一个特征量管五种楼形:尾流宽度能否统一预测涡脱频率的裁决

Written by

in

Language

Evidence level: B

上一轮证明了”St 数可以被回归”,这一轮要问更难的一问:这个回归的自变量,到底该用盖楼时量得到的尺寸,还是用只有风洞里才看得见的尾流宽度?

正文

第一层:这个想法不是我们提的,是 Roshko 1948 年提的。 涡脱频率的常规写法是 f = St·U/D,D 是迎风宽度。但 Roshko 发现,只要把 D 换成”尾流里两条自由流线的间距”d′、把 U 换成分离点处的流速 Uₛ,得到的 St* = f·d′/Uₛ 在所有测过的柱体上都落在 0.16 附近——平板、楔形、圆柱都不例外(Roshko, On the Wake and Drag of Bluff Bodies, NACA TN 3168, 1948)。这就是”统一特征量”的最初形态:不是形状决定频率,是尾流的宽度决定频率;形状只是通过改变尾流宽度间接起作用。

第二层:这条线索后来被两次加固。 Ponta & Aref 2004 年从涡街几何的稳定性出发,把 St-Re 关系从经验拟合推成了有理论骨架的式子(PMID 15447193);Roushan & Wu 2005 年给出它的结构化解释——关系的形式来自涡形成机制本身,与具体形状无关(PMID 15783649)。到 Kim & Wu 2015 年的肥皂膜实验,事情被推到最干净的一步:多种形状的障碍物生成层流涡街,Re→∞ 时 St(∞)=1/A≈0.21 与形状完全无关,形状差异全部被压缩进单个参数 B(PMID 26565336)。形状没有消失,它只是退到了一维。

第三层:但有一个反证必须摆在这里。 2025 年 PNAS 那项跨 28 个物种的巡航飞行研究,结论对”St 是普适常数”是冷水:St 与飞行条件强相关,它的普适性只在”偏好条件”内成立;真正跨物种一致的是另一个量——广义涡形成时间,用涡量注入时长比上涡量增长率来定义,与飞行条件无关(PMID 40880533)。这说明 St 本身可能是个坏因变量:它是”几何 × 工况”的混合产物,拿它当预测目标,等于逼模型去学一个本来就不该统一的量。我们真正该预测的,也许是”无量纲涡形成长度”,而 St 只是它的一个投影。

第四层:再附长度这个候选量,有直接的干预证据。 2026 年 Sci Rep 的 cross-slit 圆柱参数扫描做了件少见的事:它不改工况、只改几何(缝宽比 0.10-0.25、扫掠角),然后同时报三个数——再附长度下降 40-52.5%、阻力下降 26-32.5%、St 数在 Re=100 时下降 93.1%(PMID 42581059)。三个量同向大幅移动,说明再附长度与 St 之间不是松散相关,而是强耦合。但它同时暴露了一个麻烦:改动一个几何参数,所有候选特征量一起动,你无法从这类数据里判断谁是因、谁是果。 这正是本课题必须用”留一截面族”而不是”留一样本”来裁决的原因。

第五层:建模这一侧,两条路都有前例,而且指向相反。 一条是黑箱:SuperWing 数据集用 4,239 个参数化机翼、28,856 个 RANS 流场训练 Transformer,在留出样本上阻力误差 2.5 个阻力单位,并且对 DLR-F6、NASA CRM 这类完全没见过的复杂机翼表现出零样本外推能力(PMID 42538339)——它证明”参数化几何 + 足够大的数据 + 现代网络”确实能跨外形外推,但代价是四千级样本。另一条是可解释:微塑料碎片沉降速度那项研究里,作者从受力分析推了一个新的不规则形状因子 ISF,结果比 Corey 形状因子准,也比机器学习方法准(PMID 39419206)。一个推出来的无量纲量,赢了一个训出来的网络。 这两条路线的对照,决定了本课题的建模顺序:先找特征量,再用黑箱做上界参照,而不是反过来。

