From fdf46a2aadede85d92a3b3a57342e68295b3901b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:57:54 -0400 Subject: [PATCH 01/15] Unpublish technical synthesis until RSATB work is released --- ...rch-learned-hamiltonians-to-observables.md | 134 ------------------ 1 file changed, 134 deletions(-) delete mode 100644 _pages/research-learned-hamiltonians-to-observables.md diff --git a/_pages/research-learned-hamiltonians-to-observables.md b/_pages/research-learned-hamiltonians-to-observables.md deleted file mode 100644 index a2c2567..0000000 --- a/_pages/research-learned-hamiltonians-to-observables.md +++ /dev/null @@ -1,134 +0,0 @@ ---- -layout: research -permalink: /research/learned-hamiltonians-to-observables/ -title: "从机器学习哈密顿量到真正可计算的电子物性" -title_en: "From learned Hamiltonians to scalable electronic properties" -description: "A technical guide to turning learned electronic Hamiltonians into density of states, optical response and quantum transport, with published DeePTB, SLEM, VBCSR and NEGF examples." -description_zh: "从 DeePTB 和 SLEM 预测哈密顿量，到利用稀疏算法计算态密度、光学响应和器件输运：一份有原始论文依据的技术路线说明。" -paper_year: 2026 -project_kind: guide -tags: - - Electronic structure - - Scientific computing -paper_url: "https://doi.org/10.1038/s41467-024-51006-4" -code_url: "https://github.com/deepmodeling/DeePTB" -citation_urls: - - "https://doi.org/10.1038/s41467-024-51006-4" - - "https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html" - - "https://arxiv.org/abs/2609.04397" - - "https://doi.org/10.1103/PhysRevB.108.195143" - - "https://doi.org/10.1038/s41524-025-01853-6" - - "https://doi.org/10.1038/s41524-026-02137-3" -research_locale: en -translation_url: "/zh/research/learned-hamiltonians-to-observables/" -alternate_en_url: "/research/learned-hamiltonians-to-observables/" -alternate_zh_url: "/zh/research/learned-hamiltonians-to-observables/" ---- - -
- -An accurate Hamiltonian is not yet an electronic-structure calculation in the sense that an experimentalist or device engineer needs. It describes a system. It does not, by itself, tell us its conductivity, the charge around a defect, or what happens when we put contacts on either side of a channel. - -I have come across this distinction repeatedly while working on learned Hamiltonians, quantum transport and large sparse-matrix software. The prediction of an operator may take seconds, yet the next step still involves a numerical problem that is too large to solve naively. This note collects the **publicly documented pieces** of that route. It is a technical reading map, not a new paper or a report of unpublished benchmarks. - -## One calculation, several quite different problems - -A useful way to organize the work is: - -**Atomic structure and environment → Hamiltonian and overlap → scalable linear algebra → observable → physical validation.** - -These steps have different failure modes. Learning a matrix accurately is mostly a modeling problem. Applying it to a million-atom system is partly a data-layout and parallel-computing problem. Computing optical or transport response introduces another set of physical assumptions. The distinction matters because success at one stage is not evidence that the later stages work. - -| Layer | Main question | Something that must be checked | -| --- | --- | --- | -| Operator prediction | Can a model produce the right matrix on structures it has not seen? | Band energies, matrix symmetries, local environments outside the training set | -| Orbital representation | What is the basis in which the model acts? | Overlap matrix, orbital ordering, basis transformations and Hermiticity | -| Numerical evaluation | Can we apply or factor the needed sparse operators at the target size? | Accuracy, memory, communication, and time-to-solution | -| Physical observable | Is the chosen quantity formulated correctly for this basis and boundary condition? | Spectra, charge, optical response, or transmission against a reference | -| Engineering interpretation | Does a computed difference correspond to a measurable effect? | Contacts, electrostatics, temperature, disorder and uncertainty | - -## Learning an operator: DeePTB and SLEM solve related but different problems - -In [DeePTB](https://doi.org/10.1038/s41467-024-51006-4), we learned environment-dependent parameters of a compact Slater–Koster tight-binding Hamiltonian from first-principles electronic information. The point was not merely to interpolate one band diagram: the same trained model could be evaluated along atomic trajectories with local distortions. The published finite-temperature GaP example then used a **learned Hamiltonian and tight-binding propagation** on a structure of one million atoms. It was **not** a million-atom self-consistent DFT calculation. [Project explanation]({{ '/research/deeptb/' | relative_url }}) · [Code](https://github.com/deepmodeling/DeePTB). - -[SLEM](https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html) starts closer to the matrices used in atomic-orbital DFT: the Hamiltonian, overlap and other quantum operators. Its strictly local, equivariant architecture deals explicitly with the transformation of orbital tensors under rotations. That buys something different from compact tight binding, but it also means that the **nonorthogonal basis is part of the computational problem**, not a detail to ignore. [Project explanation]({{ '/research/slem/' | relative_url }}). - -In a nonorthogonal basis, the stationary problem is the generalized eigenproblem - -$$ -Hc_n = E_n S c_n, -$$ - -where the overlap matrix $S$ defines the orbital metric. If we simply pass the learned $H$ to an algorithm that assumes $S=I$, we may solve the wrong problem. We also need to verify that the predicted matrices obey the symmetries and numerical conditions expected by the downstream solver. Small errors in an operator need not lead to equally small errors in every derived property, particularly near delicate spectral features. - -## Getting a spectrum is not the same as diagonalizing everything - -For a finite basis with a Hermitian $H$ and positive-definite $S$, one can write the retarded resolvent and density of states as - -$$ -G^r(E)=\left[(E+i\eta)S-H\right]^{-1}, -$$ - -$$ -D_\eta(E)=-\frac{1}{\pi}\operatorname{Im}\operatorname{Tr}\!