Flower Structures & Dynamics

• Flowers are structured plant forms characterized by complex geometries such as phyllotaxis and invariant spiral patterns, ensuring robust growth and seed dispersal.
• Experimental studies reveal that off-center raindrop impacts optimize seed ejection via a sling effect, achieving dispersal distances over 1 m under lab conditions.
• Research on volatile transport highlights inter-organ fumigation mechanisms that regulate defense and reproductive success through targeted biosynthesis.

Searching arXiv for the provided flower-related papers to ground the article in current records. Flowers, in the research literature considered here, are structured plant forms whose observable behavior is governed by coupled geometry, topology, fluid dynamics, volatile transport, and material heterogeneity. They appear as cup-shaped fruit bodies that opportunistically catch raindrops to eject $0.3$-mm seeds over distances of more than $1$ m, as composite disks whose florets, scales, or leaves follow phyllotactic organization on a generative spiral, as enclosed buds that support inter-organ aerial transport of sesquiterpene volatile organic compounds, and as multi-material objects whose deformation can be inferred from video by continuum-mechanical estimation (Amador et al., 2011, Rivier et al., 2016, Boachon et al., 2024, Wada et al., 18 Dec 2025).

1. Phyllotactic organization and topological structure

Phyllotaxis describes the arrangement of florets, scales or leaves in composite flowers or plants such as daisy, aster, sunflower, pinecone, and pineapple. In the phyllotactic model, floret centers are indexed by integers $n=0,1,2,\ldots$ and placed on a planar spiral with polar coordinates

$r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$

where $a$ sets the overall scale and $\Delta\theta$ is the constant divergence angle. The classical choice is

$$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$

which yields the densest, most homogeneous disk-covering by equal-area cells. The structure is then obtained by the Voronoi tiling of the spiral lattice, so that each floret corresponds to a Voronoi cell and the flower head is represented as a geometrical foam (Rivier et al., 2016).

Within this representation, typical cells are hexagons, while pentagons and heptagons act as topological defects. The number of sides of a cell equals the number of first-neighbor points in the Delaunay graph. Although consecutive indices on the generative spiral are not Voronoi-neighbors, each hexagonal cell is crossed by exactly three visible spiral families of neighbors, the parastichies. These are labeled by successive Fibonacci numbers. Numerically, if $f_k$ is the $k$-th Fibonacci number, neighbors of site $n$ occur approximately at $1$0, $1$1, and $1$2, equivalently $1$3 whenever $1$4. Thus a large phyllotactic disk exhibits triple spiral counts such as $1$5; daisies and asters typically show $1$6 and $1$7, sunflowers $1$8 and $1$9, and pineapples $n=0,1,2,\ldots$0 and $n=0,1,2,\ldots$1.

The topological evolution of this structure is described by the elementary foam moves $n=0,1,2,\ldots$2 and $n=0,1,2,\ldots$3. A $n=0,1,2,\ldots$4 event is an edge-flip at a nearly square $n=0,1,2,\ldots$5-valent vertex, exchanging neighbors and providing the elementary plastic response to shear. A $n=0,1,2,\ldots$6 event is the disappearance or creation of a cell. The remarkable result is that the one-dimensional coding of cell-side counts and the parastichy organization are invariant under disappearance of the first Voronoi cell or, conversely, under creation of a first cell. Removing the central pentagon leaves a new first cell that is again a pentagon, while the same sequence of parastichy labels and the same concentric rings of defects persist. Grain boundaries are thin circular layers of alternating $n=0,1,2,\ldots$7-, $n=0,1,2,\ldots$8-, and $n=0,1,2,\ldots$9-sided cells, with lengths given by successive Fibonacci numbers, and each boundary sits at a topology ready for $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$0 flips. This gives composite flower heads a built-in malleability under growth and shear without destroying the global spiral counts.

