Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
e9101e2
Publish ad-negf original English article at its canonical URL
floatingCatty Oct 12, 2026
8030781
Add full Chinese ad-negf article with reciprocal hreflang
floatingCatty Oct 12, 2026
b7b1b46
Publish deeptb-negf original English article at its canonical URL
floatingCatty Oct 12, 2026
dae476a
Add full Chinese deeptb-negf article with reciprocal hreflang
floatingCatty Oct 12, 2026
75534bc
Publish ferh-magnetism original English article at its canonical URL
floatingCatty Oct 12, 2026
3d64b87
Add full Chinese ferh-magnetism article with reciprocal hreflang
floatingCatty Oct 12, 2026
b314625
Publish logdet-entropy original English article at its canonical URL
floatingCatty Oct 12, 2026
4548bf6
Add full Chinese logdet-entropy article with reciprocal hreflang
floatingCatty Oct 12, 2026
ebdd746
Preserve existing English prose at /research/moire-phonon/
floatingCatty Oct 12, 2026
be2c6c7
Add Chinese moire-phonon research article with reciprocal hreflang
floatingCatty Oct 12, 2026
1ce9f96
Preserve existing English prose at /research/nqs-impurity/
floatingCatty Oct 12, 2026
b1c92f7
Add Chinese nqs-impurity research article with reciprocal hreflang
floatingCatty Oct 12, 2026
85be2ed
Preserve existing English prose at /research/slem/
floatingCatty Oct 12, 2026
dbb4e41
Add Chinese slem research article with reciprocal hreflang
floatingCatty Oct 12, 2026
4515ede
Preserve existing English prose at /research/vbcsr/
floatingCatty Oct 12, 2026
439a615
Add Chinese vbcsr research article with reciprocal hreflang
floatingCatty Oct 12, 2026
351b372
Make original research catalog fully English and link localized resea…
floatingCatty Oct 12, 2026
618ab34
Publish complete Chinese research catalog with static links
floatingCatty Oct 12, 2026
fd6b077
Route research and About navigation to matching language pages when a…
floatingCatty Oct 12, 2026
54bd94a
Send Chinese homepage visitors directly to corresponding indexed Chin…
floatingCatty Oct 12, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions _includes/lamp-nav.html
Original file line number Diff line number Diff line change
Expand Up @@ -6,9 +6,9 @@
<div class="navlinks">
<a href="{{ '/essays/' | relative_url }}"{% if include.active == "essays" %} class="active"{% endif %}><span class="i18n-zh">随笔</span><span class="i18n-en">Essays</span></a>
<a href="{{ '/study-notes/' | relative_url }}"{% if include.active == "study" %} class="active"{% endif %}><span class="i18n-zh">学习笔记</span><span class="i18n-en">Study notes</span></a>
<a href="{{ '/research/' | relative_url }}"{% if include.active == "research" %} class="active"{% endif %}><span class="i18n-zh">研究</span><span class="i18n-en">Research</span></a>
<a href="{% if page.research_locale == 'zh' %}{{ '/zh/research/' | relative_url }}{% else %}{{ '/research/' | relative_url }}{% endif %}" data-href-en="{{ '/research/' | relative_url }}" data-href-zh="{{ '/zh/research/' | relative_url }}"{% if include.active == "research" %} class="active"{% endif %}><span class="i18n-zh">研究</span><span class="i18n-en">Research</span></a>
<a href="{{ '/publications/' | relative_url }}"{% if include.active == "pubs" %} class="active"{% endif %}><span class="i18n-zh">论文</span><span class="i18n-en">Publications</span></a>
<a href="{{ '/#about' | relative_url }}"><span class="i18n-zh">关于</span><span class="i18n-en">About</span></a>
<a href="{% if page.research_locale == 'zh' %}{{ '/zh/about/' | relative_url }}{% else %}{{ '/about/' | relative_url }}{% endif %}" data-href-en="{{ '/about/' | relative_url }}" data-href-zh="{{ '/zh/about/' | relative_url }}"><span class="i18n-zh">关于</span><span class="i18n-en">About</span></a>
<a href="{{ '/#contact' | relative_url }}"><span class="i18n-zh">联系</span><span class="i18n-en">Contact</span></a>
<button id="lang-toggle" class="lang-btn" type="button" aria-label="Switch language"></button>
</div>
Expand Down
28 changes: 28 additions & 0 deletions _pages/research-ad-negf-zh.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
---
layout: research
permalink: /zh/research/ad-negf/
title: "AD-NEGF:可微的量子输运计算"
title_en: "AD-NEGF: differentiable quantum transport calculations"
description: "AD-NEGF computes gradients through tight-binding quantum-transport calculations for sensitivity analysis and inverse Hamiltonian design."
description_zh: "AD-NEGF 为紧束缚量子输运计算增加梯度,用于灵敏度分析和哈密顿量逆向设计。"
paper_year: 2023
paper_url: "https://doi.org/10.1103/PhysRevB.108.195143"
code_url: "https://github.com/floatingCatty/ADNEGF"
research_locale: zh
translation_url: "/research/ad-negf/"
translation_of: "/research/ad-negf/"
alternate_en_url: "/research/ad-negf/"
alternate_zh_url: "/zh/research/ad-negf/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