\left[S\,G^r(E)\right],\qquad \eta>0. -$$ - -This is a broadened density of states. The appearance of $S$ in the trace is not optional: it is how the metric enters this representation. The formula describes the target quantity; it does **not** mean we should form the dense inverse of a huge matrix. - -For a well-localized sparse Hamiltonian, a matrix-vector product can be much cheaper than full diagonalization. Polynomial propagation and stochastic trace methods can exploit that distinction when their convergence, statistical error and required resolution are controlled. The important caveat is that a cheap action of $H$ does not automatically make every operation involving $S^{-1}$ cheap. Nor does it guarantee that sharply resolved local or response properties can be evaluated at the same cost as a coarse total DOS. - -The 2024 DeePTB work provides a concrete published example: DeePTB Hamiltonians were post-processed with the tight-binding propagation method to obtain finite-temperature DOS and optical properties for a million-atom GaP structure. [Nature Communications 15, 6772](https://doi.org/10.1038/s41467-024-51006-4). Later, a separate [npj Computational Materials study](https://doi.org/10.1038/s41524-026-02137-3) combined DeePMD trajectory sampling with HSE-trained DeePTB models to study phonon-assisted optical properties in supercells up to 4,096 atoms. The two papers answer different scale/accuracy questions and should not be conflated. - -## The sparse-matrix problem is physical as well as computational - -Atomic-orbital matrices carry local structure. If an atom has several orbitals, its interactions with another atom appear as a small dense block. Different elements may have different numbers of orbitals, so the block dimensions are not necessarily uniform. Storing every entry as an unrelated scalar can lose the opportunity to use efficient block kernels. - -[VBCSR](https://arxiv.org/abs/2609.04397) preserves those variable-sized blocks, groups compatible block shapes for optimized operations, and distributes matrices across MPI processes. Its public preprint reports sparse-matrix benchmarks and an application to an InP nanoparticle containing more than one million atoms. [Project explanation and runnable benchmark entry]({{ '/research/vbcsr/' | relative_url }}) · [Code](https://github.com/DeepElectron/vbcsr). - -It is worth separating three claims: a block representation saves overhead for some sparsity patterns; a particular multiplication is faster on the tested hardware; and an entire physical calculation becomes cheaper at a larger scale. These are not interchangeable. Thresholding small entries can keep intermediate matrices sparse, but it changes the approximation and needs an observable-level error check. MPI communication, memory growth and the block-size distribution matter as much as a kernel microbenchmark. - -## Transport changes the question: the system is open - -A bulk spectrum says little about the current through a device unless the contacts, boundary conditions and electrostatics are also specified. In coherent NEGF, the device Green function contains the lead self-energies: - -$$ -G^r_C(E)=\left[(E+i\eta)S_C-H_C-\Sigma^r_L(E)-\Sigma^r_R(E)\right]^{-1}. -$$ - -The transmission depends on $G^r_C$ and its coupling to the leads, not only on the isolated device's eigenvalues. In nonorthogonal formulations the lead-device coupling and related observables must also be treated consistently with overlap; the compact expression here does not replace a full implementation. - -This is why validating only the band structure is not enough. The [2025 npj Computational Materials DeePTB-NEGF paper](https://doi.org/10.1038/s41524-025-01853-6) used learned Hamiltonians for more than 10,000 break-junction configurations and for carbon-nanotube field-effect transistors, including an electrostatically gated 41 nm channel benchmarked against experimental transfer characteristics. It also reported a 180 nm nanotube **zero-bias transmission** calculation. Those last two are distinct tests; the 180 nm result should not be mistaken for a fully gated, finite-bias transistor benchmark. - -The [2D-material transport work](https://arxiv.org/abs/2512.11291) tests graphene, h-BN and MoS₂ against DFT-NEGF transmission calculations. It is a separate study of representative materials, not evidence that all defects, contacts and industrial device geometries are already covered. [DeePTB-NEGF explanation]({{ '/research/deeptb-negf/' | relative_url }}). - -A different direction is to differentiate the transport calculation itself. [AD-NEGF, Physical Review B 108, 195143 (2023)](https://doi.org/10.1103/PhysRevB.108.195143) uses automatic