第六层:所以裁决设计要卡住一个零假设。 五类截面(方/圆/锥/切角/扭转)× 关键形态参数做统一 CFD 与风洞参数扫描,每个算例同时输出 St、尾流有效宽度 d′、分离角、再附长度 Lr/D、基压系数;然后并列三个模型:A 只用盖楼量得到的原始几何参数(宽深比、锥率、切角比、扭转角);B 只用统一特征量(d′/D、Lr/D 等);C 黑箱网络。核心零假设是:B 相对 A 没有增量。 检验方式是留一截面族交叉验证——用四类训练、预测剩下那一类,看相对误差是否守在 5-10% 以内。若 B 打不过 A,”统一特征量”就只是个换了名字的形状哑变量,统一是假的;若 B 显著优于 A 且能跨族外推,才算找到了那个比”截面类型”更深的变量。

第七层:三种结局,对应三种形煞观。 ① B 胜且外推误差 <5-10%:存在一个可量、可算、可跨形使用的几何量,形煞评级可以脱离逐案风洞,进入”量三个尺寸算一个数”的阶段——但要注意,这时预测的是频率不是吉凶,中间还隔着”频率是否落在器官共振带”这一层。② B 只在部分截面族成立(扭转、切角最可能掉队):形煞的细分流派可能确实踩到了物理边界,但边界在”依赖方式”上而不在”吉凶等级”上。③ B 全面失败:形煞必须逐案验证,”一形一断”获得物理背书,而”峦头一诀、百形皆应”只能当经验口诀、不能当公式。

回到易学。 形峦派内部其实一直有两句话在打架:一句是”峦头一诀,百形皆应”,讲的是万形同理;一句是”一形一断,不可类推”,讲的是就形论形。风水师靠肉眼在两句之间来回,靠的是手感。这篇研究做的就是把这句手感翻译成一个可判定的问题:到底存不存在一个”万形同理”的量? 存在,那两句里前一句对;不存在,后一句对。古人没能力做这件事,是因为他们看不见尾流;今天我们看得见,就不该再靠手感决定用哪一句。

证据与局限

已验证: 常规 St 数可用尾流宽度与分离点流速归一为通用 St*≈0.16,且在平板/楔/圆柱等尖缘或圆形截面上成立(Roshko 1948, NACA TN 3168);St-Re 统一关系 St=1/(A+B/Re) 对多种形状障碍物成立,Re→∞ 极限 St≈0.21 与形状无关、形状差异集中在参数 B(PMID 26565336,PMID 15783649,PMID 15447193);数值方法的边界处理与阻塞比会系统性改变所测 St,必须在协议层校正(PMID 14682839);St 的跨物种普适性受工况限制,广义涡形成时间才是跨条件不变量(PMID 40880533);几何微改(缝宽比/扫掠角)可同时大幅改变再附长度、阻力与 St(PMID 42581059);参数化几何 + 大规模数据可实现对未见外形的零样本气动外推(PMID 42538339);从受力分析推导的无量纲形状因子在预测不规则体沉降速度上优于机器学习(PMID 39419206)。

部分验证: 各候选特征量(尾流有效宽度、分离角、再附长度、基压系数)与 St 的耦合关系在单个截面族内有充分证据,但”哪一个可跨五类截面统一使用”从未被同一批算例、同一套协议横向比较过;”用尾流宽度而非迎风宽度归一后 St 变常数”这一结论在高 Re、三维、强湍流来流条件下的适用性仍有争议。

待验证: ① 是否存在一个无量纲几何特征量,可跨方/圆/锥/切角/扭转五类截面把 St 的外推误差压到 5-10% 以内;② 若 St 本身不是好的因变量,改用广义涡形成长度后跨截面统一性是否提高;③ 跨截面泛化失败时,形煞分类是否对应 St 依赖方式的物理分界;④ 用盖楼可量的三个尺寸替代风洞特征量,精度损失有多大(工程可用性边界)。

相关阅读

证据等级: B
参考文献: PMID 26565336, PMID 15783649, PMID 15447193, PMID 14682839, PMID 40880533, PMID 42581059, PMID 42538339, PMID 39419206


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

Five Feature Quantities for Predicting the Strouhal Number of Vortex Shedding from Wake Width

Evidence level: B

The previous round demonstrated that “St can be regressed.” This round asks the harder question: for this regression, should the independent variable be the dimensions measured during building design, or the wake width visible only in wind tunnels?