我在天津大学做量子输运的时候,常常用 NEGF(非平衡格林函数)方法算电子通过一个器件的透射和电流。给定描述电子能量和耦合的哈密顿量,再加上电极和偏压,程序就能算出结果。但如果想知道某个参数改一点,电流会变多少,往往只能改完再跑一次。参数多的时候,这个办法很慢。

我们当时想试试,能不能直接对整个输运计算求导。比如,电流对某个轨道耦合、原子位置或其他输入变量的梯度。有了这些导数,可以分析模型对什么最敏感,也可以尝试反过来调整哈密顿量,让它接近某个目标输运性质。

AD-NEGF 用 PyTorch 实现了 NEGF 计算流程,并针对部分求解步骤写了专门的反向传播。NEGF 里有矩阵求逆和需要数值求解的中间量,直接依赖一般的自动微分不一定合适,所以论文中也用了隐式求导的方法。我们验证了梯度计算,还展示了从目标输运性质反过来优化哈密顿量的例子。

不过,这里的设计主要还在模型层面。哈密顿量参数怎样对应到实际能够制造的器件结构,散射、缺陷、接触等效应怎么处理,都需要另外建模。能对求解器求导,确实方便了很多原本只能反复试算的工作,但离直接用于真实器件设计,还有不少事情要做。

</div>
16 changes: 4 additions & 12 deletions _pages/research-ad-negf.md
Original file line number Diff line number Diff line change
Expand Up @@ -8,20 +8,12 @@ description_zh: "AD-NEGF 为紧束缚量子输运计算增加梯度,用于灵
paper_year: 2023
paper_url: "https://doi.org/10.1103/PhysRevB.108.195143"
code_url: "https://github.com/floatingCatty/ADNEGF"
research_locale: en
translation_url: "/zh/research/ad-negf/"
alternate_en_url: "/research/ad-negf/"
alternate_zh_url: "/zh/research/ad-negf/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

我在天津大学做量子输运的时候,常常用 NEGF(非平衡格林函数)方法算电子通过一个器件的透射和电流。给定描述电子能量和耦合的哈密顿量,再加上电极和偏压,程序就能算出结果。但如果想知道某个参数改一点,电流会变多少,往往只能改完再跑一次。参数多的时候,这个办法很慢。

我们当时想试试,能不能直接对整个输运计算求导。比如,电流对某个轨道耦合、原子位置或其他输入变量的梯度。有了这些导数,可以分析模型对什么最敏感,也可以尝试反过来调整哈密顿量,让它接近某个目标输运性质。

AD-NEGF 用 PyTorch 实现了 NEGF 计算流程,并针对部分求解步骤写了专门的反向传播。NEGF 里有矩阵求逆和需要数值求解的中间量,直接依赖一般的自动微分不一定合适,所以论文中也用了隐式求导的方法。我们验证了梯度计算,还展示了从目标输运性质反过来优化哈密顿量的例子。

不过,这里的设计主要还在模型层面。哈密顿量参数怎样对应到实际能够制造的器件结构,散射、缺陷、接触等效应怎么处理,都需要另外建模。能对求解器求导,确实方便了很多原本只能反复试算的工作,但离直接用于真实器件设计,还有不少事情要做。