differentiation and implicit-layer methods within a tight-binding NEGF model for response derivatives and inverse design. That does not turn an approximate Hamiltonian into an exact one; it makes the chosen model's calculations differentiable. [Project explanation]({{ '/research/ad-negf/' | relative_url }}) · [Code](https://github.com/floatingCatty/ADNEGF). - -## What I would check before believing a large calculation - -When a new model or fast solver is described as accurate and scalable, I try to separate the evidence into a few concrete questions: - -1. **Model domain:** Which structures, species, distortions and electronic states were covered by the reference data? Was the test genuinely held out? -2. **Operator consistency:** Are $H$ and, when present, $S$ Hermitian in the chosen representation? Is $S$ numerically usable, and are the basis conventions recorded? -3. **Observable accuracy:** What is the error in the **quantity being used**—DOS peak positions, charge, optical spectrum, transmission, or a response coefficient—rather than only in a training loss? -4. **Numerical convergence:** Does the answer change with broadening, polynomial order, stochastic samples, solver tolerance, sparse threshold or lead self-energy accuracy? -5. **Scaling evidence:** Are wall-clock time, memory, node count and communication reported for the **entire useful calculation** with a fixed accuracy target, not only for a favorable matrix kernel? - -That list is also useful for a coding agent. An agent can assemble a model, run a sparse calculation and generate plots. It should also record the input, training-model version, operator conventions, solver parameters, reference and failed checks. Otherwise a fast automated run can be very difficult for another scientist to interpret. My separate [flower project](https://github.com/floatingCatty/flower) focuses on the reproducibility side, and its [worked scientific-agent example](https://github.com/floatingCatty/flower/blob/main/docs/REPRODUCING_AGENT_CALCULATIONS.md) shows how a run can be exported and compared numerically. - -## Where this line of work leads - -The useful unit of scientific software is not just a model that returns a matrix. It is a workflow that turns a structure into a physically meaningful answer, with the approximations and failure modes still visible. Learned quantum operators, sparse numerical infrastructure and observable-specific solvers each remove a different bottleneck. Together they make new scales and new kinds of questions accessible—but only where the whole chain has been validated. - -This note draws on already public papers and repositories. It intentionally does not describe methods or results from manuscripts that have not yet been released. - -### Original sources and code - -- [DeePTB — Nature Communications **15**, 6772 (2024)](https://doi.org/10.1038/s41467-024-51006-4) · [GitHub](https://github.com/deepmodeling/DeePTB) -- [SLEM — ICLR 2025 proceedings](https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html) -- [VBCSR — arXiv:2609.04397 (2026 preprint)](https://arxiv.org/abs/2609.04397) · [GitHub](https://github.com/DeepElectron/vbcsr) -- [AD-NEGF — Physical Review B **108**, 195143 (2023)](https://doi.org/10.1103/PhysRevB.108.195143) -- [DeePTB-NEGF — npj Computational Materials **11**, 375 (2025)](https://doi.org/10.1038/s41524-025-01853-6) -- [Phonon-assisted optics — npj Computational Materials **12**, 268 (2026)](https://doi.org/10.1038/s41524-026-02137-3) -- [DeePTB-NEGF for 2D materials — arXiv:2512.11291](https://arxiv.org/abs/2512.11291) - -
From 35b1b543d09f420c898617eeb95eb710a76eb507 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:57:59 -0400 Subject: [PATCH 02/15] Unpublish technical synthesis until RSATB work is released --- ...-learned-hamiltonians-to-observables-zh.md | 134 ------------------ 1 file changed, 134 deletions(-) delete mode 100644 _pages/research-learned-hamiltonians-to-observables-zh.md diff --git a/_pages/research-learned-hamiltonians-to-observables-zh.md b/_pages/research-learned-hamiltonians-to-observables-zh.md deleted file mode 100644 index 3d288da..0000000 --- a/_pages/research-learned-hamiltonians-to-observables-zh.md +++ /dev/null @@ -1,134 +0,0 @@ ---- -layout: research -permalink: /zh/research/learned-hamiltonians-to-observables/ -title: "从机器学习哈密顿量到真正可计算的电子物性" -title_en: "From learned Hamiltonians to scalable electronic properties" -description: "A technical guide to turning learned electronic Hamiltonians into density of states, optical response and quantum transport, with published DeePTB, SLEM, VBCSR and NEGF examples." -description_zh: "从 DeePTB 和 SLEM 预测哈密顿量，到利用稀疏算法计算态密度、光学响应和器件输运：一份有原始论文依据的技术路线说明。" -paper_year: 2026 -project_kind: guide -tags: - - Electronic structure - - Scientific computing -paper_url: "https://doi.org/10.1038/s41467-024-51006-4" -code_url: "https://github.com/deepmodeling/DeePTB" -citation_urls: - - "https://doi.org/10.1038/s41467-024-51006-4" - - "https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html" - - "https://arxiv.org/abs/2609.04397" - - "https://doi.org/10.1103/PhysRevB.108.195143" - - "https://doi.org/10.1038/s41524-025-01853-6" - - "https://doi.org/10.1038/s41524-026-02137-3" -research_locale: zh -translation_url: "/research/learned-hamiltonians-to-observables/" -alternate_en_url: "/research/learned-hamiltonians-to-observables/" -alternate_zh_url: "/zh/research/learned-hamiltonians-to-observables/" -translation_of: "/research/learned-hamiltonians-to-observables/" ---- - - -
- -一个机器学习模型能够准确预测哈密顿量，并不等于我们已经能计算真正需要的物性。哈密顿量描述了电子之间的耦合，但实验上关心的往往是电导、缺陷附近的电子分布，或者把电极接在器件两端后会流过多少电流。 - -我做 DeePTB、SLEM、量子输运和大型稀疏矩阵时，反复碰到的是同一个断层：**预测算符可以很快，后面的数值计算却仍然可能贵得做不了**。这篇文章把几项已公开的工作连起来，讨论它们分别解决了什么问题，以及哪些事情不能直接推出来。它不是一篇新论文，也不包含未公开的算法或结果。 - -## 从结构到物性，中间缺了什么 - -整个过程可以分成五步： - -**原子结构与环境 → 哈密顿量与重叠矩阵 → 可扩展数值计算 → 物理观测量 → 物理验证。** - -预测矩阵是否准确，是模型的问题；百万原子能否跑得动，涉及矩阵格式、内存与多机通信；要计算光学和输运响应，则还需要边界条件和正确的物理表达式。不同环节有不同的误差来源。 - -| 环节 | 问题 | 必须验证的内容 | -| --- | --- | --- | -| 算符预测 | 能否泛化到未见过的结构？ | 能带、对称性、局域环境的外推 | -| 基底表示 | 算符作用在哪种轨道基底上？ | 重叠矩阵、轨道顺序、厄米性 | -| 数值算法 | 目标规模上能否高效计算？ | 误差、峰值内存、通信与总时间 | -| 物性提取 | 计算的是正确的物理量吗？ | DOS、电子密度、光学和透射的参照 | -| 实验对应 | 数值差别能否被测量？ | 电极、静电场、温度与不确定性 | - -## DeePTB 和 SLEM：预测的是不同层次的算符 - -[DeePTB](https://doi.org/10.1038/s41467-024-51006-4) 使用紧凑的 Slater–Koster 紧束缚表示，从第一性原理电子信息中学习随局域环境变化的参数。原子热运动和结构畸变会改变跃迁参数，因此不必给每个键固定一套值。2024 年的工作将学习到的哈密顿量交给紧束缚传播算法，对 **100 万原子 GaP** 计算有限温度的态密度和光学性质。这是百万原子的学习型哈密顿量计算，**不是百万原子自洽 DFT**。[项目介绍]({{ '/zh/research/deeptb/' | relative_url }}) · [代码](https://github.com/deepmodeling/DeePTB)。 - -[SLEM](https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html) 更直接地面向原子轨道 DFT 的矩阵，包括哈密顿量和重叠矩阵。严格局域的等变网络处理了轨道张量在空间旋转下的变换。相比紧凑的正交紧束缚模型，这里多了一个不能忽视的问题：**原子轨道基底通常不正交**。[项目介绍]({{ '/zh/research/slem/' | relative_url }})。 - -非正交基底中的本征问题是 - -$$ -Hc_n=E_n S c_n, -$$ - -其中 $S$ 是重叠矩阵。如果直接把 $H$ 送入假定 $S=I$ 的算法，求解的可能就是另一个问题。模型还必须满足相应的对称性和数值条件。即使两个哈密顿量的误差很小，某些对能级简并、带隙或费米能附近细节敏感的物性，也未必有同样小的误差。 - -## 求态密度不必先得到全部本征态 - -对有限维、厄米的 $H$ 与正定的 $S$，可以写出推迟格林函数和展宽的态密度： - -$$ -G^r(E)=\left[(E+i\eta)S-H\right]^{-1}, -$$ - -$$ -D_\eta(E)=-\frac{1}{\pi}\operatorname{Im}\operatorname{Tr}\!\left[S\,G^r(E)\right],\qquad \eta>0. -$$ - -迹里的 $S$ 是非正交度量的一部分，不应省去。不过写出逆矩阵，并不意味着要显式构造巨大的稠密逆矩阵。局域哈密顿量往往是稀疏的，矩阵向量乘法比完整对角化便宜得多；多项式传播和随机迹方法可以利用这一点。 - -但这里有一个经常被忽略的前提：**应用 $H$ 很便宜，不代表所有涉及 $S^{-1}$ 的运算都便宜**。能量展宽、随机误差、展开阶数和最终需要的谱分辨率也都需要检查。粗分辨率的总态密度，与极高精度的局域态密度或线性响应，难度并不相同。 - -[DeePTB 的 Nature Communications 论文](https://doi.org/10.1038/s41467-024-51006-4) 给出了百万原子 GaP 的 TBPM 后处理案例。而另一篇 [2026 年 npj Computational Materials 论文](https://doi.org/10.1038/s41524-026-02137-3) 将 DeePMD 热运动采样与 HSE 训练的 DeePTB 结合，研究最大 4096 原子超胞的声子辅助光学谱。两者解决的是不同尺度与精度的问题，不能混成“百万原子 HSE 自洽计算”。 - -## VBCSR：矩阵的存储方式也会决定能算多大 - -一个原子的若干轨道与另一个原子的轨道之间，对应的往往是一个小稠密块。不同元素的轨道数不同，矩阵块尺寸也可能不同。把所有数据拆成单独标量，既增加索引开销，也不利于利用块内的高效计算。 - -[VBCSR 预印本](https://arxiv.org/abs/2609.04397) 描述了一套保留可变大小原子块的稀疏矩阵工具：相同形状的块可以成组计算，矩阵通过 MPI 分布在多个进程之间。论文报告了多个稀疏矩阵 benchmark，并展示了一个超过百万原子的 InP 纳米颗粒应用。[项目介绍与实际 benchmark]({{ '/zh/research/vbcsr/' | relative_url }}) · [代码](https://github.com/DeepElectron/vbcsr)。 - -判断这种方法的价值，至少要分开看三件事：某种矩阵格式是否适合目标块结构；一个乘法 kernel 在特定硬件上是否更快；**完整的物理计算**是否真的以可接受的误差更快完成。阈值截断可以阻止中间矩阵不断变稠密，但也引入近似；通信、内存和物性误差要一起报告。 - -## 量子输运：有了哈密顿量，还得有电极和边界 - -材料的能带不能直接告诉我们器件电流。输运需要考虑接触、边界条件，以及有栅极时的静电势。相干 NEGF 中，中心区域格林函数包含左右电极的自能： - -$$ -G^r_C(E)=\left[(E+i\eta)S_C-H_C-\Sigma^r_L(E)-\Sigma^r_R(E)\right]^{-1}. -$$ - -透射与电极耦合和格林函数共同有关；非正交基底中的耦合也必须一致处理。因此，**能带误差小不等于透射误差小**。 - -[2025 年 npj Computational Materials 的 DeePTB-NEGF 工作](https://doi.org/10.1038/s41524-025-01853-6) 用学习型哈密顿量计算了超过一万种断裂结结构，并研究碳纳米管晶体管。41 nm 沟道的栅控器件与实验转移曲线进行对照；另一个 180 nm 碳管案例计算的是**零偏压透射**，并非 180 nm 的完整有限偏压栅控模拟。类似地，[二维材料研究](https://arxiv.org/abs/2512.11291) 在石墨烯、h-BN 和 MoS₂ 上与 DFT-NEGF 透射结果进行比较。[项目介绍]({{ '/zh/research/deeptb-negf/' | relative_url }})。这些结果不能直接证明任意缺陷、界面和真实工业器件都具有相同精度。 - -另一个方向是让输运计算本身可微。我更早做的 [AD-NEGF（Physical Review B 2023）](https://doi.org/10.1103/PhysRevB.108.195143) 使用自动微分和隐式层，在紧束缚 NEGF 框架里计算导数，用于灵敏度分析和逆向设计。它没有消除物理模型的近似，只是让给定模型的计算与梯度保持一致。[项目介绍]({{ '/zh/research/ad-negf/' | relative_url }})。 - -## 我会怎样检查一项“大规模高精度”计算 - -看一个模型或求解器是否可信，我通常会问： - -1. **训练范围：** 参考数据包括哪些元素、结构和电子态？测试集是真正独立的吗？ -2. **算符一致性：** $H$、$S$ 的厄米性、正定性、轨道约定是否与数值求解器匹配？ -3. **物性误差：** 不能只看训练损失；DOS 峰位、局域电荷、光学谱或透射本身，与参考结果差多少？ -4. **数值收敛：** 答案对展宽、随机采样、多项式阶数、稀疏阈值和电极自能有多敏感？ -5. **端到端开销：** 报告的是整套有用计算的时间、内存、节点数与通信，还是一个特别适合某种硬件的矩阵 kernel？ - -这套问题也适合给 Coding Agent。Agent 可以训练模型、运行计算、生成图片，但还应该留下输入、模型版本、算符约定、数值参数、参考结果以及失败的检查。否则一次自动化执行得再快，换个人也不容易复现。我做的 [flower](https://github.com/floatingCatty/flower) 就关注这个问题；仓库已经有一个[可复现实验案例](https://github.com/floatingCatty/flower/blob/main/docs/REPRODUCING_AGENT_CALCULATIONS.md)。 - -## 最后 - -我越来越觉得，科学计算真正有价值的单位不是“模型输出一个矩阵”，而是一条能把原子结构变成**可验证的物理答案**的工作流。学习型算符、稀疏数值算法和物性求解器解决的是不同瓶颈。它们确实让新的规模和科学问题变得可能，但前提是整条计算链的误差和适用范围都能交代清楚。 - -本文仅梳理已公开的论文与代码，没有涉及未发布的手稿结果。 - -### 原始资料 - -- [DeePTB — Nature Communications **15**, 6772 (2024)](https://doi.org/10.1038/s41467-024-51006-4) -- [SLEM — ICLR 2025](https://proceedings.iclr.cc/paper_files/paper/2025/hash/7db2ffcbfd0bd361d47b7fa612bd2ba2-Abstract-Conference.html) -- [VBCSR — arXiv:2609.04397](https://arxiv.org/abs/2609.04397) -- [AD-NEGF — Physical Review B **108**, 195143 (2023)](https://doi.org/10.1103/PhysRevB.108.195143) -- [DeePTB-NEGF — npj Computational Materials **11**, 375 (2025)](https://doi.org/10.1038/s41524-025-01853-6) -- [声子辅助光学谱 — npj Computational Materials **12**, 268 (2026)](https://doi.org/10.1038/s41524-026-02137-3) -- [二维材料量子输运 — arXiv:2512.11291](https://arxiv.org/abs/2512.11291) - -
From 391807583968690c63a42ea513b9752d56b54231 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:58:11 -0400 Subject: [PATCH 03/15] Remove unpublished Hamiltonians-to-observables guide from research index --- _pages/research.html | 6 ------ 1 file changed, 6 deletions(-) diff --git a/_pages/research.html b/_pages/research.html index b24e20c..5a5b193 100644 --- a/_pages/research.html +++ b/_pages/research.html @@ -41,12 +41,6 @@