2. Raindrop capture and seed ejection

Several species have raindrop-sized flowers that catch raindrops opportunistically in order to spread their $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$1-mm seeds distances of over $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$2 m. The fruit bodies are shallow, bowl-like cups between $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$3 and $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$4 mm in diameter, comparable to typical raindrop diameters in the $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$5–$r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$6 mm range. Each small bowl carries dozens of seeds. Cross-sections reveal a roughly spherical-cap profile with rim angles tuned to guide splash; seeds sit near the bottom of the well, just above the focus of the incoming drop; and the small aperture with gentle curvature discourages radial leakage and promotes upward, directed ejection. To isolate geometry and impact dynamics, researchers scanned wild fruit bodies with micro-computed tomography, converted them to CAD models, and $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$7D-printed rigid resin replicas. Single water drops of controlled diameter $r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$8 ($r_n=a\sqrt{n}, \qquad \theta_n=n\Delta\theta,$9–$a$0 mm) and impact speed $a$1 ($a$2–$a$3 m/s) were then generated with a drop-on-demand system, while the target mimic was translated laterally to vary the nondimensional impact parameter

$a$4

where $a$5 is the horizontal offset between drop and cup axes and $a$6. High-speed cameras operating at up to $a$7 fps recorded splash, seed entrainment, ejection angles, and landing distances (Amador et al., 2011).

The governing mechanics are expressed in terms of the incident drop’s momentum and kinetic energy,

$a$8

These must be redistributed into the splash sheet and the entrained seeds. When $a$9, symmetry directs the splash upward but with a limited horizontal component. For off-center impacts, especially $\Delta\theta$0–$\Delta\theta$1, the rim of the drop skirts down one side of the cup and generates a locally higher tangential velocity. Empirically,

$\Delta\theta$2

with $\Delta\theta$3 peaking near $\Delta\theta$4. Decomposing $\Delta\theta$5 into vertical and horizontal components gives

$\Delta\theta$6

and the projectile-range estimate

$\Delta\theta$7

predicts a theoretical optimum at $\Delta\theta$8 when air resistance is neglected.

High-speed analysis shows that seeds, often encapsulated briefly within a tiny secondary droplet, leave the cup at up to $\Delta\theta$9 m/s, with typical values around $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$0 m/s. Off-center impacts produce launch angles in the $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$1–$$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$2 band, whereas central impacts cluster below $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$3. Measured seed flight distances exceed $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$4 m for $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$5–$$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$6, peaking around $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$7–$$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$8 m under laboratory gravity and still surpassing $$\Delta\theta=\frac{2\pi}{\phi}, \qquad \phi=\frac{1+\sqrt{5}}{2}\simeq 1.618\ldots,$$9 m in most trials; a nearly $f_k$0 reduction in dispersal occurs if $f_k$1 dips below $f_k$2 or exceeds $f_k$3. The reason off-center impacts maximize dispersal is threefold: speed amplification through a sling effect, seed encapsulation within a coherent globule, and launch angles near the theoretical optimum. From an evolutionary standpoint, natural selection appears to have tuned flower size and curvature to exploit common raindrop sizes and impact offsets in the habitat. In engineering terms, the same mechanism motivates microfluidic splash actuators or droplet ejection devices with no moving parts, including targeted drug delivery, inkjet printing, and environmental sensing.

3. Intrafloral volatile transport by natural fumigation

A distinct line of work identifies a new physiological phenomenon in petunia flowers: inter-organ aerial transport of volatile organic compounds via natural fumigation. In closed, pre-anthesis Petunia hybrida buds, three sesquiterpene synthases, PhTPS1, 3, and 4, and one monoterpene synthase, PhTPS2, are expressed, but only PhTPS1 is strictly tube-localized in the top $f_k$4 cm of the floral tube, just beneath the unexpanded corolla. PhTPS1 converts the universal $f_k$5 precursor farnesyl diphosphate (FPP) into a mixture of four sesquiterpenes: germacrene D, bicyclogermacrene, $f_k$6-cadinene, and germacrene D-4-ol. In vitro and yeast expression assays show classical Michaelis–Menten kinetics with FPP, summarized by

$f_k$7

Crude protein assays from dissected tubes, but not from pistils, regenerate the same four products when supplied with $f_k$8 $f_k$9-FPP, confirming that PhTPS1 alone is sufficient for this bouquet (Boachon et al., 2024).

The physical mechanism depends on the enclosed bud headspace. Closed buds create a small chamber above the tube, and PhTPS1-generated sesquiterpenes are emitted predominantly from the inner, adaxial surface of the top tube, accounting for $k$0–$k$1 of total emission. Once in headspace, these VOCs diffuse and partition onto neighboring surfaces. Local diffusive flux in air is described by Fick’s first law,

$k$2

with $k$3 for a sesquiterpene in air. In a well-mixed enclosed bud of volume $k$4, the average headspace concentration rises until adsorption and desorption equilibria are reached. Surface adsorption onto the stigmatic cuticle follows

$k$5

and the stigma contains approximately $k$6–$k$7 more cuticular wax than the tube, ensuring preferential trapping of tube-generated sesquiterpenes on the stigma surface.

Biochemical and genetic evidence supports inter-organ aerial transport. Removing the tube from flower buds four days before anthesis eliminates accumulation of germacrene D, bicyclogermacrene, and germacrene D-4-ol in pistils on day $k$8 post-anthesis, while geraniol and $k$9-cadinene remain. Feeding only excised tubes of closed buds with $n$0-mevalonolactone results in labeled sesquiterpenes in detached pistils after $n$1 h, directly demonstrating gas-phase transfer. PhTPS1-RNAi lines with $n$2–$n$3 reduction of PhTPS1 mRNA emit approximately $n$4 less of the four tube products and show severely diminished pistil accumulation. Gas-phase complementation is decisive: placing PhTPS1-RNAi pistils inside wild-type tubes for $n$5 h fully restores internal pools of the missing sesquiterpenes, whereas wild-type pistils placed in mutant tubes lose those compounds.

The functional consequences extend beyond transport. Wild-type stigmas harbor very low bacterial loads, approximately $n$6 OTUs on day $n$7 and even less on day $n$8 post-anthesis, whereas PhTPS1-RNAi stigmas show a significant rise in a Pseudomonas OTU. PhTPS1-RNAi pistils on day $n$9 post-anthesis weigh only $1$00–$1$01 of wild type and have smaller stigmas with thinner, shorter styles; these defects are rescued by gas complementation and reciprocally induced in wild-type pistils placed in mutant tubes. Despite normal pollen germination and tube growth, PhTPS1-RNAi flowers produce up to $1$02 fewer seeds per flower, with seed weight unaltered. This establishes pre-anthesis fumigation as a mechanism for stigma defense, pistil maturation, and seed yield, and raises explicit open questions about conservation across angiosperms, the biochemical details of VOC perception, and whether bud morphology or cuticle composition modulates fumigation efficiency.

4. Flowers as multi-material mechanical systems

Flowers also serve as a representative common object for estimating physical material parameters from video. The Multi-material Physical Gaussians framework, M-PhyGs, addresses the fact that real-world objects are often complex in material composition and geometry and therefore lie outside assumptions of homogeneous single-material objects, pre-learned dynamics, or simplistic topologies. From a short video captured in a natural setting, M-PhyGs jointly segments the object into similar materials and recovers continuum mechanical parameters while accounting for gravity. The representation is hybrid. Appearance and geometry are encoded by a set of $1$03D Gaussians

$1$04

recovered by $1$05D Gaussian-splatting in the rest pose. Dynamics are encoded by a volumetric grid of material particles

$1$06

uniformly filling the shape. Each particle carries density $1$07, Young’s modulus $1$08, and Poisson’s ratio $1$09, and is linked to nearby Gaussians. Joint material segmentation uses DINO and GARField features, $1$10 and $1$11, first to over-segment in feature space and then to optimize material grouping into homogeneous-material regions (Wada et al., 18 Dec 2025).

The continuum-mechanical model assigns density $1$12 and Young’s modulus $1$13 per segment, while holding Poisson’s ratio constant, for example $1$14. The constitutive law is corotated linear elasticity in MLS-MPM. For particle deformation gradient $1$15 and rotation $1$16, the strain energy density is

$1$17

with Lamé parameters

$1$18

Stress is updated by standard particle-to-grid transfers, and gravity and contact are treated as external forces. Gravity-aware MPM also uses initial internal forces $1$19 rotated by $1$20, allowing the framework to account for sag before any external interaction.

To avoid early divergence from fully non-rigid simulation, M-PhyGs uses cascaded $1$21D and $1$22D objectives. The $1$23D tracking loss is

$1$24

where $1$25 are tracks from Dynamic $1$26D Gaussians. The image-plane refinement combines $1$27, $1$28, a DINO feature loss $1$29, and a silhouette-boundary distance-transform loss

$1$30

combined as

$1$31

with $1$32, $1$33, and $1$34. Material grouping is regularized by

$1$35

with affinity weights $1$36, plus $1$37 and $1$38 for $1$39. Efficiency and stability are improved by temporal mini-batching over overlapping subsequences of $1$40 frames. The Phlowers dataset contains $1$41 real flowers such as rose, carnation, daisy, and lily, each filmed at at least $1$42 frames with $1$43 synchronized cameras calibrated by COLMAP and ChArUco alignment, while a person inserts or arranges the flower into a flower frog. Because direct measurement of $1$44 and $1$45 on delicate petals is infeasible, evaluation uses future-frame prediction on held-out frames $1$46–$1$47 with PSNR, $1$48D IoU, and $1$49D Chamfer distance against SAM2 pseudo-masks.

5. Geometric and symbolic classification of flower images

Flower image classification has also been treated as a geometric recognition problem. Kumar et al. proposed a triangle-based method in which flowers are first segmented by whorl-based region-merging segmentation. A bank of Gabor filters $1$50 is applied to the grayscale image $1$51, producing a whorl-response map

$1$52

Quick-shift over-segmentation partitions the image into regions $1$53, and maximal region merging propagates foreground labels from high-response whorl seeds while using image-border super-pixels as background seeds. A binary flower mask is then thinned to a raw skeleton and pruned by Discrete Curve Evolution, which recursively removes vertices causing minimal change in a context-sensitive relevance measure and retains only skeleton branches terminating at convex DCE vertices. On the pruned skeleton, each foreground pixel $1$54 is classified by its $1$55-connected neighborhood degree: degree $1$56 identifies endpoints, degree at least $1$57 identifies junction points, and degree $1$58 lies on simple strokes (Kumar et al., 2016).

Let

$1$59

be the endpoint-and-junction set. The Delaunay triangulation $1$60 is defined by the empty-circumcircle property: triangle $1$61 belongs to $1$62 if and only if there is a circle through those points containing no other point of $1$63 in its interior. Equivalently, $1$64 can be computed from the Voronoi diagram, and the Quickhull-based implementation runs in $1$65 time. Each triangle contributes side lengths

$1$66

together with internal angles computed by the law of cosines. Thus each triangle produces the $1$67-dimensional feature vector

$1$68

A flower sample with $1$69 triangles, where $1$70 is the number of points on the convex hull, is then compressed into six intervals: $1$71 This interval-valued representation preserves intra-class variability while reducing storage from $1$72 scalars to six intervals.

Classification uses a voting-based symbolic classifier. For each class $1$73, a test flower’s crisp triangle features are checked against the stored class intervals, and each successful membership test contributes to an acceptance count $1$74. The final label is

$1$75

On a dataset of $1$76 flower classes with $1$77 samples, using $1$78 random trials for each split, the reported average classification accuracies are $1$79 for $1$80 train/test, $1$81 for $1$82, and $1$83 for $1$84. The method consistently outperforms earlier skeleton-based baselines cited in the paper because it preserves skeletal topology via endpoints and junctions, captures spatial configuration through Delaunay triangles, and models intra-class spread through interval features.

6. Conceptual synthesis, corrected intuitions, and open directions

Across these studies, several simpler intuitions are explicitly revised. In splash-mediated dispersal, central impacts are not the most effective; off-center impacts maximize ejection speed, encapsulation, and launch angle (Amador et al., 2011). In volatile biology, accumulation and emission of VOCs do not necessarily occur only from the tissue of biosynthesis; natural fumigation demonstrates inter-organ aerial transport from the floral tube to the stigma in closed buds (Boachon et al., 2024). In phyllotaxis, growth or removal of the first floret does not destroy the global spiral count; the parastichy organization remains invariant under $1$85 events and is mechanically buffered by Fibonacci grain boundaries (Rivier et al., 2016). In mechanics, flowers cannot always be modeled as homogeneous objects; M-PhyGs is built on the contrary assumption that flowers are materially heterogeneous and that petals, leaves, and stems should not be collapsed into a single averaged material (Wada et al., 18 Dec 2025).

Taken together, these results suggest a unifying interpretation of flowers as systems in which local geometry converts generic physical processes into structured outcomes. A cup only a few millimetres across can reshape a raindrop into a seed-carrying jet; a divergence angle of $1$86 can organize dense floret packing while preserving robustness under $1$87 and $1$88; an enclosed bud headspace can transform tube-localized sesquiterpene biosynthesis into stigma-localized defense and developmental control; and a short video can be used to infer part-wise mechanical parameters when appearance, segmentation, and continuum dynamics are jointly modeled. The open questions stated in the literature remain correspondingly broad: whether natural fumigation is conserved in flowering plants, what receptors and second messengers mediate VOC perception on the pistil surface, whether bud morphology or cuticle composition modulates fumigation efficiency, and how far part-wise physical parameter estimation from video can generalize beyond the current flower datasets.