</div>

<div class="post-en" lang="en" markdown="1">

When I was working on quantum transport at Tianjin University, I often used the non-equilibrium Green's function (NEGF) method to calculate electron transmission and current through a device. Given a Hamiltonian describing electronic energies and couplings, along with electrodes and an applied bias, the program produces a result. But if we want to know how much the current changes when a parameter is adjusted, we usually have to run the calculation again. That gets slow when there are many parameters.
Expand Down
28 changes: 28 additions & 0 deletions _pages/research-deeptb-negf-zh.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
---
layout: research
permalink: /zh/research/deeptb-negf/
title: "DeePTB-NEGF:从学习到的哈密顿量计算量子输运"
title_en: "DeePTB-NEGF: transport from learned Hamiltonians"
description: "A DeePTB and DPNEGF workflow for quantum transport in graphene, h-BN and MoS2, compared with first-principles DFT-NEGF calculations."
description_zh: "DeePTB 和 DPNEGF 结合,计算石墨烯、氮化硼和二硫化钼的量子输运,并与 DFT-NEGF 对照。"
paper_year: 2025
paper_url: "https://arxiv.org/abs/2512.11291"
code_url: "https://github.com/deepmodeling/dpnegf"
research_locale: zh
translation_url: "/research/deeptb-negf/"
translation_of: "/research/deeptb-negf/"
alternate_en_url: "/research/deeptb-negf/"
alternate_zh_url: "/zh/research/deeptb-negf/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

做量子输运计算时,NEGF 通常需要先拿到体系的哈密顿矩阵。用 DFT 直接构造当然比较可靠,但材料一大,或者想换很多个结构反复计算,这一步往往就是主要开销。DeePTB 能从原子结构预测紧束缚哈密顿量,于是一个自然的问题是:能不能把它直接接到量子输运求解器里?

这篇合作工作尝试了这件事。我们把 DeePTB 预测的哈密顿量接到 DPNEGF,研究石墨烯、六方氮化硼和 MoS₂ 这三种二维材料,并与传统 DFT-NEGF 得到的能带和电子透射谱做比较。预印本报告了较好的一致性,以及在这些测试体系上的计算效率提升。

这里实际上有两层需要分开验证的近似。一层是学习到的哈密顿量是否准确,另一层是输运求解器能否正确处理接触、器件几何和边界条件。即使能带看起来相近,透射也不一定自然就会准确。所以我觉得把模型和物性计算接在一起,并做完整的输运对照,比单独展示一个好的哈密顿量误差更有意义。

目前公开的是预印本。它说明了这条计算路线在几个典型二维材料上的可行性,但不能据此说所有缺陷、界面或实际器件都已有同等的精度。

</div>
16 changes: 4 additions & 12 deletions _pages/research-deeptb-negf.md
Original file line number Diff line number Diff line change
Expand Up @@ -8,20 +8,12 @@ description_zh: "DeePTB 和 DPNEGF 结合,计算石墨烯、氮化硼和二硫
paper_year: 2025
paper_url: "https://arxiv.org/abs/2512.11291"
code_url: "https://github.com/deepmodeling/dpnegf"
research_locale: en
translation_url: "/zh/research/deeptb-negf/"
alternate_en_url: "/research/deeptb-negf/"
alternate_zh_url: "/zh/research/deeptb-negf/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

做量子输运计算时,NEGF 通常需要先拿到体系的哈密顿矩阵。用 DFT 直接构造当然比较可靠,但材料一大,或者想换很多个结构反复计算,这一步往往就是主要开销。DeePTB 能从原子结构预测紧束缚哈密顿量,于是一个自然的问题是:能不能把它直接接到量子输运求解器里?

这篇合作工作尝试了这件事。我们把 DeePTB 预测的哈密顿量接到 DPNEGF,研究石墨烯、六方氮化硼和 MoS₂ 这三种二维材料,并与传统 DFT-NEGF 得到的能带和电子透射谱做比较。预印本报告了较好的一致性,以及在这些测试体系上的计算效率提升。

这里实际上有两层需要分开验证的近似。一层是学习到的哈密顿量是否准确,另一层是输运求解器能否正确处理接触、器件几何和边界条件。即使能带看起来相近,透射也不一定自然就会准确。所以我觉得把模型和物性计算接在一起,并做完整的输运对照,比单独展示一个好的哈密顿量误差更有意义。