Research

- - 2026 · Technical perspective -

From learned Hamiltonians to scalable electronic properties

-

Why accurate learned quantum operators are only the beginning: overlap metrics, sparse solvers, spectra, optical response and open-system transport.

- Read the technical guide ↗ - 2026 · arXiv

Moiré phonon condensation in twisted bilayer graphene

From 3245dbbe6ff953325aab42e09b8701d73b1b353e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:58:16 -0400 Subject: [PATCH 04/15] Remove unpublished Hamiltonians-to-observables guide from research index --- _pages/research-zh.html | 6 ------ 1 file changed, 6 deletions(-) diff --git a/_pages/research-zh.html b/_pages/research-zh.html index 1d161f6..16c37f8 100644 --- a/_pages/research-zh.html +++ b/_pages/research-zh.html @@ -41,12 +41,6 @@

研究

- - 2026 · 技术综述 -

从机器学习哈密顿量到真正可计算的电子物性

-

把 DeePTB、SLEM、VBCSR 和量子输运连接起来：从学习算符、非正交重叠到稀疏求解和物性验证。

- 阅读技术文章 ↗ - 2026 · arXiv

魔角石墨烯里的声子凝聚

From db8d0680617cdf2622f833055e0a9701dbaa361a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:58:27 -0400 Subject: [PATCH 05/15] Remove promotional homepage link to postponed RSATB-adjacent review --- index.html | 4 ---- 1 file changed, 4 deletions(-) diff --git a/index.html b/index.html index 4659594..3d01d26 100644 --- a/index.html +++ b/index.html @@ -393,10 +393,6 @@

It began with a small flame.

All publications → 研究Research → -

- 如果想看这些工作的共同技术脉络：For the technical thread connecting these projects: - 从学习哈密顿量到可计算物性From learned Hamiltonians to scalable electronic properties ↗ -

From 42612cfeb2235a09a6ce0ff51b845584f6f1d939 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:58:48 -0400 Subject: [PATCH 06/15] Assert postponed guide stays out of generated site, sitemap and research feeds --- scripts/audit_site.py | 56 ++++++++++++++++--------------------------- 1 file changed, 20 insertions(+), 36 deletions(-) diff --git a/scripts/audit_site.py b/scripts/audit_site.py index 68bf69e..a7a9f91 100644 --- a/scripts/audit_site.py +++ b/scripts/audit_site.py @@ -256,43 +256,27 @@ def json_file(name): papers = json_file("publications.json") notes = research.get("items", []) entries = papers.get("papers", []) - expect(len(notes) >= 11, "Fewer than eleven research/software and technical guide entries") + expect(len(notes) >= 10, "Fewer than ten released research and software notes") expect(len(entries) >= 11, "Fewer than eleven verified publications in generated feed") - # The high-level technical guide should be discoverable in both - # languages and cite all its public primary sources, not just the - # reference DeePTB paper. It is a BlogPosting, not a fictitious journal work. - guide_slug = "learned-hamiltonians-to-observables" - guide_en = "https://floatingcatty.github.io/research/" + guide_slug + "/" - guide_zh = "https://floatingcatty.github.io/zh/research/" + guide_slug + "/" - guide_record = next((x for x in notes if x.get("url") == guide_en), None) - expect(guide_record is not None, "High-level technical perspective missing from research feed") - if guide_record: - expect(guide_record.get("url_zh") == guide_zh, - "High-level guide missing Chinese URL in research feed") - for url, expected_lang in ((guide_en, "en"), (guide_zh, "zh-Hans")): - guide = parse_page(url) - if guide: - posts = [x for x in guide.schemas if isinstance(x, dict) - and x.get("@type") == "BlogPosting"] - expect(len(posts) == 1, "Guide should have one BlogPosting schema: " + url) - if posts: - source_urls = posts[0].get("citation") - expect(isinstance(source_urls, list) and len(source_urls) >= 6, - "Guide does not cite the primary papers: " + url) - expect(posts[0].get("inLanguage") == expected_lang, - "Guide has wrong structured-data language: " + url) - expect(any("s41524-025-01853-6" in (href or "") for href in guide.links), - "Guide missing published device-transport source: " + url) - expect(any("2609.04397" in (href or "") for href in guide.links), - "Guide missing VBCSR source: " + url) - for url in (guide_en, guide_zh): - index_url = "https://floatingcatty.github.io/" + ( - "research/" if url == guide_en else "zh/research/") - idx = parse_page(index_url) - if idx: - expect(any(urlsplit(href or "").path == urlsplit(url).path - for href in idx.links), - "Localized research index omits flagship guide: " + index_url) + # The author postponed the standalone cross-project synthesis until RSATB + # is public; no public page, sitemap entry, JSON feed item or index link + # may expose either localized draft in the meantime. + for slug in ("research/learned-hamiltonians-to-observables/", + "zh/research/learned-hamiltonians-to-observables/"): + withheld = "https://floatingcatty.github.io/" + slug + expect(resolve_url(withheld) is None, "Postponed guide still rendered: " + withheld) + expect(withheld.rstrip("/") not in sitemap_urls, "Postponed guide in sitemap: " + withheld) + expect(not any(slug.rstrip("/") in (x.get("url", "") or "") or + slug.rstrip("/") in (x.get("url_zh", "") or "") + for x in notes), + "Postponed guide leaked into research.json: " + withheld) + for location in ("https://floatingcatty.github.io/", + "https://floatingcatty.github.io/research/", + "https://floatingcatty.github.io/zh/research/"): + p = resolve_url(location) + if p: + expect(slug.rstrip("/") not in p.read_text(encoding="utf-8"), + "Postponed guide still linked from public navigation: " + location) conference = next((x for x in entries if "2025-12-12-DeePTB-NEGF-preprint" in x.get("url", "")), None) From 69e06f7edf82878b4c44f9a2386d9b6c98c85127 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:59:31 -0400 Subject: [PATCH 07/15] Ground CV multiscale research arc in reference PDF and expose original paper links --- _pages/cv.html | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/_pages/cv.html b/_pages/cv.html index 2c333d9..5dba044 100644 --- a/_pages/cv.html +++ b/_pages/cv.html @@ -32,7 +32,8 @@

