DeepH-2：通过等变局部坐标Transformer增强深度学习电子结构
Universal Hamiltonian learning with rotational and basis equivariance
AI总结：
针对现有神经网络架构难以广泛通用的问题，本研究提出等变局部坐标Transformer框架DeepH-2，融合局部坐标变换与等变神经网络优势，在电子结构计算的效率和精度上均超越前代，达到最优性能，为通用大材料模型研究开辟新路径。
摘要：
深度学习电子结构计算在革新计算材料研究格局方面展现出巨大潜力。然而，当前的神经网络架构被认为不适合广泛的通用应用。本文提出一种等变局部坐标Transformer框架，旨在增强深度学习密度泛函理论哈密顿量，该方法被命名为DeepH-2。与DeepH、DeepH-E3等以往模型不同，DeepH-2将局部坐标变换的简洁性与等变神经网络的数学优雅性无缝融合，有效克服了二者各自的缺陷。综合实验表明，DeepH-2在效率与精度上均优于前代模型，展现出当前最优的性能。这一进展为探索通用神经网络模型乃至大材料模型开辟了机遇。
英文摘要：
Deep learning has the potential to transform computational materials research by making electronic-structure calculations more efficient and scalable. The Hamiltonian matrix in the localized atomic-orbital basis is a promising target for deep-learning electronic structure, yet enforcing rotational equivariance can be computationally costly, and the matrix depends on the choice of basis set. Here we introduce DeepH-2, a physics-guided neural-network framework that reduces the computational cost of enforcing rotational equivariance through a local-coordinate formulation and incorporates basis equivariance by mapping a shared latent Hamiltonian to the target atomic-orbital basis. DeepH-2 achieves sub-meV test-set error on a diverse materials dataset and demonstrates transferability across material dimensions. These results highlight the value of encoding physics priors into deep-learning electronic structure models, and pave the way for developing the electronic-structure foundation models.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