Main Text

Layer 1: This idea was not proposed by us; it originated with Roshko in 1948. The conventional formulation of the Strouhal frequency is $f = St \cdot U/D$, where $D$ is the windward width. However, Roshko found that replacing $D$ with “the spacing between two free streamlines in the wake,” denoted as $d’$, and replacing $U$ with the velocity at the separation point $U_s$, yields a modified Strouhal number $St^ = f \cdot d’/U_s$ that clusters near 0.16 across all tested bluff bodies—flat plates, wedges, and cylinders are no exception (Roshko, On the Wake and Drag of Bluff Bodies, NACA TN 3168, 1948). This represents the initial form of a “unified feature quantity”: not shape determines frequency, but wake width determines frequency; shape acts indirectly by altering wake width.*

Layer 2: This line of inquiry was subsequently reinforced twice. Ponta & Aref (2004) derived a theoretically grounded equation for the $St-Re$ relationship starting from the geometric stability of vortex streets (PMID 15447193); Roushan & Wu (2005) provided a structural explanation— the form of the relationship stems from the vortex formation mechanism itself and is independent of specific shapes (PMID 15783649). By Kim & Wu’s 2015 soap film experiments, the matter was pushed to its cleanest step: obstacles of various shapes generated laminar vortex streets, and as $Re \to \infty$, $St(\infty) = 1/A \approx 0.21$ was independent of shape, with all shape differences compressed into a single parameter $B$ (PMID 26565336). Shape did not disappear; it merely receded to one dimension.

Layer 3: However, a counter-evidence must be presented here. The 2025 PNAS study on cruising flight across 28 species concluded that “St is a universal constant” is a cold water: $St$ is strongly correlated with flight conditions, and its universality holds only within “preferred conditions”; what remains consistent across species is another quantity—the generalized vortex formation time, defined as the ratio of vorticity injection duration to vorticity growth rate, which is independent of flight conditions (PMID 40880533). This suggests St itself may be a poor dependent variable: it is a mixed product of “geometry × operating conditions,” and using it as a prediction target forces the model to learn a quantity that should not be unified. What we should truly predict is perhaps the “dimensionless vortex growth length,” with $St$ merely being one projection of it.

Layer 4: Regarding this candidate quantity, there is direct intervention evidence. The 2026 Sci Rep cross-slit cylinder parameter scan did something rare: it kept operating conditions constant while only altering geometry (slit width ratio 0.10-0.25, sweep angle), then reported three metrics simultaneously—reattachment length decreased by 40-52.5%, drag decreased by 26-32.5%, and $St$ decreased by 93.1% at $Re=100$ (PMID 42581059). The three quantities moved significantly in the same direction, indicating a strong coupling between reattachment length and $St$, rather than loose correlation. However, it simultaneously exposed a complication: altering one geometric parameter causes all candidate feature quantities to move together, making it impossible to determine cause versus effect from such data. This is precisely why this study must use a “leave-one-section-family” approach rather than “leave-one-sample” to adjudicate.

Layer 5: On the modeling side, both paths have precedents pointing in opposite directions. One is a black box: the SuperWing dataset trained a Transformer on 4,239 parameterized wings and 28,856 RANS flow fields; on held-out samples, it achieved a drag error of 2.5 drag units and demonstrated zero-shot extrapolation capability for complex wings unseen in DLR-F6, NASA CRM, etc. (PMID 42538339)—proving that “parameterized geometry + sufficient data + modern networks” can indeed extrapolate across forms, at the cost of thousands of samples. The other is interpretable: in a study on microplastic fragment settling velocity, authors derived a new irregular shape factor ISF from force analysis, which proved more accurate than the Corey shape factor and also more accurate than machine learning methods (PMID 39419206). A derived dimensionless quantity wins over a trained network. This comparison determines the modeling sequence for this project: first identify feature quantities, then use black boxes as upper bound references, rather than the reverse.