Curriculum vitae

Zhanghao Zhouyin 周寅张皓

Physics PhD candidate · McGill University · Montréal, Canada

-

I develop AI-native scientific computing systems that connect machine-learned physical models, high-performance simulation, and engineering-scale prediction, spanning numerical infrastructure, electronic and device simulation, differentiable physics, and multiscale software.

+

I develop AI-native scientific computing systems that connect machine-learned physical models, high-performance simulation, and engineering-scale prediction. My work bridges scales—from atomistic and electronic-structure simulation through quantum transport and physical observables to device-level TCAD, coupled multiphysics, and circuit-aware modeling—using scalable numerical solvers and differentiable physics.

+

Across scales: Atomistic / electronic structure → quantum operators and observables → quantum transport → device TCAD / multiphysics → circuit models.

Some publications use the author name Yinzhanghao Zhou or YinZhangHao Zhou.

中文版本 ↗ @@ -63,7 +64,7 @@

AI-native scientific computing and research projects

Industry and research experience

-
Oct 2025–Sep 2026

Research Intern · Nanoacademic Technologies Inc. (QTCAD)

Montréal. AI-enhanced modeling for superconducting quantum devices, linking first-principles Hamiltonians, finite-element electrostatics, quantum dynamics, and circuit-level simulation in QTCAD workflows.

+
Oct 2025–Sep 2026

Research Intern · Nanoacademic Technologies Inc. (QTCAD)

Montréal. AI-enhanced multiscale modeling for Nanoacademic's QTCAD, including learned first-principles Hamiltonian surrogates for superconducting-junction and circuit quantities, and device workflows that couple finite-element electrostatics, atomistic tight binding, time-dependent quantum dynamics and circuit-level models.

May 2022–Jun 2024

Research Intern · AI for Science Institute

Beijing; machine learning for materials simulation. Research mentors: Dr. Qiangqiang Gu and Dr. Linfeng Zhang.

Oct 2020–Mar 2021

Editorial Intern · Beijing Academy of Artificial Intelligence

Tracked machine-learning research and wrote technical reviews.

@@ -85,6 +86,10 @@

Publications and preprints

  • {% if pub.citation_date %}{{ pub.citation_date }}{% else %}{{ pub.date | date: "%Y" }}{% endif %}
    {{ pub.title | escape }}
    +
    + {% if pub.paperurl %}Original paper / preprint ↗{% endif %} + {% if pub.research_url %}  ·  Research note ↗{% endif %} +
    {{ pub.venue | escape }}
    • {% endfor %} From a4755472439716ce8228ad7a2a6f38e553ea4eee Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:59:37 -0400 Subject: [PATCH 08/15] Ground CV multiscale research arc in reference PDF and expose original paper links --- _pages/cv-zh.html | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/_pages/cv-zh.html b/_pages/cv-zh.html index 3255a36..a8f2cad 100644 --- a/_pages/cv-zh.html +++ b/_pages/cv-zh.html @@ -29,7 +29,8 @@

    学术履历

    Zhanghao Zhouyin 周寅张皓

    Physics PhD candidate · McGill University · Montréal, Canada

    -

    我开发 AI-native 科学计算系统，将机器学习物理模型、高性能模拟和工程尺度预测连接起来。研究涵盖稀疏数值基础设施、电子结构和器件模拟、可微物理及多尺度工程软件。

    +

    我开发 AI-native 科学计算系统，将机器学习物理模型、高性能模拟与工程尺度预测连接起来。研究从原子和电子结构、量子输运与物性计算，延伸到器件级 TCAD、多物理场耦合以及电路模型；核心是可扩展数值求解器、可微物理计算和可验证的跨尺度模拟流程。

    +

    跨尺度计算： 原子与电子结构 → 量子算符和物性 → 量子输运 → 器件 TCAD／多物理场 → 电路模型。

    部分论文采用 Yinzhanghao Zhou 或 YinZhangHao Zhou 署名。

    English version ↗ @@ -60,7 +61,7 @@

    AI 原生科学计算与研究项目

    工业与研究经历

    -
    Oct 2025–Sep 2026

    Research Intern · Nanoacademic Technologies Inc. (QTCAD)

    蒙特利尔。面向超导量子器件的多尺度模拟，将第一性原理、有限元静电学、量子动力学与电路层建模连接起来；开展 AI 增强的器件建模。

    +
    Oct 2025–Sep 2026

    Research Intern · Nanoacademic Technologies Inc. (QTCAD)

    蒙特利尔。在 Nanoacademic QTCAD 中开展 AI 增强的多尺度量子器件建模，包括以第一性原理哈密顿量代理模型估计超导结与电路相关物理量，以及耦合有限元静电学、原子尺度紧束缚和含时量子动力学的器件流程。

