Adaptive Node Positioning in Biological Transport Networks

Albert Alonso, Lars Erik J. Skjegstad, Julius B. Kirkegaard · 2024

Read Adaptive Node Positioning in Biological Transport Networks by Albert Alonso, Lars Erik J. Skjegstad, Julius B. Kirkegaard (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2411-00692.txt.

Description

Abstract Biological transport networks are highly optimized structures that ensure power-efficient distribution of fluids across various domains, including animal vasculature and plant venation. Theoretically, these networks can be described as space-embedded graphs, and rich structures that align well with observations emerge from optimizing their hydrodynamic energy dissipation.

Catalog record

CREATORAlbert Alonso, Lars Erik J. Skjegstad, Julius B. Kirkegaard
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2411-00692

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Adaptive Node Positioning in Biological Transport Networks Albert Alonso, Lars Erik J. Skjegstad, Julius B. Kirkegaard Abstract Biological transport networks are highly optimized structures that ensure power-efficient distribution of fluids across various domains, including animal vasculature and plant venation. Theoretically, these networks can be described as space-embedded graphs, and rich structures that align well with observations emerge from optimizing their hydrodynamic energy dissipation. Studies on these models typically use regular grids and focus solely on edge width optimization. Here, we present a generalization of the hydrodynamic graph model which permits additional optimization of node positioning. We achieve this by defining sink regions, accounting for the energy dissipation of delivery within these areas, and optimizing by means of differentiable physics. In the context of leaf venation patterns, our method results in organic networks that adapt to irregularities of boundaries and node misalignment, as well as overall improved efficiency. We study the dependency of the emergent network structures on the capillary delivery conductivity and identify a phase transition in which the network collapses below a critical threshold. Our findings provide insights into the early formation of biological systems and the efficient construction of transport networks. Transport networks are ubiquitous in nature and in living systems. The efficiency of such networks is crucial for the evolutionary fitness of organisms such as those observed in leaf venation [ 1 , 2 , 3 ] and blood vasculature systems [ 4 , 3 , 5 ] , and emerges in complex systems such as river networks [ 6 ] and human transport systems [ 7 , 8 , 9 ] . Understanding the structure and morphogenesis of such network structures has been facilitated by studying energetically optimal solutions to static, hydrodynamic networks [ 10 , 11 , 12 , 13 ] . The optimization of hydrodynamic transport networks is traditionally approached as an edge-optimization problem [ 10 , 5 , 2 , 13 ] , assuming systems where network nodes serve as sinks for their local area. Thus, optimizing the energy dissipation leads to optimal edge conductivities [ 10 ] , while the positions of the nodes themselves are considered fixed. In principle, this approach can be used to model any bounded system if enough nodes are used. However, for a finite number of nodes confined within a bounded system, it is evident that the node placement itself influences the optimality of the fluid delivery system. In many systems, the finite number of nodes is a physical fact and must be imposed, e.g., due to a lower bound on the vein thicknesses determined by the capillary size. Here, we consider bounded systems of finite nodes and generalize the hydrodynamic model to have well-defined optima both in edge conductivities and node positioning. Specifically, we model a bounded leaf venation network and let the boundary represent the leaf margin. Each node represents a source or sink (we set a single node at the leaf base as the source, and let all other nodes be sinks) and each edge represents a vein between two nodes. Fig. 1 a shows the optimal solution obtained for a fixed, hexagonal grid in a bounded system. The imposition of the domain boundary and the inability of the nodes to adapt positionally results in non-uniform areas associated with each node. As fluid dissipation will typically be