目前公开的是预印本。它说明了这条计算路线在几个典型二维材料上的可行性,但不能据此说所有缺陷、界面或实际器件都已有同等的精度。

</div>

<div class="post-en" lang="en" markdown="1">

A quantum transport calculation usually begins with a Hamiltonian. Constructing it directly from DFT can be reliable, but it becomes expensive when the system is large or when we want to compare many different structures. DeePTB can predict a tight-binding Hamiltonian from an atomic structure, which raises a natural question: can we use that Hamiltonian directly in a quantum transport solver?
Expand Down
25 changes: 25 additions & 0 deletions _pages/research-ferh-magnetism-zh.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
---
layout: research
permalink: /zh/research/ferh-magnetism/
title: "FeRh 中交换关联泛函对磁性的影响"
title_en: "Exchange-correlation functionals and magnetism in FeRh"
description: "A note on why density-functional predictions of magnetism in FeRh change with lattice constant and exchange-correlation functional."
description_zh: "关于 FeRh 合金的磁性预测为什么依赖晶格常数和 DFT 交换关联泛函。"
paper_year: 2025
paper_url: "https://doi.org/10.1016/j.commatsci.2024.113561"
research_locale: zh
translation_url: "/research/ferh-magnetism/"
translation_of: "/research/ferh-magnetism/"
alternate_en_url: "/research/ferh-magnetism/"
alternate_zh_url: "/zh/research/ferh-magnetism/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

FeRh 是铁和铑组成的一种合金,磁性对晶格常数特别敏感。原子之间的距离只要有一点变化,不同磁性状态之间的能量关系就可能跟着改变。这给 DFT 计算出了一个难题:算出来的结果,到底有多少来自材料本身,有多少来自我们使用的近似?

在 DFT 里,电子的交换关联效应需要用一个泛函来近似。实际做计算时,可选的方法有很多,例如 LDA、GGA 和 SCAN。它们在一些材料上可以给出很接近的结果,换一个体系却不一定如此。FeRh 就是一个很适合拿来比较的例子,因为结构、电子和磁性之间联系得很紧。

这篇工作比较了几种交换关联泛函对 FeRh 的描述,包括晶格常数和磁性状态等结果。我们发现,没有哪一个泛函能在所有相关性质上都给出同样可靠的描述。因此研究 FeRh 这类体系时,需要关心计算对泛函和结构参数的依赖,不能只报出一次计算得到的磁性结果。

</div>
14 changes: 4 additions & 10 deletions _pages/research-ferh-magnetism.md
Original file line number Diff line number Diff line change
Expand Up @@ -7,18 +7,12 @@ description: "A note on why density-functional predictions of magnetism in FeRh
description_zh: "关于 FeRh 合金的磁性预测为什么依赖晶格常数和 DFT 交换关联泛函。"
paper_year: 2025
paper_url: "https://doi.org/10.1016/j.commatsci.2024.113561"
research_locale: en
translation_url: "/zh/research/ferh-magnetism/"
alternate_en_url: "/research/ferh-magnetism/"
alternate_zh_url: "/zh/research/ferh-magnetism/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

FeRh 是铁和铑组成的一种合金,磁性对晶格常数特别敏感。原子之间的距离只要有一点变化,不同磁性状态之间的能量关系就可能跟着改变。这给 DFT 计算出了一个难题:算出来的结果,到底有多少来自材料本身,有多少来自我们使用的近似?

在 DFT 里,电子的交换关联效应需要用一个泛函来近似。实际做计算时,可选的方法有很多,例如 LDA、GGA 和 SCAN。它们在一些材料上可以给出很接近的结果,换一个体系却不一定如此。FeRh 就是一个很适合拿来比较的例子,因为结构、电子和磁性之间联系得很紧。

这篇工作比较了几种交换关联泛函对 FeRh 的描述,包括晶格常数和磁性状态等结果。我们发现,没有哪一个泛函能在所有相关性质上都给出同样可靠的描述。因此研究 FeRh 这类体系时,需要关心计算对泛函和结构参数的依赖,不能只报出一次计算得到的磁性结果。