    May 2022–Jun 2024

    Research Intern · AI for Science Institute

    北京；机器学习材料模拟。研究指导：Dr. Qiangqiang Gu 和 Dr. Linfeng Zhang。

    Oct 2020–Mar 2021

    Editorial Intern · Beijing Academy of Artificial Intelligence

    跟踪机器学习研究并撰写技术综述。

    @@ -82,6 +83,10 @@

    论文与预印本

  • {% if pub.citation_date %}{{ pub.citation_date }}{% else %}{{ pub.date | date: "%Y" }}{% endif %}
    {{ pub.title | escape }}
    +
    + {% if pub.paperurl %}正式论文／预印本 ↗{% endif %} + {% if pub.research_url %}  ·  研究介绍 ↗{% endif %} +
    {{ pub.venue | escape }}
    • {% endfor %} From 4ceeb46354209eed8d8dc3b4c8a213e60afbedda Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 20:59:50 -0400 Subject: [PATCH 09/15] Add clear icon-labeled academic CV link to shared navigation in both languages --- _includes/lamp-nav.html | 2 ++ 1 file changed, 2 insertions(+) diff --git a/_includes/lamp-nav.html b/_includes/lamp-nav.html index c63d82f..03a50e6 100644 --- a/_includes/lamp-nav.html +++ b/_includes/lamp-nav.html @@ -8,6 +8,8 @@ 学习笔记Study notes 研究Research 论文Publications + CV + 关于About 联系Contact From 7ee4ce9ff75515d42da8ba103906851b6b56b2de Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:00:05 -0400 Subject: [PATCH 10/15] Style prominent CV navigation pill and keep navigation usable on narrow screens --- _includes/lamp-head.html | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/_includes/lamp-head.html b/_includes/lamp-head.html index 61ddec1..3674e0f 100644 --- a/_includes/lamp-head.html +++ b/_includes/lamp-head.html @@ -73,6 +73,9 @@ .navlinks a{font-size:13.5px;color:#AFA497} .navlinks a:hover{color:#E6DDCF} .navlinks a.active{color:#E6DDCF} + .navlinks a.nav-cv{display:inline-flex;align-items:center;gap:6px;padding:5px 11px;color:#E9C784;border:1px solid rgba(224,173,88,.48);border-radius:999px;font-family:'IBM Plex Mono',monospace;font-size:12px;letter-spacing:.25px;white-space:nowrap} + .navlinks a.nav-cv:hover{border-color:#E0AD58;background:rgba(224,173,88,.1);color:#F4DFC2} + .nav-cv svg{flex:none;width:14px;height:14px} .lang-btn{cursor:pointer;font-family:'IBM Plex Mono',monospace;font-size:12px;letter-spacing:.5px;color:#E6DDCF;background:transparent;border:1px solid rgba(224,173,88,.28);border-radius:999px;padding:5px 12px;transition:border-color .3s,background .3s} .lang-btn:hover{border-color:rgba(224,173,88,.6);background:rgba(224,173,88,.08)} /* bilingual toggle — default Chinese; [data-lang=en] flips to English */ @@ -114,8 +117,14 @@ :root.js .reveal{opacity:0;transform:translateY(18px);transition:opacity 1s ease,transform 1s cubic-bezier(.2,.7,.2,1)} :root.js .reveal.is-visible{opacity:1;transform:translateY(0)} + @media (max-width:900px){ + .nav{flex-wrap:wrap;row-gap:10px} + .navlinks{flex-wrap:wrap;gap:9px 15px;justify-content:flex-end;margin-left:auto} + } @media (max-width:640px){ + .navlinks{width:100%;justify-content:flex-start;gap:8px 12px} .navlinks a{font-size:12.5px} + .navlinks a.nav-cv{padding:4px 9px;font-size:11.5px} } @media (prefers-reduced-motion: reduce){ html{scroll-behavior:auto} From f3d1fb8c7d5e77af476654cbbede20969f8fc675 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:00:17 -0400 Subject: [PATCH 11/15] Expose CV as a first-screen homepage CTA with correct bilingual URLs --- index.html | 1 + 1 file changed, 1 insertion(+) diff --git a/index.html b/index.html index 3d01d26..9bd5f83 100644 --- a/index.html +++ b/index.html @@ -285,6 +285,7 @@

    Zhanghao Zhouyin

    What I'm building Notes by the lamp → + 学术 CVAcademic CV ↗
    scroll
    From d23dd3bb80bc289019284fe06d94696515569baf Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:00:42 -0400 Subject: [PATCH 12/15] Render existing paper URLs in current lamp design while preserving scholarly citation metadata --- _layouts/publication.html | 89 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 89 insertions(+) create mode 100644 _layouts/publication.html diff --git a/_layouts/publication.html b/_layouts/publication.html new file mode 100644 index 0000000..f5a09ce --- /dev/null +++ b/_layouts/publication.html @@ -0,0 +1,89 @@ +--- +layout: raw +--- + + + + + +{{ page.title | escape }} | Zhanghao Zhouyin — Publications + +{% include lamp-head.html %} +{% include lamp-profile-style.html %} +{% comment %}Keep the Highwire/Google Scholar metadata that the older AcademicPages layout supplied.{% endcomment %} + +{% for paper_author in page.paper_authors %} + +{% endfor %} + +{% if page.journal %}{% endif %} +{% if page.conference %}{% endif %} +{% if page.volume %}{% endif %} +{% if page.doi %}{% endif %} + + + + +
    + {% include lamp-nav.html active="pubs" %} +
    +

    Publication record

    +

    {{ page.title | escape }}

    + {% if page.paper_authors %} +

    {% for paper_author in page.paper_authors %}{{ paper_author | escape }}{% unless forloop.last %} · {% endunless %}{% endfor %}

    + {% endif %} +

    {% if page.venue %}{{ page.venue | escape }}{% endif %}

    +
    + {% if page.paperurl %}Original publication / preprint ↗{% endif %} + {% if page.preprint_url and page.preprint_url != page.paperurl %}ArXiv version ↗{% endif %} + {% if page.research_url %}Research background ↗{% endif %} + All publications ↗ + Academic CV ↗ +
    +
    +
    +
    +

    Abstract and research overview

    +

    Paper record · Original source linked above

    +
    {{ content }}
    +

    This page is part of the author's maintained research bibliography. The text may be an author-written overview or an abstract from an earlier manuscript version; consult the linked publisher or preprint for the version of record.

