training and testing happens on the same system using trajectories of molecular dynamics. Although this type of evaluation may be also used in the baselines that the author compare to, I feel it is not sufficient. A good generalization is not surprising if the MD trajectories have a good coverage of different atomic geometric configurations. We run into the chicken egg dilemma, if DFT is already calculated on a system, why wo
By focusing on a strictly localized equivariant message-passing framework, the authors present a creative way to address the challenges of efficiency and scalability in quantum mechanical computations. The use of SO(2) convolutions to manage high-order tensor complexity is particularly novel, as it reduces the computational burden associated with f and g orbitals. The methodological rigor of the paper is detailed theoretical justifications for the design choices of SLEM. The authors provide mat
The experimental comparisons presented in the paper are limited to only two other models, and these comparisons are not consistently provided for all experiments. Expanding the range of baseline models, including more well-established methods, would strengthen the validation of SLEM’s computational efficiency, scalability, and accuracy. Incorporating additional well-known benchmark datasets, such as QH9 [1], nablaDFT [2], and potentially QM9 [3, 4] (used in models like HamGNN), could provide a m
The main innovation and strength of this paper is combining strict locality, with an architecture heavily inspired in Allegro, with the SO(2) convolution trick, offering a compelling solution to handling heavy atoms and large systems efficiently while maintaining accuracy. The work can be considered well validated through comprehensive benchmarks, and they demonstrate the previous claims of both improved accuracy and reduced computational costs compared to state-of-the-art methods. The authors a
A major limitation of the paper is its insufficient demonstration of transferability. While the authors show good performance on individual systems, they train and evaluate on the same type of material (e.g., training on Si and testing on Si configurations). There's no evaluation of cross material transferability. For instance, training on light elements and testing on heavy elements, or training on one crystal structure and testing on another. Also, the parallelization benefits, while promising
Code & Models
Videos
Taxonomy
TopicsGaussian Processes and Bayesian Inference · Quantum Computing Algorithms and Architecture · Quantum Mechanics and Applications
MethodsFocus · Convolution