</div>

<div class="post-en" lang="en" markdown="1">

FeRh is an alloy of iron and rhodium whose magnetism is unusually sensitive to the lattice constant. A small change in the distance between atoms can change the relative energies of different magnetic states. That makes it a difficult system for DFT: how much of a prediction comes from the material, and how much comes from the approximation we chose?
Expand Down
32 changes: 32 additions & 0 deletions _pages/research-logdet-entropy-zh.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,32 @@
---
layout: research
permalink: /zh/research/logdet-entropy/
title: "LogDet:用矩阵估计神经网络里的信息量"
title_en: "LogDet: estimating information in neural networks"
description: "The LogDet entropy estimator measures high-dimensional information using matrix operations and was applied to the analysis of neural networks."
description_zh: "LogDet 熵估计器利用矩阵运算处理高维数据,用于分析神经网络特征中的信息变化。"
paper_year: 2025
paper_url: "https://doi.org/10.1016/j.neucom.2025.131520"
paper_title: "Understanding neural networks with logarithm determinant entropy estimator"
paper_doi: "10.1016/j.neucom.2025.131520"
paper_authors:
- "Zhanghao Zhouyin"
- "Ding Liu"
research_locale: zh
translation_url: "/research/logdet-entropy/"
translation_of: "/research/logdet-entropy/"
alternate_en_url: "/research/logdet-entropy/"
alternate_zh_url: "/zh/research/logdet-entropy/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

我本科时做过一段时间神经网络理论,最感兴趣的是网络内部的信息到底发生了什么变化。比如一层层特征传递下去,哪些东西保留下来了,哪些被丢掉了。信息论给我们提供了熵和互信息这些工具,但真要把它们用在高维神经网络特征上,估计本身就不容易。

直接估计高维概率分布需要大量样本。一些常见的熵估计器,在维度很高或者数据分布比较特殊时也容易出现偏差。我们于是试了一种从矩阵出发的办法:利用样本构成的矩阵及其对数行列式(LogDet)来估计信息量,省去直接重建完整概率密度的步骤。

论文中我们做了几组对比实验,也把估计器用来观察神经网络训练过程中不同层的特征变化,希望借此分析信息瓶颈理论里讨论的现象。这篇工作很早就开始了,后来经过几轮修改,发表在 2025 年的 Neurocomputing。

这类指标的意义也要放在具体估计方法里看。它可以帮助比较和观察网络,但不能简单地把一个估计数值当作对网络内部信息的完整解释。我后来转到物理计算,仍然常常碰到类似的问题:先要弄清楚我们用的量到底测量了什么,才谈得上从结果里得出结论。

</div>
16 changes: 4 additions & 12 deletions _pages/research-logdet-entropy.md
Original file line number Diff line number Diff line change
Expand Up @@ -12,20 +12,12 @@ paper_doi: "10.1016/j.neucom.2025.131520"
paper_authors:
- "Zhanghao Zhouyin"
- "Ding Liu"
research_locale: en
translation_url: "/zh/research/logdet-entropy/"
alternate_en_url: "/research/logdet-entropy/"
alternate_zh_url: "/zh/research/logdet-entropy/"
---

<div class="post-zh" lang="zh-Hans" markdown="1">

我本科时做过一段时间神经网络理论,最感兴趣的是网络内部的信息到底发生了什么变化。比如一层层特征传递下去,哪些东西保留下来了,哪些被丢掉了。信息论给我们提供了熵和互信息这些工具,但真要把它们用在高维神经网络特征上,估计本身就不容易。

直接估计高维概率分布需要大量样本。一些常见的熵估计器,在维度很高或者数据分布比较特殊时也容易出现偏差。我们于是试了一种从矩阵出发的办法:利用样本构成的矩阵及其对数行列式(LogDet)来估计信息量,省去直接重建完整概率密度的步骤。

论文中我们做了几组对比实验,也把估计器用来观察神经网络训练过程中不同层的特征变化,希望借此分析信息瓶颈理论里讨论的现象。这篇工作很早就开始了,后来经过几轮修改,发表在 2025 年的 Neurocomputing。

这类指标的意义也要放在具体估计方法里看。它可以帮助比较和观察网络,但不能简单地把一个估计数值当作对网络内部信息的完整解释。我后来转到物理计算,仍然常常碰到类似的问题:先要弄清楚我们用的量到底测量了什么,才谈得上从结果里得出结论。

</div>

<div class="post-en" lang="en" markdown="1">

As an undergraduate, I spent some time studying neural-network theory. I was interested in what happens to information inside a network: what survives as features pass through layers, and what gets discarded. Information theory gives us entropy and mutual information, but estimating them for high-dimensional neural-network features is difficult in practice.
Expand Down
Loading
Loading