proportional to area, this thus implies that the sinks become non-uniform. By removing the spatial constraint on the node positions, we find more energy-efficient solutions, which yield networks that appear organic and more consistent with networks observed in nature (Fig. 1 b). Figure 1: Optimization and relaxation on an inclined leaf ( \varphi_{0}=60^{\circ} , \beta=0.4 , N=175 ). Non-suppressed edges and delivery regions of each sink node are shown. The source is marked with a red dot. (a) Result after transport network optimization solving Eq. ( 3 ) on a hexagonal grid. (b) Result after relaxation, where the network optimizes P=P_{\mathrm{t}}+P_{\mathrm{d}} . The power of the transport network is given by [ 10 ] P_{\mathrm{t}}=\sum_{e\in\text{edges}}\left(\frac{F_{e}^{2}}{C_{e}}+c_{t}\,C_{e}^{\gamma}\right)L_{e}, (1) where F_{e} is the flux associated with edge e , which has length L_{e} and conductivity C_{e} . The constant c_{t} defines the metabolic cost of maintaining the edge, which is equivalent to considering a system with a finite amount of resources. \gamma is a parameter that determines how the material cost scales with the conductivities, which we set to \gamma=1/2 in this Letter. The flow is pressure-driven, such that the flux over a (directed) edge e=i\rightarrow j is given by F_{e}=(p_{i}-p_{j})C_{e}/L_{e} . The pressures are indirectly determined by the need to satisfy Kirchhoff’s law \sum_{e\in i}\pm_{e}F_{e}=s_{i} , where s_{i} is the source/sink at node i and the sign is set by the direction of the edge. This can be solved efficiently with hardware acceleration [ 14 ] . Typically, the sink magnitudes are assumed to be equal. However, once the nodes are permitted to move, these values must change as well. Thus, taking the source s_{0}=1 , we define s_{i}=-A_{i}/\sum{A_{i}} ( i\geq 1) , where A_{i} is the area surrounding a node, implicitly defined by the associated Voronoi cell. In order for a transport system to have stable and well-defined optima in node positioning, the above model must be expanded to account for the power dissipation within the Voronoi cells. To achieve this, we add a simple power-delivery term, P_{\mathrm{d}} , to the conventional formulation for transport power [ 10 ] , and thus assume that the total power can be described by the sum of two contributions P=P_{\mathrm{t}}+P_{\mathrm{d}} . With the goal of defining a self-consistent model that equally considers both transport and delivery costs, we take P_{\mathrm{d}}=\sum_{i\in\text{sinks}}\frac{s_{i}^{2}}{\hat{C}_{i}}\langle\ell\rangle_{i}. (2) This is analogous to the first term in the transport formulation (Eq. ( 1 )), but considers the power dissipation due to the delivery of the sink fluid s_{i} over an average Voronoi distance \langle\ell\rangle_{i} . \hat{C}_{i} is the delivery conductivity, which effectively models a capillary system. Physically, this term favors equally sized and isotropic sink areas. We assume that the delivery conductivity is an intrinsic property of the material, such that \hat{C}_{i}=\hat{C} is fixed and equal for all nodes. Optimization — Our scheme to minimize P consists of the combination of an edge conductivity optimization and a node relaxation process. We employ a fully connected graph and initialize all edge conductivities to the same value. For a given node positioning \{\boldsymbol{x}_{i}\} , we optimize the conductivities by using an adaptation model [ 13 , 2 ] \dfrac{\mathrm{d}C_{e}}{\mathrm{d}t}=\left(\dfrac{F^{2}_{e}}{C_{e}^{\gamma+1}}-\gamma c_{t}\right)C_{e}+c_{0}e^{-\lambda t}, (3) where the term F^{2}_{e}\mathbin{/}C_{e}^{\gamma+1} is the squared wall shear stress, and \gamma c_{t} represents the optimal squared shear stress. Furthermore, we include a growth term characterized by the area growth rate \lambda , which has proven to be a robust strategy to achieve better optima [ 2 ] . To facilitate smooth convergence, we also decrease exponentially the magnitude of the growth term c_{0}(\tau)\propto e^{-\nu\tau} during the node positioning relaxation process . Here, we use \tau to denote the timescale of node relaxation and t for the conductivity optimization. Crucially, even though our transport network is fully connected, optimization results in a sparse, planar graph (Fig. 1 ). While networks with static nodes can simplify the incidence matrix by,

Read the complete text