研究Research

这里放一些我做过的研究和当时的想法。欢迎交流。A few projects I've worked on, and some thoughts about them. Feel free to get in touch.

2026 · arXiv

魔角石墨烯里的声子凝聚Moiré phonon condensation in twisted bilayer graphene

把上万个原子的结构变化投影到声子模式上,发现主要由两个集体模式描述。Projecting a reconstruction of more than ten thousand atoms onto phonon modes reveals that two collective modes account for most of it.

阅读Read ↗
2026 · Physical Review B

用神经网络求解量子嵌入里的杂质问题Neural quantum states for quantum embedding

用神经量子态处理强关联电子的杂质模型,也研究了自洽迭代里高精度采样的开销。Using neural quantum states for an impurity model of interacting electrons, including the cost of accurate sampling in self-consistent calculations.

阅读Read ↗
2025 · Computational Materials Science

FeRh 中交换关联泛函对磁性的影响Exchange-correlation functionals and magnetism in FeRh

比较几种 DFT 泛函对 FeRh 结构和磁性的预测,看看结果为什么会不同。Comparing how several DFT functionals describe the structure and magnetism of FeRh.

阅读Read ↗
2025 · ICLR

SLEM:从原子结构预测量子算符SLEM: predicting quantum operators from atomic structures

在网络里加入局域性和旋转对称性的约束,预测哈密顿量、重叠矩阵等量子算符。Predicting Hamiltonian, overlap and other quantum-operator matrices using a model built around locality and rotational symmetry.

阅读Read ↗
2025 · Neurocomputing

LogDet:用矩阵估计神经网络里的信息量LogDet: estimating information in neural networks

用对数行列式构造高维熵估计器,观察神经网络训练时的信息变化。A matrix-based entropy estimator for high-dimensional data, used to study information changes during neural-network training.

阅读Read ↗
2024 · Nature Communications

DeePTB:用机器学习做大规模电子结构计算DeePTB: machine learning for large-scale electronic structure

从第一性原理数据学习随原子环境变化的紧束缚参数,用来计算更大材料的电子结构。Learning environment-dependent tight-binding parameters from first-principles data to calculate electronic structure in larger materials.

阅读Read ↗
2023 · Physical Review B

AD-NEGF:可微的量子输运计算AD-NEGF: differentiable quantum transport calculations

让 NEGF 的输运结果可以对模型参数求导,用于灵敏度分析和哈密顿量逆向设计。Computing derivatives of NEGF transport results with respect to model parameters for sensitivity analysis and Hamiltonian design.

阅读Read ↗

软件Software

Open source

flower:把 Agent 跑过的计算整理清楚flower: keeping scientific-agent calculations reproducible

记录 Agent 的计算步骤、参数和结果,把有用的流程导出成可以重复运行的 protocol。Recording an agent's steps, parameters and results, then exporting useful runs as reproducible protocols.

阅读Read ↗
Open source

VBCSR:给大规模稀疏矩阵写的一套工具VBCSR: distributed sparse matrices for scientific computing

支持 MPI 的变块稀疏矩阵与乘法工具,主要用于电子结构等科学计算。Variable-block sparse matrices and multiplication across MPI processes, mainly for electronic structure and related numerical work.

阅读Read ↗

完整论文见The papers are also listed under 论文Publications ↗.