    +
    + {% if page.citation %} +
    +

    Citation

    +

    {{ page.citation | escape }}

    + {% if page.doi %}

    DOI: {{ page.doi | escape }}

    {% endif %} +
    + {% endif %} +
    + {% include lamp-footer.html %} +
    +{% include lamp-embers.html %} + + From 36c8207b2a203f8c69e8c7d64ac2119ddfd96db9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:00:54 -0400 Subject: [PATCH 13/15] Migrate all existing publication URLs from AcademicPages single layout to the current research site --- _config.yml | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/_config.yml b/_config.yml index 4efa7cc..f22fb90 100644 --- a/_config.yml +++ b/_config.yml @@ -205,15 +205,16 @@ defaults: values: layout: single author_profile: true - # _publications + # _publications: original URLs remain stable, but use the unified + # oil-lamp layout and preserve independent Scholar/Highwire citation tags. - scope: path: "" type: publications values: - layout: single - author_profile: true - share: true - comments: true + layout: publication + author_profile: false + share: false + comments: false # Sass/SCSS sass: From 3331217748a28469fb46405010c85a27d10da54a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:01:21 -0400 Subject: [PATCH 14/15] Test visible CV navigation, cross-scale description and unified publication pages --- scripts/audit_site.py | 34 ++++++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+) diff --git a/scripts/audit_site.py b/scripts/audit_site.py index a7a9f91..02d973a 100644 --- a/scripts/audit_site.py +++ b/scripts/audit_site.py @@ -142,6 +142,15 @@ def json_file(name): "Homepage should expose the researcher identity") expect(any("/research/" in (href or "") for href in home.links), "Homepage does not link to research") + expect(any("/cv/" in (href or "") for href in home.links), + "Homepage must visibly link the academic CV") + home_path = resolve_url("https://floatingcatty.github.io/") + if home_path: + markup = home_path.read_text(encoding="utf-8") + expect('class="nav-cv"' in markup, + "Shared navigation lacks visible icon-labeled CV item") + expect('data-href-zh="/zh/cv/"' in markup, + "CV navigation does not expose the Chinese URL") # The human-facing researcher and CV pages should not expose the old 2023-era # resume or stale biography, and their canonical URL must stay unchanged. @@ -167,6 +176,20 @@ def json_file(name): "Alea iacta est.", "Zhanghao's Blog")), slug + " still contains content from the retired CV/website") + # CV deliberately links papers to original scientific sources as well + # as same-theme author-owned detail pages (not an old AcademicPages view). + for locale in ("cv/", "zh/cv/"): + cv_path = resolve_url("https://floatingcatty.github.io/" + locale) + if cv_path: + cv_html = cv_path.read_text(encoding="utf-8") + expect('10.1038/s41467-024-51006-4' in cv_html, + "CV missing direct DOI link to published DeePTB paper") + expect('/publication/' in cv_html, + "CV should preserve stable local paper detail links") + expect('TCAD' in cv_html and 'circuit' in cv_html.lower() or + 'TCAD' in cv_html and '电路' in cv_html, + "CV lost atomistic-to-device multiscale positioning") + # Full-language profile URLs must be crawlable independently. for slug in ("cv", "about"): en_url = "https://floatingcatty.github.io/" + slug + "/" @@ -336,6 +359,17 @@ def json_file(name): expect("ScholarlyArticle" in page.schema_types(), "ScholarlyArticle missing in " + url) expect(bool(page.meta.get("citation_title")), "citation_title missing in " + url) + published_page = resolve_url(url) + if published_page: + markup = published_page.read_text(encoding="utf-8") + expect('class="publication-title"' in markup and + 'class="profile-intro"' in markup, + "Paper detail still uses retired academic theme: " + url) + expect('class="page__inner-wrap"' not in markup, + "Paper detail unexpectedly retains AcademicPages HTML: " + url) + source = paper.get("paper_url") + expect(bool(source) and source in markup, + "Paper detail must link to original published paper: " + url) authors = paper.get("authors") or [] expect(bool(authors) and len(page.meta.get("citation_author", [])) == len(authors), "Citation author count mismatch in " + url) From ba835329f56c62784e59bf32cb5277d12743323e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Zhanghao=20Zhouyin=20=28=E5=91=A8=E5=AF=85=E5=BC=A0?= =?UTF-8?q?=E7=9A=93=29?= <64253517+floatingCatty@users.noreply.github.com> Date: Sun, 11 Oct 2026 21:01:39 -0400 Subject: [PATCH 15/15] Document stable paper permalinks, refreshed citation design and conservative publication policy --- README.md | 15 ++++++--------- 1 file changed, 6 insertions(+), 9 deletions(-) diff --git a/README.md b/README.md index a3de68f..b158062 100644 --- a/README.md +++ b/README.md @@ -6,7 +6,7 @@ Live at . ## Structure -The site is a Jekyll project. The homepage, research, essays, study notes and publication index use the oil-lamp design. The CV, About page and individual publication abstracts use the underlying [academicpages](https://github.com/academicpages/academicpages.github.io) / Minimal Mistakes theme. +The site is a Jekyll project. The homepage, research, essays, study notes and publication index use the oil-lamp design. The CV, About, research notes, publication index and individual paper records now share the site's oil-lamp visual design; old paper permalinks are retained, with Highwire and ScholarlyArticle metadata. | Path | Purpose | | --- | --- | @@ -19,7 +19,8 @@ The site is a Jekyll project. The homepage, research, essays, study notes and pu | `_pages/research-*.md` | Bilingual research and software notes (`layout: research`) | | `publications.json`, `research.json` | Jekyll-generated public machine-readable indexes | | `scripts/audit_site.py` | Checks generated pages, research links, citation metadata, profile canonicals and legacy redirects | -| `_includes/lamp-profile-style.html` | Shared accessible/printable About and CV presentation | +| `_includes/lamp-profile-style.html` | Shared accessible/printable profile and citation presentation | +| `_layouts/publication.html` | Current-styled paper records at existing permalinks, preserving citation metadata and direct source links | | `_includes/lamp-*.html` | Shared markup for the lamp-styled pages (head, nav, footer, embers) | | `_layouts/raw.html` | Pass-through layout for the lamp landing/listing pages | | `_layouts/post.html`, `_layouts/research.html` | Lamp-styled article layouts, incl. MathJax | @@ -84,13 +85,9 @@ each, linking directly to the corresponding translation. The site homepage uses locale-aware links where a real Chinese equivalent exists. A research note is a personal explanation of a project, not the original -paper abstract. A broader **technical perspective** may use -`project_kind: guide` with a first-party source URL and -`citation_urls:` listing several fully verified primary references. -This renders `BlogPosting.citation` as a list of URLs and labels the note -as a technical perspective, not a new research paper. Do not describe -unpublished numerical results, or add SEO-only translations that do not -contain a substantial independent-language body. +paper abstract. A research synthesis should only be published when its sources and intended +scope are ready for public discussion. Never disclose unreleased numerical +results or create SEO-only translations with no substantive body. An academic paper belongs in `_publications/`. Its front matter should include the **published** author list, journal or conference, date/year, DOI