Jadwal Sholat

Memuat jadwal sholat…

Fisika & Matematika editorial

Open AccessOA2023

Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian

QHNet is a SE(3)-equivariant graph network that predicts quantum Hamiltonian matrices efficiently by reducing tensor products by 92% and preventing exponential channel growth, achieving state-of-the-art performance with faster speed and 50% less memory on MD17 datasets.
Hai-Yang Yu; Zhao Xu; X. Qian; Xiao-Ning Qian; Shui-Wang Ji· International Conference on Machine Learning· 2023· DOI 10.48550/arXiv.2306.04922

The core problem

Predicting the Hamiltonian matrix is a fundamental task in quantum chemistry and condensed matter physics, as it encodes the interactions and energetics of molecular systems. Traditional methods, such as density functional theory (DFT), are computationally expensive, motivating machine learning approaches. However, two critical factors—efficiency and equivariance—are often in conflict: equivariant models that respect the underlying symmetries (e.g., rotation and translation) tend to be computationally heavy due to the extensive use of tensor products. This work introduces QHNet, a SE(3)-equivariant graph network that achieves both efficiency and equivariance. The key innovation lies in its architecture, which reduces the number of tensor products by 92% and prevents the exponential growth of channel dimensions when handling multiple atom types. Experiments on MD17 datasets demonstrate that QHNet achieves comparable performance to state-of-the-art methods at significantly faster speeds and with 50% less memory usage. The code is publicly available as part of the AIRS library.

Innovation

QHNet is designed as a SE(3)-equivariant graph neural network that predicts the Hamiltonian matrix directly from atomic coordinates and types. The architecture leverages the symmetry properties of the Hamiltonian under rotations and translations, ensuring that predictions transform equivariantly. The key methodological advance is the reduction of tensor products—operations that are computationally expensive and scale poorly with the number of channels. By innovatively restructuring the network, QHNet decreases the number of tensor products by 92% compared to conventional equivariant networks. Additionally, QHNet introduces a mechanism to prevent the exponential growth of channel dimensions as the number of atom types increases, which is a common bottleneck in multi-component systems. This is achieved through a careful design of the feature aggregation and message passing schemes. The network processes molecular graphs where nodes represent atoms and edges represent interactions, and outputs the Hamiltonian matrix in a block-wise manner. The overall architecture can be visualized as:

The equivariance is enforced by using spherical harmonics and Clebsch-Gordan coefficients, but with a streamlined tensor product scheme that avoids redundant computations.

Introduction
Predicting the Hamiltonian matrix is a fundamental task in quantum chemistry and condensed matter physics, as it encodes the interactions and energetics of molecular systems. Traditional methods, such as density functional theory (DFT), are computationally expensive, motivating machine learning approaches. However, two critical factors—efficiency and equivariance—are often in conflict: equivariant models that respect the underlying symmetries (e.g., rotation and translation) tend to be computationally heavy due to the extensive use of tensor products. This work introduces QHNet, a SE(3)-equivariant graph network that achieves both efficiency and equivariance. The key innovation lies in its architecture, which reduces the number of tensor products by 92% and prevents the exponential growth of channel dimensions when handling multiple atom types. Experiments on MD17 datasets demonstrate that QHNet achieves comparable performance to state-of-the-art methods at significantly faster speeds and with 50% less memory usage. The code is publicly available as part of the AIRS library.

Why it matters

The key insight of QHNet is that equivariance does not necessarily require an excessive number of tensor products. By carefully designing the network to share computations and avoid redundant tensor products, QHNet maintains the expressive power needed for accurate Hamiltonian prediction while drastically reducing computational cost. The 92% reduction in tensor products is achieved through a combination of factorized operations and a novel message passing scheme that exploits the sparsity of molecular graphs. The prevention of exponential channel growth is particularly important for systems with many atom types, as it allows the network to scale linearly with the number of types rather than exponentially. This is achieved by using a type-specific embedding that is shared across tensor operations, and by aggregating messages in a way that does not require separate channels for each type combination. The authors also discuss the limitations: QHNet currently focuses on molecular systems and may need adaptation for periodic systems. Future work includes extending QHNet to larger molecules and exploring its application to other quantum properties. Overall, QHNet represents a significant step towards practical equivariant models for quantum chemistry, offering a blueprint for efficient architecture design.

Who should read this

Pembaca Research Digest dan praktisi rumpun ini.

Opening member content…