Layer 6: Therefore, the adjudication design must hinge on a null hypothesis. Five section types (square/circular/conical/truncated/torsioned) × key morphological parameters undergo unified CFD and wind tunnel parameter scans, with each case simultaneously outputting $St$, effective wake width $d’$, separation angle, reattachment length $Lr/D$, and base pressure coefficient. Then three models are presented in parallel: A uses only raw geometric parameters measured during building design (width-depth ratio, taper ratio, truncation ratio, torsion angle); B uses only unified feature quantities ($d’/D$, $Lr/D$, etc.); C is a black box network. The core null hypothesis is: B provides no incremental gain relative to A. The testing method is leave-one-section-family cross-validation—train on four families, predict the remaining one, and check if relative error stays within 5-10%. If B fails against A, “unified feature quantities” are merely shape dummy variables with renamed labels, and unification is false; if B significantly outperforms A and can extrapolate across families, then a variable deeper than “section type” has been found.

Layer 7: Three outcomes correspond to three views of Form Sha (shape煞). ① B wins with extrapolation error <5-10%: a quantifiable, calculable, cross-shape geometric quantity exists, allowing Form Sha ratings to detach from case-specific wind tunnels and enter a stage of “calculating one number from three dimensions”—but note, at this stage, the prediction is frequency, not auspiciousness, with an intermediate layer of “whether frequency falls within organ resonance bands.” ② B holds only for partial section families (torsion and truncation most likely to fail): the sub-schools of Form Sha may indeed have hit physical boundaries, but the boundary lies in “dependency mode” rather than “auspiciousness level.” ③ B fails comprehensively: Form Sha must be verified case-by-case, “one shape one judgment” gains physical endorsement, while “one principle for all shapes” can only serve as an empirical mnemonic, not a formula.

Returning to Yijing Studies. Within the Form Sha school, two sentences have always been at odds: one is “one principle for all shapes,” referring to universal similarity; the other is “one shape one judgment, no analogy,” referring to judging based strictly on form. Feng Shui practitioners oscillate between these two via eye contact, relying on tactile intuition. This research translates that intuition into a determinable question: Does a “universal similarity” quantity actually exist? If it exists, the first sentence is correct; if not, the second is correct. Ancient scholars lacked the capability for this task because they could not see the wake; today we can see it, so we should not rely on intuition to decide which sentence to use.

Evidence and Limitations

Verified: Conventional $St$ numbers can be normalized into a universal $St^* \approx 0.16$ using wake width and separation point velocity, holding true for sharp-edged or circular cross-sections like flat plates/wedges/cylinders (Roshko 1948, NACA TN 3168); the unified $St-Re$ relationship $St=1/(A+B/Re)$ holds for various shape obstacles, where the $Re \to \infty$ limit $St \approx 0.21$ is shape-independent and shape differences concentrate in parameter $B$ (PMID 26565336, PMID 15783649, PMID 15447193); boundary handling and blockage ratios in numerical methods systematically alter measured $St$, requiring correction at the protocol level (PMID 14682839); cross-species universality of $St$ is constrained by operating conditions, with generalized vortex formation time being the invariant across conditions (PMID 40880533); geometric micro-modifications (slit width ratio/sweep angle) can simultaneously and significantly alter reattachment length, drag, and $St$ (PMID 42581059); parameterized geometry + large-scale data enables zero-shot aerodynamic extrapolation for unseen forms (PMID 42538339); derived dimensionless shape factors outperform machine learning in predicting settling velocities of irregular bodies (PMID 39419206).

Partially Verified: Coupling relationships between candidate feature quantities (effective wake width, separation angle, reattachment length, base pressure coefficient) and $St$ have sufficient evidence within single section families, but “which one can be uniformly applied across five section types” has never been compared laterally using the same set of cases and protocols; the applicability of the conclusion “normalizing $St$ with wake width rather than windward width yields a constant” remains debated under high $Re$, three-dimensional, and strong turbulence inflow conditions.

To Be Verified: ① Does a dimensionless geometric feature quantity exist that can extrapolate $St$ error across square/circular/conical/truncated/torsioned five section types to within 5-10%? ② If $St$ itself is not a good dependent variable, does switching to generalized vortex growth length improve cross-section unification? ③ When cross-section generalization fails, does Form Sha classification correspond to physical boundaries in $St$ dependency modes? ④ What is the precision loss when replacing wind tunnel feature quantities with three dimensions measurable during building design (engineering usability boundary).

Related Reading

Evidence Level: B
References: PMID 26565336, PMID 15783649, PMID 15447193, PMID 14682839, PMID 40880533, PMID 42581059, PMID 42538339, PMID 39419206


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *