2017-用于植物学树模型的交互式木材燃烧

本文属于翻译文章，如有侵权，请联系删除。
原文地址

摘要

我们提出了一种植物树模型燃烧的新方法。树模型表示为分支结构的连接粒子和燃烧的多边形表面网格。每一个粒子都存储着生物和物理属性，这些属性驱动植物的动力学行为和燃烧的放热反应。再加上棒的现实物理学，粒子可以实现动态分支运动。我们对材料特性（如水分和碳化行为）进行建模，并将其与单个颗粒相关联。燃烧在多边形网格上的树模型的表面域中被有效地处理。用户可以通过引发火灾和在树枝上施加压力来动态地与模型交互。火焰通过消耗可用的资源在树模型中真实地传播。我们的方法以交互速率运行，并支持并行的多个树实例。我们通过大量实例证明了我们方法的有效性，并评估了其对真实木材样品燃烧的合理性。

1 简介

 As frequent objects in everyday scenes, models of trees and plants serve as important content in various application domains, ranging from architectural design and urban modeling to computer games and more recently the motion planning of autonomous agents. While many approaches exist to readily model the structure and visual appearance of vegetation, modeling the dynamic behavior of plants and their interaction with the environment still pose challenging problems. Faithfully modeling the plants’ response to physical effects, such as wind, rain, or snow arguably play a key role in many situations. However, the complexity of these natural phenomena renders their joint simulation impossible in most processing pipelines.Moreover, artists often require a nuanced means of control to adapt such effects as part of the content creation and the storytelling in their applications.

Traditionally, the adaptivity of trees and plants is modeled by explicitly considering the environment [Měch and Prusinkiewicz 1996], inverse procedural modeling [Stava et al. 2014], modeling the competition for resources [Palubicki et al. 2009], or simulated adaptation [Pirk et al. 2012]. Only recently, efforts in plant modeling concentrate on the dynamic behavior by simulating the physics response of plants. Realistic motions, also including the faithful simulation of growth [Longay et al. 2012], the adaptation to surfaces [Hädrich et al. 2017], and as induced by realistic materials [Wang et al.2017], play a key role for the convincing behavior of plants.

Current methods in computer graphics for simulating the combustion of solids ignore essential parts of tree combustion and cannot be easily adapted to botanical models of trees [Melek and Keyser 2002]. Physically-based simulations of wood combustion explored in other research areas, such as material sciences [Thi et al. 2016],are computationally demanding and only focus on the combustion of wood samples in laboratory conditions. In forestry, where the study of tree combustion is a key topic, no models have been devised at the scale of tree geometry [Seidl et al. 2012].

In this paper we advance tree modeling by introducing a combustion model for trees that is biologically plausible while also being computationally efficient. The key idea to our approach is to jointly simulate tree motion, fire, and the combustion of plant tissue. Fire is simulated through a volumetric grid-based fluid solver that transfers heat to plant tissue and initiates the combustion of the model.Conversely, a burning tree model releases heat to its environment,resulting in a positive feedback loop which leads to fire spread and maintains the combustion (Fig. 1).

A tree is represented by a dual formulation that employs particles for modeling the branching structure and a surface mesh for simulating the combustion. We use a position-based dynamics approach with connected particles to enable dynamic and realistic motions of branches [Kugelstadt and Schömer 2016]. Each particle stores biological and physical attributes to drive the kinetic behavior of the plant and the exothermic reaction of the combustion. Burning tree models results in inhomogeneous transformations that are difficult to simulate jointly with the fluid dynamics of fires. Moreover,
changes in the structure and the biomechanical properties of the plants directly affect the combustion.

In summary our key contributions are: (1) we propose a method for the combustion of botanical tree models based on heat transfer and integration of plausible physics of tree models and fluid dynamics; (2) we model trees based on position-based dynamics and Cosserat physics to support realistic plant motions and jointly simulate them with fluid dynamics for fire; (3) our framework maintains animation-ready plants at all stages of the simulation and supports the efficient interaction with plants in real-time; (4) we evaluate our implementation against the combustion of real wood samples.

作为日常场景中的常见对象，树木和植物模型在各种应用领域中都是重要的内容，从建筑设计、城市建模到计算机游戏，最近还包括自主代理的运动规划。尽管有许多方法可以很容易地对植被的结构和视觉外观进行建模，但对植物的动态行为及其与环境的相互作用进行建模仍然是一个具有挑战性的问题。忠实地模拟植物对物理效应（如风、雨或雪）的反应，可以说在许多情况下都起着关键作用。然而，这些自然现象的复杂性使得它们的联合模拟在大多数处理管道中是不可能的。此外，艺术家通常需要一种微妙的控制手段来适应这些效果，作为其应用程序中内容创作和故事讲述的一部分。

传统上，树木和植物的适应性是通过明确考虑环境[Měch和Prusinkiewicz 1996]、逆过程建模[Stava等人2014]、资源竞争建模[Palobicki等人2009]或模拟适应[Pirk等人2012]来建模的。直到最近，植物建模的工作才集中于通过模拟植物的物理响应来研究动态行为。真实的运动，也包括对生长的忠实模拟[Longay等人，2012]，对表面的适应[Hädrich等人，2017]，以及由真实材料诱导的[Wang等人，2017]，对植物令人信服的行为起着关键作用。

当前用于模拟固体燃烧的计算机图形学方法忽略了树木燃烧的基本部分，并且不能容易地适应树木的植物学模型[Meek和Keyser 2002]。在其他研究领域（如材料科学[Thi等人，2016]）探索的基于物理的木材燃烧模拟需要计算，并且只关注实验室条件下木材样品的燃烧。在森林中，树木燃烧的研究是一个关键课题，目前尚未设计出树木几何尺度的模型[Seidl等人，2012年]。

在本文中，我们通过为树木引入燃烧模型来推进树木建模，该模型在生物学上是合理的，但在计算上也是有效的。我们方法的关键思想是联合模拟树木运动、火灾和植物组织的燃烧。通过基于体积网格的流体解算器模拟火灾，该解算器将热量传递到植物组织并启动模型的燃烧。相反，燃烧树模型向其环境释放热量，导致正反馈回路，导致火灾蔓延并维持燃烧（图1）。

图1.树模型的燃烧：树暴露在火中，直到分支结构达到其点燃温度（a）。燃烧释放储存在树器官中的能量，并在整个树模型中传播，直到达到峰值（b）。燃烧导致树枝弯曲和断裂（c），而火焰征服更多树枝（d），最终烧毁整个树模型（e）

树由双重公式表示，该公式使用粒子来模拟分支结构，使用表面网格来模拟燃烧。我们使用连接粒子的基于位置的动力学方法来实现分支的动态和真实运动[Kugelstadt和Schömer 2016]。每个粒子都存储生物和物理属性，以驱动植物的动力学行为和燃烧的放热反应。燃烧树模型导致了不均匀的变换，难以与火灾的流体动力学一起模拟。此外，植物结构和生物力学特性的变化直接影响燃烧。

总之，我们的主要贡献是：（1）我们提出了一种基于传热的植物树模型燃烧方法，并结合了树木模型和流体动力学的物理合理性；（2） 我们基于位置的动力学和Cosserat物理对树木进行建模，以支持真实的植物运动，并与火灾的流体动力学共同模拟它们；（3） 我们的框架在模拟的所有阶段维护动画就绪的植物，并支持与植物的实时有效交互；（4） 我们针对真实木材样品的燃烧评估了我们的实施

2 RELATED WORK

Early approaches of modeling vegetation focus on the refined definition of branching structures through a variety of approaches,such as fractals and repetitive patterns [Oppenheimer 1986] or L-Systems [Prusinkiewicz 1986]. Later methods focus on the interaction of plants with their environment, through query modules [Měch and Prusinkiewicz 1996], inverse procedural modeling [Stava et al. 2014] and through explicitly modeling the competition for resources [Palubicki et al. 2009]. Data-driven and sketch based approaches focus on modeling plants from sensor data [Livny et al. 2011; Neubert et al. 2007] or user-defined sketches [Ijiri et al.2006; Okabe et al. 2007].

Recent approaches provide more nuanced ways to efficiently model branching structures with an emphasis on simulating the response of trees to their environment [Pirk et al. 2014, 2012], interactive growth [Pirk et al. 2012] or FEM simulation [Zhao and Barbič 2013]. Longay et al. [2012] introduce an interactive tree modeling system that jointly models trees based on the competition for resources and user defined sketches. These approaches have in common to lift tree models from static branching structures to more dynamic representations, however, they do not model the interaction of plants with plausible physics.

Another line of work invests in techniques for increasing the realism of the physics response of elastic filaments, such as plant tendrils or assemblies of fibers and nearly one-dimensional slender structures in general [Casati and Bertails-Descoubes 2013]. Selle et al. [2008] simulate the dynamics of thinned out hair with several thousand individual fibers. A single fiber is constructed using a tetrahedral mass-spring system that is furnished with additional ad-hoc springs in order to prevent from a volumetric collapse. Such mass-spring models are rather popular in graphics because of their efficiency, however, physically accurate results are complicated to produce since realistic material parameters are hard to integrate without running into numerical issues [Michels et al. 2015].Pai [2002] introduce the theory of Cosserat rods to the computer graphics community and it has been shown that this leads to physically accurate results for modeling the dynamics of rods and fibers [Bertails et al. 2006]. More recently, Kugelstadt and Schömer [2016] showed that even bending and torsion of elastic rods can be approximated with a position-based dynamics (PBD) framework [Müller et al. 2007], by introducing additional constraints.

Elastic rods play a key role in simulating the physics response of tree models. Hädrich et al. [2017] model the growth behavior and physics response of climbing plants with an emphasis on dynamic control and interactive authoring.The simulation and animation of fire has a long tradition in computer graphics. Commonly, fire is simulated through grid-based fluid solvers that allow capturing laminar and turbulent features in flames [Hong et al. 2007; Nguyen et al. 2002; Stam 1999] and in smoke [Fedkiw et al. 2001]. Pegoraro and Parker [2006] introduce a method for the physically-based rendering of fire dynamics that builds on molecular physics and Horvath and Geiger [2009] propose an efficient implementation that combines a coarse particle simulation with high resolution view-dependent grids.

The combustion of rigid bodies can be modeled through volumetric grids that allow to track disconnected propagating fronts [Liu et al. 2012; Melek and Keyser 2005; Zhao et al. 2003] and to spread fires on surfaces [Chiba et al. 1994]. Hong et al. [2010] propose a method to model fires under general geometric constraints that even allows to burn plant models. Finally, Stomakhin et al. [2014] introduce a point-based technique for melting and solidifying materials through simulating heat transfer. Their method allows to capture the thermodynamic properties in different materials and to alter their mechanical properties, however, it does not support interactive modeling and heat diffusion in real-time.

Research in forestry and botany focuses on the heat transfer of entire stands of trees [Bohren and Thorud 1973; Encinas et al. 2007],on models to describe the biomechanical properties of wood with respect to charring behavior [Lizhong et al. 2002], fire resistance of wood [Lawes et al. 2011], or the pyrolysis of trees and plants in general [Demirbaş 2000]. The resistance of trees against fires is mostly determined by the thickness of their bark and several approaches try to quantify the resistance behavior of certain species [Lawes et al. 2011]. Many of these techniques describe accurate models of fire propagation and heat transfer in plants, however, they cannot directly be applied to detailed geometric models of trees and do not support the interactive exploration of combustion in real-time.

2 相关工作

早期的植被建模方法侧重于通过各种方法对分支结构进行精确定义，如分形和重复模式[Openheimer 1986]或L-Systems[Prusinkiewicz 1986]。后来的方法通过查询模块[Měch和Prusinkiewicz 1996]、逆过程建模[Stava等人2014]和明确建模资源竞争[Paubicki等人2009]，关注植物与其环境的相互作用。数据驱动和基于草图的方法侧重于从传感器数据（Livny等人，2011年；Neubert等人，2007年）或用户定义的草图（Ijiri等人，2006年；Okabe等人，2007）对植物进行建模。

最近的方法提供了更精细的方法来有效地建模分支结构，重点是模拟树木对其环境的响应[Pirk等人，2014年，2012年]、交互生长[Pirk et al.2012年]或FEM模拟[Zhao和Barbič2013年]。Longay等人【2012】介绍了一种交互式树建模系统，该系统基于资源和用户定义草图的竞争对树进行联合建模。这些方法的共同点是将树模型从静态分支结构提升为更动态的表示，然而，它们没有用合理的物理来模拟植物的相互作用。

另一项工作投资于提高弹性细丝的物理响应真实性的技术，如植物卷须或纤维组件以及大致一维细长结构[Casati和Bertails Descoubes 2013]。Selle等人【2008】用几千根单独的纤维模拟稀疏头发的动力学。使用四面体质量弹簧系统构建单个纤维，该系统配备有额外的自组织弹簧，以防止体积塌陷。这种质量弹簧模型由于其效率而在图形中相当流行，然而，由于真实的材料参数很难集成而不涉及数值问题，因此物理精确的结果很难产生[Michels等人，2015]。Pai[2002]将Cosserat棒的理论引入计算机图形界，并已表明，这将导致棒和纤维动力学建模的物理精确结果[Bertails等人2006]。最近，Kugelstadt和Schömer【2016】表明，通过引入额外的约束，即使弹性杆的弯曲和扭转也可以通过基于位置的动力学（PBD）框架来近似[Müller等人2007]。

弹性杆在模拟树木模型的物理响应中起着关键作用。Hädrich等人[2017]建模了攀援植物的生长行为和物理响应，重点是动态控制和交互式创作。火灾的模拟和动画在计算机图形学中有着悠久的传统。通常，火灾通过基于网格的流体求解器进行模拟，该求解器允许捕捉火焰中的层流和湍流特征[Hong等人，2007；Nguyen等人，2002；Stam 1999]和烟雾中的层流特征[Fedkiw等人，2001]。Pegoraro和Parker[2006]介绍了一种基于分子物理的火灾动力学物理渲染方法，Horvath和Geiger[2009]提出了一种将粗糙粒子模拟与高分辨率视图相关网格相结合的有效实现。

刚体的燃烧可以通过体积网格进行建模，该网格允许跟踪不连续的传播前沿[Liu等人2012；Melek和Keyser 2005；Zhao等人2003]，并在表面上蔓延火焰[Chiba等人1994]。Hong等人【2010】提出了一种在一般几何约束条件下模拟火灾的方法，甚至允许燃烧植物模型。最后，Stomakhin等人【2014】介绍了一种基于点的技术，通过模拟传热来熔化和固化材料。他们的方法允许捕捉不同材料的热力学性质并改变其力学性质，但是，它不支持实时交互建模和热扩散。

林业和植物学的研究侧重于整个树木林分的热传递[Bohren和Thorud 1973；Encinas等人2007]，以及描述木材在炭化行为方面的生物力学特性的模型[Lizhong等人2002]，木材的耐火性[Lawes等人2011]，或一般树木和植物的热解[Damirbaş2000]。树木对火灾的抵抗力主要取决于树皮的厚度，有几种方法试图量化某些物种的抵抗行为[Lawes等人，2011]。这些技术中的许多描述了火灾在植物中传播和传热的精确模型，然而，它们不能直接应用于树木的详细几何模型，并且不支持实时交互探索燃烧。

# 3  OVERVIEW
The realistic combustion of tree models requires simulating fluid dynamics for fire, modeling dynamic tree models – including geometric deformations of the branch mesh, and the simulation of the heat transfer. Fig. 2 shows an overview of our system. The input to our method is a tree model represented as a skeletal graph, and a set of parameters for both, the fire simulation and the combustion. Fire is simulated through a grid-based fluid solver that approximates the Navier-Stokes equations. For the implementation we rely on established solvers that allow to simulate characteristic features of flames at interactive rates [Fedkiw et al. 2001]. The combustion of flames is simulated based on an approach similar to Melek and Keyser et al. [2002], that supports laminar and turbulent flames as well as the burning of solid and gaseous fuels.

Trees are modeled as environmental-sensitive particles that allow plausible branch motions. Branches are formed by connecting strings of particles in a hierarchical relationship. We couple the particles with Cosserat physics [Kugelstadt and Schömer 2016],which enables more realistic motions for solitary trees, including the bending and torsion of branches. Moreover, each tree particle is associated with physical and biological attributes, such as strain,moisture, and insulation, that define the models’ response to the fire simulation.

The branch mesh is dynamically updated according to attribute changes of the particles and the combustion simulation. Our method propagates heat on the surface of branches and deforms the branch geometry according to the combustion. The combustion starts when plant tissue is exposed to a fire until the ignition temperature is reached. The temperature propagates through the plant tissue depending on the local material properties.

3 概述

树模型的真实燃烧需要模拟火灾的流体动力学，建模动态树模型——包括分支网格的几何变形，以及热传递的模拟。图2显示了我们系统的概述。我们方法的输入是以骨架图表示的树模型，以及火灾模拟和燃烧的一组参数。通过基于网格的流体解算器模拟火灾，该解算器近似于Navier-Stokes方程。对于实现，我们依赖于已建立的求解器，该求解器允许以交互速率模拟火焰的特征[Fedkiw等人2001]。火焰的燃烧基于类似于Melek和Keyser等人[2002]的方法进行模拟，该方法支持层流和湍流火焰以及固体和气体燃料的燃烧。

树木被建模为环境敏感粒子，允许合理的树枝运动。分支是通过以层次关系连接粒子串而形成的。我们将粒子与Cosserat物理结合[Kugelstadt和Schömer 2016]，这使得孤立的树木能够进行更真实的运动，包括树枝的弯曲和扭曲。此外，每个树粒子都与物理和生物属性相关联，例如应变、湿度和绝缘，这些属性定义了模型对火灾模拟的响应。

根据粒子的属性变化和燃烧模拟动态更新分支网格。我们的方法在树枝表面传播热量，并根据燃烧使树枝几何结构变形。当植物组织暴露于火中直到达到点燃温度时，燃烧开始。温度通过植物组织传播取决于局部材料特性。

# 4 DYNAMIC TREE MODELS
Dynamic motions are essential for the realistic simulation of combustion. While temperature differences of flames and surrounding air produce wind that sets branches and leaves in motion, the combustion of plant tissue gradually reduces the thickness of branches and causes them to bend, break, and eventually fall down on the ground. Our system is able to efficiently simulate these effects by exploiting a particle-based representation for trees and by using simple algebraic expressions for the integration of position and orientation updates for branching structures.

Similar to other approaches [Livny et al. 2011], the input to our method is a graph-based representation of a tree model. The branching structure is defined as a skeletal graphG = {N,H}, where N are the nodes and H the edges. Each edge connects two nodes ns and nt to produce a hierarchal and non-cyclic relationship, where ns is the parent of nt . One node nroot is the root node of the tree model.Single branches are defined as chains of edges C = {h1,h2, . . . ,hn },where n denotes the length of a given branch.

For the simulation of trees we use a discretized Cosserat rod tree representation. Cosserat rods are described by a smooth curve r(s) : [s0,s1] → IR3. Each point on the curve has an orthonormal material frame with the attached basis vectors {d1(s), d2(s), d3(s)}.The material frame vectors dk are called directors. The orthonormal frame can be described by a quaternion as a rotation of the world space basis vectors {e1, e2, e3} (Fig. 3, left). The rods are discretized by a set of particles and material frames between them. Two particles and one quaternion that stores the orientation of the material frame
will be called rod element (Fig. 3, right). Adjacent rod elements share one particle.

We convert the graph-based representation into rod elements by recursively traversing the tree graph starting at nroot. Nodes in the tree graph are represented by particles and edges are converted into rod elements. We compute the quaternion of a rod element by rotating the quaternion un−1 of the parent rod element so that the director d3n−1 aligns with the current edge direction:

where un is the quaternion for the new rod element, un−1 the orientation of the previous rod element, R returns the rotation quaternion for a given axis and angle, and d3n−1 is the director of the previous rod element, hˆn the unit vector of the new edge, hn−1 the edge of the parent rod element, and ∠ denotes the angle between two vectors. The length of the rod element is given by the edge length from the tree graph. We associate the thickness of a graph node with the corresponding particle for generating the surface mesh.

Tree dynamics are realized by employing position and orientation based Cosserat rods [Kugelstadt and Schömer 2016] using connected particles. Their method accounts for bending and torsion of elastic rods, which makes it well suited for the dynamic modeling of branching structures. In contrast to the original approach, we additionally define branching structures. Each particle can have multiple outgoing rod elements to its children (Fig. 3, right) and each child rod element stores an orientation and the corresponding length. Cosserat physics can be applied to branching structures by summing up the corrections for particle positions and rod element orientations in a first step followed by the application of these changes to the particles and rod elements respectively.

4 动态树模型

动态运动对于燃烧的真实模拟至关重要。当火焰和周围空气的温差产生风，使树枝和树叶运动时，植物组织的燃烧会逐渐减少树枝的厚度，导致树枝弯曲、断裂，最终落在地面上。我们的系统能够通过利用基于粒子的树表示以及使用简单的代数表达式来集成分支结构的位置和方向更新来有效地模拟这些效果。

与其他方法类似[Livny等人，2011年]，我们方法的输入是基于图的树模型表示。分支结构被定义为骨架图G={N，H}，其中N是节点，H是边。每条边连接两个节点ns和nt，以产生一个层次和非循环关系，其中ns是nt的父节点。一个节点nroot是树模型的根节点。单个分支定义为边链C={h1，h2，…，hn}，其中n表示给定分支的长度。

对于树的模拟，我们使用离散化的Cosserat Rod树表示。Cosserat rods 由光滑曲线$r(s):[s_0,s_1]{\mathbb{R}}^3$描述.曲线上的每个点都具有正交材料框架，该框架具有附加的基向量$d_1(s),d_2(s),d_3(s)$。材质帧向量$d_k$称为导向器。正交帧可以用四元数描述为世界空间基向量{e1，e2，e3}的旋转（图3，左侧）。杆由一组粒子集合和它们之间的材料框架离散化。存储材质帧方向的两个粒子和一个四元数，将被称为杆元件（图3，右侧）。相邻的杆元素共享一个粒子。

我们通过递归遍历从$n_{root}$开始的树图，将基于图的表示转换为杆元素。树图中的节点由粒子表示，边转换为杆元素。我们通过旋转四元数$u_{n-1}$来计算父杆元素的四元数，使得导向器$d_{3_{n-1}}$与当前边方向对齐：

$$u_n=R\left({\mathbf{d}}_{3_{n-1}}\times {\hat{\mathbf{h}}}_n,\angle \left({\mathbf{h}}_n,{\mathbf{h}}_{n-1}\right)\right)u_{n-1}\text{\hspace{1em}(1)}$$

其中，$u_n$是新杆元素的四元数，$u_{n-1}$上一个杆元件的方向，$R$返回给定轴和角度的旋转四元数，${\mathbf{d}}_{3_{n-1}}$是前一个杆元素的指向矢，${\hat{\mathbf{h}}}_n$是新边的单位向量，${\mathbf{h}}_{n-1}$母杆元件的边缘，以及∠ 表示两个向量之间的角度。杆元件的长度由树形图的边缘长度给出。我们将图形节点的厚度与用于生成曲面网格的相应粒子相关联。

通过使用基于位置和方向的Cosserat杆[Kugelstadt和Schömer 2016]使用连接粒子实现树木动力学。他们的方法考虑了弹性杆的弯曲和扭转，这使得它非常适合于分支结构的动态建模。与原始方法不同，我们还定义了分支结构。每个粒子可以有多个输出杆元素到其子元素（图3，右侧），每个子杆元素存储一个方向和相应的长度。Cosserat物理学可以通过在第一步中对粒子位置和杆元件取向的校正进行总结，然后将这些变化分别应用于粒子和杆元件，从而应用于分支结构。

# 5 FIRE SIMULATION
Colloquially, fire denotes a set of flames and its evolution during the process of combustion. More precisely, fire can be defined as chemical reaction of oxygen and flammable gases (fuel) that releases various reaction products like water and carbon dioxide, as well as heat and light, which together poses the fire.

We simulate fire as a temperature field evolving in a gaseous fluid according to Fedkiw et al. [2001]. After the velocity update, the temperature is updated according to a modified diffusion process.For that, the distribution of the temperature T in the fluid domain is governed by

where u® denotes the fluid’s velocity field, Tamb the ambient temperature, and α the thermal diffusivity of air. The term involving the fourth power of the temperature difference and a non-negative constant γ is the radiative cooling term adapted from Nguyen et al. [2002]. The time- and position-dependent function sT represents the input of heat generated by the burning process into the fire simulation. It is provided by the combustion simulation. Furthermore,
the combustion process generates smoke. Its density S is updated according to a passive advection process with the fluid described by

The source term sS is also provided by the combustion model. Both,sT and sS are described in Sec. 6.2.

The fluid equations are discretized on a rectangular Eulerian grid using finite differences. We use Lie operator splitting to numerically solve equations with multiple terms. All advection terms are handled by the unconditionally stable version of the MacCormack method [1969], while the projection of the tentative velocity is performed using a Jacobi iteration. Velocity damping and temperature cooling terms are integrated analytically. All remaining terms use forward Euler integration. We do not use trees as a boundary condition for the fluid simulation, but instead use the heat and smoke generation terms sT (Eq. 2) and sS , (Eq. 3), provided by our combustion model.Furthermore, we define the outer boundary of our fluid domain as p = 0, t = Tamb, S = 0 and ∂nu® = 0 and do not explicitly model the ground boundary.

5 火灾模拟

通俗地说，火是指一组火焰及其在燃烧过程中的演变。更准确地说，火灾可以定义为氧气和可燃气体（燃料）的化学反应，释放出各种反应产物，如水和二氧化碳，以及热和光，这些反应产物共同引发火灾。

根据Fedkiw等人[2001]，我们将火灾模拟为气态流体中的温度场演变。在速度更新之后，根据修改的扩散过程更新温度。因此，温度T在流体域中的分布由

$${\partial }_{t}T+\vec{u}\cdot \nabla T=\alpha {\nabla }^{2}T+\gamma {\left(T-T_{\mathrm{a}\mathrm{m}\mathrm{b}}\right)}^4+s_{\mathrm{T}},\text{\hspace{1em}(2)}$$

其中$\vec{u}$表示流体的速度场，$T_{amb}$表示环境温度，$\alpha$表示空气的热扩散率。涉及温差四次方和非负常数$\gamma$的项是根据Nguyen等人[2002]改编的辐射冷却项。时间和位置相关函数$s_T$表示燃烧过程产生的热量输入到火灾模拟中。它由燃烧模拟提供。此外燃烧过程产生烟雾。其密度S根据被动平流过程进行更新，流体描述如下：

$${\partial }_{t}S+\vec{u}\cdot \nabla S=s_T\text{\hspace{1em}(3)}$$

源项$s_S$也由燃烧模型提供。第6.2节描述了$s_T$和$s_S$。

使用有限差分在矩形欧拉网格上离散流体方程。我们使用Lie算子分裂来数值求解具有多个项的方程。所有平流项均由MacCormack方法[1969]的无条件稳定版本处理，而暂定速度的投影则使用雅可比迭代进行。对速度阻尼和温度冷却项进行了解析积分。所有剩余术语均使用前向欧拉积分。我们不使用树作为流体模拟的边界条件，而是使用燃烧模型提供的热量和烟雾生成项$s_T$（方程式2）和$s_S$（方程式3）。此外，我们将流体域的外边界定义为$p=0，t=T_{amb}，S=0$和$\partial n\vec{u}$=0，并且不显式建模地面边界。

# 6 COMBUSTION OF TREE MODELS
The combustion of a solid fuel starts when it is exposed to heat until its ignition temperature is reached. For charring materials, such as wood, cloth, and paper, combustion additionally involves the decomposition of material into char and flammable gases (fuel). This process, which can be considered as a part of the combustion, is called pyrolysis [Emmons and Atreya 1982]. At a high level of abstraction the phenomenon can be divided into three separable components: (1) thermal decomposition of material (pyrolysis); (2) conditions for flame spread (ignition); (3) heat and mass transfer from the burning zone into the plant tissue (fire spread). For our method, we assume pyrolysis to occur in a relatively thin zone at the branch tissue, instead of operating on the full volume of a branch. This propagating zone is referred to as the pyrolyzing front. In the following sections we discuss a mathematical model that includes these components and allows for modeling the realistic combustion of woody tissue for botanical tree models.

6.1 Interactive Wood Combustion
We aim to create a computationally efficient model for the combustion of wood that can be used for three-dimensional models of trees.Specifically, our model expresses the pyrolysis of wood, including the evaporation of moisture, the formation of a char layer, and the reduction of virgin wood. Moreover, we model the ignition of woody tissue and the subsequent flame spread including the heat transfer across the tree as well as in the surrounding air.

Our combustion model relies on a discrete representation of the branch geometry. We use a polygonal surface mesh, where each surface element (triangle) represents a volumetric portion of the branch,which can be defined as a set of points traced out by a continuous projection of each surface point onto the closest point on the branch axi (Fig. 4, left and inset, right). Each surface element is associated with information about the total mass M of the underlying branch volume, the surface temperature at this location TS, the water content W ,the thickness of a possible char layer HC,the thickness of the virgin wood H and the area of the inwardly moving, pyrolyzing front A. 
Spatial transfer of information happens only in terms of heat in our model. This means, we neglect the spatial mass transfer, such as the release of flammable gases, and assume an immediate effect of pyrolysis on e.g. air temperature. Given this scale of abstraction we can summarize the most important assumptions in our combustion model as follows:

• The tissue layer currently undergoing pyrolysis is approximated by a surface (Pyrolyzing Front).
• The thickness of the char layer is determined by the amount of pyrolyzed wood (Char Contraction).
• The char layer has an insulating effect proportional to its thickness (Char Insulation).
• The rate of wood combustion is proportional to surface temperature (Reaction Rate Coefficient).
• Heat transfer from air to surface is inversely proportional to the amount of water content of surface elements (Moisture).
• Pyrolysis adds heat to the surrounding air proportional to the mass loss rate of surface elements (Heat Transfer).

6 树木模型燃烧

当固体燃料暴露于热中直到达到其点燃温度时，其燃烧开始。对于烧焦材料，如木材、布料和纸张，燃烧还包括将材料分解为炭和可燃气体（燃料）。这一过程被认为是燃烧的一部分，被称为热解[Emmons和Atreya 1982]。在高度抽象的情况下，该现象可分为三个可分离的部分：（1）材料的热分解（热解）；（2） 火焰蔓延（点燃）条件；（3） 热量和质量从燃烧区转移到植物组织（火灾蔓延）。对于我们的方法，我们假设热解发生在分支组织的相对薄的区域，而不是在分支的整个体积上进行操作。这个传播区被称为热解前沿。在下面的章节中，我们讨论了一个包含这些组件的数学模型，并允许为植物树模型建模木质组织的真实燃烧。

6.1 交互式木材燃烧

我们的目标是为木材燃烧创建一个计算效率高的模型，该模型可用于树木的三维模型。具体来说，我们的模型表达了木材的热解，包括水分的蒸发、炭层的形成和原始木材的还原。此外，我们还模拟了木质组织的点燃和随后的火焰传播，包括树木和周围空气中的热传递。

我们的燃烧模型依赖于分支几何结构的离散表示。我们使用多边形曲面网格，其中每个曲面元素（三角形）表示分支的体积部分，可以将其定义为通过每个曲面点到分支轴上最近点的连续投影所描绘的一组点（图4，左侧和右侧插图）。每个表面元素与关于下面的分支体积的总质量$M$、该位置处的表面温度$T_S$、水含量W、可能的炭层$H_C$的厚度、原始木材$H$的厚度以及向内移动的热解前缘$A$的面积的信息相关联。

在我们的模型中，信息的空间传递只发生在热量方面。这意味着，我们忽略了空间传质，如可燃气体的释放，并假设热解对空气温度的直接影响。鉴于这种抽象的规模，我们可以将燃烧模型中最重要的假设总结如下：

•当前正在进行热解的组织层近似于表面（热解前沿）。

•炭层的厚度取决于热解木材的数量（炭收缩）。

•炭层具有与其厚度成比例的绝缘效果（炭绝缘）。

•木材燃烧速率与表面温度（反应速率系数）成比例。

•从空气到表面的传热与表面元素的含水量（水分）成反比。

•热解将热量添加到周围空气中，与表面元素的质量损失率成比例（热传递）。

6.1.1 Pyrolyzing Front. Pyrolysis occurs in a thin layer of the branch tissue. We assume a fixed temperature for pyrolysis and a single step global reaction scheme (infinite rate kinetics) – the virgin wood is decomposed into volatiles and char [Galgano and Di Blasi 2005]:

We represent the pyrolyzing front as a two-dimensional triangular surface with area A, which is given by A = A0 H/H0. The original depth of the volumetric portion represented by a surface element is denoted by H0 and the current height of the pyrolyzing front is given by H. At the onset of pyrolysis the pyrolyzing front is located at the branch surface and matches exactly the area A0 of the surface element (H = H0). As pyrolysis continues, the pyrolyzing front will move from the surface towards the branch axis (the value of H decreases), resulting in a decreasing area A (Fig. 4, middle and right). The current height H of the pyrolyzing front is given by


where ρ is the density of the virgin wood. M denotes the current mass of virgin wood expressed as a function of time. Thus, the current height H of the pyrolyzing front in each time step of the simulation can be obtained from the value of the current mass of virgin wood M, as the remaining factors in Eq. (4) are constant.

6.1.2 Char Contraction Factor. As the pyrolyzing front is moving towards the branch center in our model, we calculate the formation of a char layer. Studies in wood engineering show that a reasonable approximation for the char layer thickness can be obtained from the current height of the pyrolyzing frontH. We use the char contraction factor κ to obtain an estimate for the char layer thickness Hc. The char contraction factor is defined as the fraction of the thickness of the char layer, when the pyrolysis is complete, and the original thickness of the virgin wood. Various species of trees and types of wood can have different char contraction factors [Parker 1989].We use an experimentally derived char contraction factor κ, in the range of [0.5, 1.0], to calculate the char layer thickness


6.1.3 Char Insulation. Various experiments in wood engineering indicate that, with the formation of a char layer, the mass loss rate of virgin wood starts decreasing significantly [Friquin 2011].Therefore, the insulation effect of char is an important phenomenon to include in a combustion model of wood. However, a detailed physical simulation of this phenomenon would necessitate a description of volumetric heat transfer across the char layer and different woody tissues. We therefore approximate the insulating effect of the char layer with the non-dimensional parameter c, which is non-linearly proportional to the char layer thickness Hc. It is given by 


At the start of pyrolysis, c equals 1 as the char layer thickness Hc is equal to 0. Subsequently, as Hc increases, the char insulation parameter c decreases exponentially to cmin, where cmin represents the maximum insulation effect of char in the range of [0, 1].

6.1.4 Reaction Rate Coefficient. Finally, we assume that the mass loss rate is proportional to the surface temperature T . As we only define temperatures on the branch surface, we obtain the rate of change of mass from the branch surface temperature. We introduce a temperature dependent reaction rate k(T ) and define the rate of change of mass for virgin wood as an ordinary differential equation:


The rate of change of mass depends on the dimensionless char insulation parameter c, the area of the pyrolyzing front A, and the reaction rate given by 

in which S(x) = 3x 2 − 2x 3 describes a sigmoid-like function interpolating smoothly from zero to one for temperatures between T0 = 150◦C and T1 = 450◦C.

Unlike our definition of k, the mass loss can also be computed based on the reaction rate coefficient defined in the Arrhenius equation [Mcnaught and Wilkinson 1997]. However, we found that a first order Arrhenius reaction scheme did not suffice as approximation of wood pyrolysis in our case. For our model we require an unignited state of wood which is not supported by the Arrhenius equation. Our reaction rate coefficient k describes the rate of mass change adequately for both, the unignited and ignited, states of wood. Moreover, it simplifies the mathematical description of our model, while still qualitatively capturing the effects of temperature on mass loss rate.The boundary condition for Eq. (7) is given by the density of wood ρ and the initial mass M0 = A0/2H0ρ of the volumetric portion of wood associated with the surface elements.

6.1.5 Moisture and Smoke. One of the main parameters of wood combustion is moisture, as it affects the temperature changes on the branch surface. A piece of wood with high water content can prolong pyrolysis to the point of prohibiting ignition altogether. We represent moisture in our model explicitly by a separate parameter W , with values in the range of [0, 1]. High values of W denote a piece of wood with a high level of water content and a low value represents dry wood. In our model we assume that moisture affects the rate of change of surface temperatures via the coefficient

where bdry and bwet are coefficients representing the heat transfer rates for dry and wet wood. We calculate the value of the water content W based on the evapotation ratew and the surface temperature Ts by solving the ordinary differential equation:

Our fire simulation (Sec. 5) expresses smoke by the smoke density parameter S (Eq. 3). We control the amount of smoke in our fire simulation by a combustion dependent source term sS . Specifically,we assume that denser smoke is the result of wood combustion with higher moisture content W :

where SM and SW are coefficients to express the rate at which smoke is added proportionally to the mass loss and the evaporation of water.

6.1.1 热解前部。热解发生在分支组织的薄层中。我们假设热解温度固定，采用单步全局反应方案（无限速率动力学）——原始木材分解为挥发物和炭[Galgano和Di Blasi 2005]：

我们将热解前沿表示为具有面积a的二维三角形表面，面积a由a=A0 H/H0表示。由表面元素表示的体积部分的原始深度由H0表示，热解前沿的当前高度由H表示。在热解开始时，热解前沿位于分支表面，并与表面元素的面积A0精确匹配（H＝H0）。随着热解的继续，热解前沿将从表面向分支轴移动（H值减小），导致面积a减小（图4，中间和右侧）。热解前沿的当前高度H由下式给出

其中ρ是原始木材的密度。M表示原始木材的当前质量，表示为时间的函数。因此，由于方程（4）中的其余因素是恒定的，因此可以从原始木材M的当前质量值获得模拟的每个时间步骤中热解前缘的当前高度H。

6.1.2 炭化收缩系数。在我们的模型中，随着热解前沿向分支中心移动，我们计算了炭层的形成。木材工程研究表明，可以从热解前沿$H$的当前高度获得炭层厚度的合理近似值。我们使用炭化收缩因子$k$来获得炭层厚度$H_c$的估计值。炭化收缩系数定义为热解完成时炭层厚度与原始木材原始厚度的比值。不同种类的树木和木材可能具有不同的炭化收缩因子[Parker 1989]。我们使用实验得出的炭收缩因子$k$在[0.5，1.0]范围内来计算炭层厚度。

$$H_c=k(H_0-H)\text{\hspace{1em}(5)}$$

6.1.3炭绝缘。木材工程中的各种实验表明，随着炭层的形成，原始木材的质量损失率开始显著降低[Friquin 2011]。因此，木炭的隔热效果是木材燃烧模型中的一个重要现象。然而，这一现象的详细物理模拟将需要描述炭层和不同木质组织之间的体积传热。因此，我们用与炭层厚度Hc成非线性比例的无量纲参数c来近似炭层的绝缘效应。由以下人员给出

$$c=c_{min}+(1-c_{min})e^{-c_x{H}_c}\text{\hspace{1em}(6)}$$

热解开始时，当炭层厚度Hc等于0时，c等于1。随后，随着Hc的增加，炭绝缘参数c呈指数下降至cmin，其中cmin表示炭在[0，1]范围内的最大绝缘效果。

6.1.4反应速率系数。最后，我们假设质量损失率与表面温度$T$成正比。由于我们只定义了分支表面上的温度，我们从分支表面温度获得了质量变化率。我们引入了温度相关的反应速率$k(T)$，并将原始木材的质量变化率定义为一个常微分方程：

$$d_{t}M+kcA=0\text{\hspace{1em}(7)}$$

质量变化率取决于无量纲炭绝热参数c、热解前沿A的面积和下式给出的反应速率

$$k(T)=\Delta m\cdot \{\begin{array}{ll}0& T<T_0\\ S\left(\left(T-T_0\right)/\left(T_1-T_0\right)\right)& T_0\le T\le T_1\\ 1& T>T_1\end{array}\text{\hspace{1em}(8)}$$

其中$S(x)=3x^2-2x^3$描述了在$T_0=150C$之间的温度下从零到一平滑插值的S形函数$T_1=450^{◦}C$

与我们对k的定义不同，质量损失也可以根据Arrhenius方程[Mcanot和Wilkinson 1997]中定义的反应速率系数来计算。然而，我们发现，在我们的情况下，一级阿伦尼乌斯反应方案不足以近似木材热解。对于我们的模型，我们需要一个未知的木材状态，这是阿伦尼乌斯方程不支持的。我们的反应速率系数k充分描述了木材未点燃和点燃状态的质量变化速率。此外，它简化了模型的数学描述，同时定性地捕捉了温度对质量损失率的影响。方程（7）的边界条件由木材密度ρ和与表面元素相关的木材体积部分的初始质量M0=A0/2H0ρ给出。

6.1.5 湿气和烟雾。木材燃烧的主要参数之一是水分，因为它影响树枝表面的温度变化。一块含水量高的木材可以将热解时间延长到完全禁止点燃的程度。我们用单独的参数W明确表示模型中的水分，值在[0，1]范围内。W值高表示含水量高的木材，而W值低表示干木材。在我们的模型中，我们假设水分通过系数
$$b=(1-W)b_{dry}+W{b}_{wet}\text{\hspace{1em}(9)}$$
其中$b_{dry}$和$b_{wet}$是表示干木材和湿木材的传热速率的系数。我们通过求解常微分方程，根据蒸发速率$w$和表面温度$T_s$计算含水量$W$的值：
$$d_{t}W+w{T}_s=0\text{\hspace{1em}(10)}$$
我们的火灾模拟（第5节）通过烟雾密度参数S（方程式3）表示烟雾。我们通过燃烧相关源项sS控制火灾模拟中的烟雾量。具体而言，我们假设更密集的烟雾是含水量W更高的木材燃烧的结果：
$$s_S=-S_M{d}_{t}M-S_W{d}_{t}W\text{\hspace{1em}(11)}$$
其中SM和SW是表示烟与质量损失和水蒸发成比例增加的速率的系数。

## 6.2 Heat Transfer
Combustion of wood is expressed in terms of bidirectional heat transfer between tree models and the surrounding air. A comparatively higher air temperature raises branch surface temperatures,which, when high enough, results in an increase of the mass loss rate (Eq. 7). The mass loss of surface elements, on the other hand,induces an increase in air temperature (Eq. 2, 13). Therefore, the heat transfer constitutes a positive feedback loop that enables the spread of flames in our model.

6.2.1 Surface Heat Transfer. The temperature Ts of a surface element is determined by taking into account the heat diffusion on the branch surface and the conduction of heat between air and surface. The heat conduction between air and branch surface (surface elements) is proportional to the difference in temperature between them. The rate of change of surface temperature Ts is thus given by

The left-most term on the right side of the equation describes the heat diffusion on the branch surface using the Laplace operator and a temperature diffusion coefficient a. The right-most term expresses Newton‘s Law of Cooling. We assume that the heat transfer coefficient b for surface elements can vary with different amounts of water content W , defined in Eq. (9). Fig. 5 shows the surface temperature on the branching structure of a tree model.

The radial heat transfer is not modeled explicitly by diffusion equations. Our model decouples the combustion in radial and tangential directions by introducing a pyrolyzing front for the radial direction and surface heat conduction for the tangential. Thus, the rate of combustion in those principal directions is controlled by separate Eqs. (7) and (12).

6.2.2 Air Heat Transfer. The heat transfer from branch surface to air is expressed by the surface heat transfer term sT in Eq. (2). The surface heat transfer term expresses the combustion of flammable gases released by the process of pyrolysis. In our model this complex movement and combustion of gases is approximated by the mass loss dt M determined in Eq. (7). Due to the combustion of flammable gases, the temperature increases proportionally to the mass loss:

where TM is a coefficient expressing the rate at which heat is added to the air by the combustion of gases released by the process of pyrolysis. We do not model the conduction of heat from wood tissue to air. The Laplace-Beltrami operator in Eq. (12) is discretized using the graph Laplacian. We use forward Euler integration for all differential equations in this section.

## 6.3 Structural Deformation
The pyrolysis of wood causes geometric deformations of the branching structure. We capture these effects by exploiting our particle based tree representation (Sec. 4) and the associated surface mesh.

6.3.1 Branch Surface Deformation. Due to the contraction of char, the volume of wood decreases. We model this effect by deforming the surface mesh of the branches (Fig. 6). Vertices affected by pyrolysis are moved along their original normal direction, where the offset can be obtained by the thickness of the char layer Hc, the remaining virgin wood H, and the original thickness of the branch H0:

The strength of the deformation is controlled by the contraction factor κ, which determines the amount the char layer contracts as a result of the pyrolysis (Eq. 5). Since Hoff is always less than H0 (the char layer stops forming when the total mass of the element is pyrolyzed), a vertex never moves further than to its projection point on the branch, which guarantees that the mesh does not degenerate.

6.3.2 Tree Model Deformation. We compute the bending and breaking of branches by computing strains for twisting, stretching,bending, and shearing for each particle [Kugelstadt and Schömer 2016]. A branch breaks when the strain ε exceeds a threshold εmax (ε > εmax ). Then the rod element between the particles is removed and the physics of the detached particles are further simulated. To influence the dynamics of tree motions through the combustion
process, we update the particles according to geometric changes of the model. Moreover, the mass loss caused by the combustion influences the stiffness of a branch, which causes it to bend and break more easily. Gravity is applied as an external force to all particles.

The mass of overhanging branches increases the strain, however,thicker branches are stiffer than thinner branches and therefore experience less strain. We approximate the maximum strain εmax a branch can endure using the following equation:

where d is the thickness of the branch at a particle, M0 denotes the mass at the start of the simulation, M the current mass of the surface elements associated with the particle, and cd a user defined coefficient. Branches that undergo combustion are penalized inversely proportional to their thickness to account for the effects of the mass of the overhanging sub-branches.

6.3.3 Leaves. Similar to branches, leaves burn individually and transfer heat to the surrounding environment. We represent leaves as two-dimensional surfaces with associated mass and area. Leaves are attached to branch particles (Sec. 7), thereby fires can spread from branches to leaves and vice versa. We simulate the mass loss based on Eq. (7) without the char insulation term and the heat diffusion based on Eq. (12). During the combustion process, the area of leaves decreases proportionally to their mass loss analogously to Eq. (14)

6.2 传热

木材的燃烧用树木模型和周围空气之间的双向传热来表示。相对较高的空气温度会升高分支表面温度，当足够高时，会导致质量损失率的增加（等式7）。另一方面，表面元素的质量损失会导致空气温度升高（方程式2，13）。因此，热传递构成了一个正反馈回路，使火焰在我们的模型中传播。

6.2.1 表面传热。通过考虑分支表面上的热扩散以及空气和表面之间的热传导来确定表面元件的温度$T_s$。空气和分支表面（表面元件）之间的热传导与它们之间的温度差成比例。因此，表面温度$T_s$的变化率由下式给出：
$${\partial }_t{T}_{\mathrm{s}}=a{\nabla }^2{T}_{\mathrm{s}}+b\left(T_{\mathrm{a}}-T_{\mathrm{s}}\right)\text{\hspace{1em}(12)}$$

方程右侧最左边的项使用拉普拉斯算子和温度扩散系数$\alpha$描述了分支表面上的热扩散。最右边的项表达了牛顿冷却定律。我们假设表面元素的传热系数b可以随不同的含水量W而变化，如方程（9）所定义。图5显示了树模型分支结构上的表面温度。

径向传热没有通过扩散方程明确建模。我们的模型通过在径向引入热解前沿和在切向引入表面热传导，使燃烧在径向和切向上解耦。因此，这些主要方向上的燃烧速率分别由方程式(7) 和(12)控制。

6.2.2 空气传热。从分支表面到空气的传热由方程式(2)中的表面传热项$s_T$表示。表面传热术语表示热解过程中释放的可燃气体的燃烧。在我们的模型中，气体的这种复杂运动和燃烧近似于方程(7)中确定的质量损失$d_{t}M$。由于可燃气体的燃烧，温度与质量损失成比例增加：
$$s_T=-T_M{d}_{t}M\text{\hspace{1em}(13)}$$
其中$T_M$是表示通过热解过程释放的气体的燃烧将热量添加到空气中的速率的系数。我们没有模拟从木质组织到空气的热传导。方程(12)中的拉普拉斯-贝尔特拉米算子使用拉普拉斯图离散化。在本节中，我们对所有微分方程使用正欧拉积分。

6.3 结构变形

木材的热解导致树枝结构的几何变形。我们通过利用基于粒子的树表示（第4节）和关联的曲面网格来捕捉这些效果。

6.3.1分支表面变形。由于炭的收缩，木材的体积减小。我们通过变形分支的表面网格来模拟这种效果（图6）。受热解影响的顶点沿其原始法线方向移动，其中可以通过炭层$H_c$的厚度、剩余原始木材$H$和分支$H_0$的原始厚度获得偏移：

$$H_{off}=H_0-H-H_c\text{\hspace{1em}(14)}$$

变形强度由收缩因子κ控制，该因子决定热解后炭层收缩的量（方程式5）。由于$H_{o}ff$始终小于$H_0$（当元素的总质量被热解时，炭层停止形成），顶点永远不会移动到分支上的投影点，这保证了网格不会退化。

6.3.2 树模型变形。我们通过计算每个粒子的扭曲、拉伸、弯曲和剪切应变来计算分支的弯曲和断裂[Kugelstadt和Schömer 2016]。当应变$\epsilon$超过阈值${\epsilon }_{\max }$（$\epsilon$>${\epsilon }_{\max }$）时，分支断裂。然后移除颗粒之间的杆元素，并进一步模拟分离颗粒的物理性质。通过燃烧处理影响树木的运动，在此过程中，我们根据模型的几何变化来更新粒子。此外，燃烧造成的质量损失会影响支管的刚度，从而使支管更容易弯曲和断裂。重力作为外力作用于所有粒子。

悬垂树枝的质量增加了应变，然而，较厚的树枝比较薄的树枝更硬，因此承受的应变较小。我们使用以下公式来近似分支可以承受的最大应变${\epsilon }_{\max }$：
$${\epsilon }_{\max }=\frac{c_d}{d}\cdot \left(1-\frac{\sum M_0-\sum M}{\sum M_0}\right)\text{\hspace{1em}(15)}$$

其中d是粒子处分支的厚度，M0表示模拟开始时的质量，M是与粒子相关的表面元素的当前质量，cd是用户定义的系数。经历燃烧的分支与其厚度成反比，以考虑悬垂分支质量的影响。

6.3.3 树叶。与树枝类似，树叶单独燃烧并将热量传递到周围环境。我们将树叶表示为具有相关质量和面积的二维曲面。树叶附着在树枝颗粒上（第7节），因此火灾可以从树枝蔓延到树叶，反之亦然。我们基于方程（7）模拟了质量损失，没有碳绝缘项，基于方程（12）模拟了热扩散。在燃烧过程中，叶片的面积与它们的质量损失成比例地减少，类似于方程式（14）

# 7 IMPLEMENTATION AND RESULTS
Our framework is implemented in C++ using OpenGL 4.2 and CUDA 8.0 on a desktop computer with an Intel Xenon CPU clocked at 3.7 Ghz and 32 GB of RAM. We did not specifically optimize our code and rendered all results shown in the paper and the accompanying video with a NVIDIA Titan X GPU in our framework.

7.0.1 Fire and Smoke. We simulate fire with a grid based fluid solver and render it with a GPU-based raymarching approach to capture the characteristic volumetric features of flames. To maintain interactive frame rates, we compute the involved grids for fire velocity, temperature, and smoke at different resolutions. For most of our renderings we used a fire velocity grid of size 643 and grids of size 1283 to compute temperature and smoke. The size of the fluid domain is chosen to keep dx small, but big enough to avoid effects from boundary conditions. For most of our experiments dx is 5-10 cm. We use a single time step of 30 ms per frame. The final color and opacity of the fire at a given voxel is determined by the temperature value and a transfer function [Nguyen et al. 2002]. We use a two pass Gaussian filter to approximate light scattering around the flames to add the impression of brightness and glow. Finally, we advect point sprites with the fire simulation to mimic the effect of sparks.

7.0.2 Branch and Leaf Geometry. We compute a triangular surface mesh for the branches of tree models based on the hierarchical rod structure defined in Sec. 4. For each chain of rod elements we compute a consecutive and uniformly tessellated mesh based on generalized cylinders. The start and end of a cylinder is defined by the radius and orientation of two subsequent particles. If more than one rod element starts at a particle, the rod element with the closest change of orientation is chosen to be the consecutive element to form a smooth branch, all other rod elements start a new branch with a corresponding mesh.

The surface heat transfer is directly computed on the tree mesh.It is important to note that the computed mesh is not a smooth manifold; meshes of child branches are not connected with the geometry of parent branches. We use an index buffer to store the adjacency information of mesh triangles to simulate heat diffusion.For vertices at the starting end of disconnected child branch meshes,adjacency information is not available. For these vertices, we select the closest vertex of the closest branch as the neighbor. The index buffer is computed as a preprocessing step.

7.0.3 Animation and Rendering. We use Linear Blend Skinning [Le and Deng 2012] to animate the tree mesh according to the positional changes of the branch particles. Each vertex is associated with the two particles of its corresponding rod element. The vertex is animated by computing its position updates based on the weights given by its associated particles. We further pronounce the geometry of branches with displacement mapping and define leaves as points that are rendered as textured quad surfaces in the geome try shader. To animate leaf motion, we associate each leaf with its closest branch particle; positional changes of branch particles cause updates of the corresponding leaf positions. Branch collisions are resolved by particle-particle collision tests through PBD.

We use texture blending on the branch mesh to animate the combustion process. A parameter for charring describes the visual degree of darkening of the surface, which is proportional to the thickness of the char layer. To animate glowing of the surface, we use a glow parameter that is set to 1 when the surface temperature of the surface reaches a threshold; it decays separately from the temperature to retain more control of the visualization.

7.1 Performance and Parameters
We measured the performance of our system for the figures shown in the paper; Tab. 1 shows timings and parameters of our system. All three components (dynamic tree motions, fire simulation, and combustion model) can be efficiently computed at interactive rates. Our method uses 9 user-controlled parameters that allow us to control the physical and biological properties of the combustion process.These parameters can be adjusted and explored interactively in our system. Additionally, we rely on seven physically-motivated parameters in our equations. These parameters are initialized once and remain constant across different simulations and tree models. Tab. 2 lists all parameters, their explanation, ranges, and units.

Our goal is to provide a mathematical model for wood combustion with a minimal set of parameters, while still allowing to model the key properties of burning trees, such as for pyrolysis, moisture, and charring. The introduced parameter space captures a diversity of realistic combustion patterns (Fig. 7, a). For example, the resistance of trees to fire is defined in two ways. First, by modulating the charring parameters, i.e. char contraction factor κ and maximum insulation parametercmin (Fig. 7, b), second, by changing the rate of mass loss using ∆m (Fig. 7, e). The first approach captures different wood charring dynamics, whereas the second allows to express wood combustion dynamics due to, e.g., the lack of wood resins such as in Redwood trees. Thus, a user has control over the rate and extent of flame spread. Further, our method exposes a parameter,called ϵmax, to control the brittleness of branches as part of the wood combustion.

In Fig. 7 (c), a high value for the maximum strain ϵmax has been chosen, leading to frequent branch breaking events and a visibly damaged tree structure after the combustion. One of the most important specifiers of wood combustion in real trees is the water content as it majorly influences both, the flame spread as well as the amount and type of smoke produced. Our model takes this important phenomenon into account by controlling the smoke generation via the moisture parameter W , illustrated in Fig. 7 (d).


7.2 Results
The examples in Figs. 1 and 8 illustrate the capabilities of our framework. Different tree models are set on fire by igniting branches with a fire emitter. Due to the structure of the branches and individual settings for the parameters for each of the models, our method allows to plausibly simulate the combustion of botanical tree models (Fig. 8, a-e). The geometric features of the model affect the combustion process: the sparser branching structure of the deciduous tree model (Fig. 8, f-i) and higher moisture of its leaves hinder the fire to fully propagate through the model.

Fig. 8 and Fig. 9 show the interaction with burning tree models.The user can drag the branches of a model. They sway back and the flames propagate through the tree model. Dynamic motions of branches causes changes in the fire simulation and thereby affect the combustion of the model. Due to the combustion and externally induced stress, branches can bend and break, while they continue to burn. As the combustion progresses, the insulation of char hinders the combustion to proceed (Fig. 9, e, f).

In Fig. 10 we ignited the tree models in different locations. While the flames slowly propagate from the top towards the lower branches (a-c), the model rapidly catches fire when it is initially ignited at the bottom (d-g). We model heat transfer in wood and air, which allows fires to spread without the direct contact of flames and branches Moreover, as we rely on a grid-based fluid solver it is possible to affect our fire simulation and the combustion model with wind. Fig. 11 (right) shows how the flames of a burning tree are extinguished by a wind gust.

Finally, Fig. 14 shows the combustion of a group of trees consisting of five models. A model is set on fire and starts to burn.The combustion propagates across the branching structure and to other models in the scene until all tree models are affected. The combustion slows down when no further fuel is available.

7 实施和结果

我们的框架是在桌面计算机上使用OpenGL 4.2和CUDA 8.0在C++中实现的，该计算机具有Intel Xenon CPU，时钟为3.7 Ghz，RAM为32 GB。我们没有特别优化我们的代码，并在我们的框架中使用NVIDIA Titan X GPU渲染了论文和附带视频中显示的所有结果。

7.0.1 火灾和烟雾。我们使用基于网格的流体解算器模拟火灾，并使用基于GPU的光线行进方法渲染火灾，以捕捉火焰的特征体积特征。为了保持交互式帧速率，我们计算了不同分辨率下的火灾速度、温度和烟雾的相关网格。对于我们的大多数渲染图，我们使用了643大小的火焰速度网格和1283大小的网格来计算温度和烟雾。选择流体域的大小以保持dx小，但足够大以避免边界条件的影响。对于我们的大多数实验，dx为5-10厘米。我们使用每帧30ms的单个时间步长。给定体素处火焰的最终颜色和不透明度由温度值和传递函数决定[Nuyen等人，2002]。我们使用两通高斯滤波器来近似火焰周围的光散射，以增加亮度和辉光的印象。最后，我们将点精灵与火灾模拟相结合，以模拟火花的效果。

7.0.2 枝叶几何结构。我们基于第4节中定义的分层杆结构，为树模型的分支计算三角形曲面网格。对于每个杆元素链，我们基于广义圆柱体计算连续且均匀细分的网格。圆柱体的起点和终点由两个后续粒子的半径和方向定义。如果多于一个杆元素开始于一个粒子，则选择方向变化最接近的杆元件作为连续元件以形成平滑分支，所有其他杆元件开始具有相应网格的新分支。

在树网格上直接计算表面传热。需要注意的是，计算的网格不是平滑的流形；子分支的网格不与父分支的几何体连接。我们使用索引缓冲区来存储网格三角形的邻接信息，以模拟热扩散。对于断开连接的子分支网格起始端的顶点，相邻信息不可用。对于这些顶点，我们选择最近分支的最近顶点作为邻居。作为预处理步骤计算索引缓冲区。

7.0.3 动画和渲染。我们使用线性混合蒙皮[Le和Deng 2012]根据分支粒子的位置变化为树网格设置动画。每个顶点都与其对应的杆元素的两个粒子相关联。顶点通过基于其关联粒子给出的权重计算其位置更新来设置动画。我们进一步使用置换贴图来发音分支的几何体，并将树叶定义为在geometry着色器中渲染为纹理四边形曲面的点。为了设置树叶运动的动画，我们将每个树叶与其最近的分支粒子相关联；分支粒子的位置变化导致相应叶片位置的更新。分支碰撞通过粒子-粒子碰撞测试通过PBD解决。

我们在分支网格上使用纹理混合来设置燃烧过程的动画。炭化参数描述表面的视觉变暗程度，其与炭化层的厚度成比例。要设置曲面发光的动画，我们使用当曲面的表面温度达到阈值时设置为1的发光参数；它与温度分开衰减，以保持对可视化的更多控制。

7.1 性能和参数

我们根据论文中所示的数字测量了系统的性能；表1显示了系统的时序和参数。所有三个组成部分（动态树木运动、火灾模拟和燃烧模型）都可以以交互速率进行有效计算。我们的方法使用9个用户控制的参数，这些参数允许我们控制燃烧过程的物理和生物特性。这些参数可以在我们的系统中进行交互调整和探索。此外，我们依赖于方程中的七个物理激励参数。这些参数初始化一次，并在不同的模拟和树模型中保持不变。表2列出了所有参数及其说明、范围和单位。

我们的目标是为木材燃烧提供一个具有最小参数集的数学模型，同时仍然允许对燃烧树木的关键特性进行建模，例如热解、水分和炭化。引入的参数空间捕捉了多种真实的燃烧模式（图7，a）。例如，树木对火的抵抗力有两种定义。首先，通过调节炭化参数，即炭化收缩因子κ和最大绝热参数cmin（图7，b），其次，通过使用∆m（图7，e）。第一种方法捕捉不同的木材炭化动力学，而第二种方法允许表达木材燃烧动力学，例如，由于缺少木材树脂，如红木树。因此，用户可以控制火焰传播的速度和范围。此外，我们的方法公开了一个参数，称为εmax，以控制树枝的脆性，作为木材燃烧的一部分。

在图7（c）中，选择了最大应变εmax的高值，导致频繁的树枝断裂事件和燃烧后明显受损的树木结构。真实树木中木材燃烧的最重要指标之一是含水量，因为它主要影响火焰传播以及产生的烟雾量和类型。我们的模型通过通过湿度参数W控制烟雾产生来考虑这一重要现象，如图7（d）所示。

7.2 结果

图1和图8说明了我们框架的功能。不同的树模型通过用火焰发射器点燃树枝来点燃。由于每个模型的分支结构和参数的单独设置，我们的方法允许模拟植物树模型的燃烧（图8，a-e）。该模型的几何特征影响燃烧过程：落叶树模型的稀疏分支结构（图8，f-i）及其叶片的较高湿度阻碍了火焰在模型中充分传播。

图8和图9显示了与燃烧树模型的交互作用。用户可以拖动模型的分支。它们向后摆动，火焰通过树模型传播。树枝的动态运动会导致火灾模拟发生变化，从而影响模型的燃烧。由于燃烧和外部引起的应力，树枝在继续燃烧时会弯曲和断裂。随着燃烧的进行，炭的绝缘阻碍了燃烧的进行（图9，e，f）。


在图10中，我们在不同位置点燃了树模型。当火焰从顶部向下部分支（a-c）缓慢传播时，当最初在底部点燃时，模型会迅速起火（d-g）。我们对木材和空气中的热传递进行建模，这允许火焰在没有火焰和树枝直接接触的情况下蔓延。此外，由于我们依赖基于网格的流体解算器，因此有可能影响我们的火灾模拟和有风的燃烧模型。图11（右）显示了一棵燃烧的树的火焰是如何被一阵风扑灭的。

最后，图14显示了由五个模型组成的一组树木的燃烧。一个模型被点燃并开始燃烧。燃烧通过分支结构传播到场景中的其他模型，直到所有树模型都受到影响。当没有其他燃料可用时，燃烧速度减慢。

8 EVALUATION, DISCUSSION, AND LIMITATIONS
The pyrolysis of trees is a complex natural phenomenon that is difficult to observe and to sufficiently quantify. While many studies investigate the combustion of different tree species and the pyrolysis of woody materials in confined settings, such as in laboratories,there exists no dataset that provides a baseline for the combustion of geometric models of botanical trees.

8.1 Evaluation
To evaluate our computational model we provide visual and quantitative comparisons to experimental data of real wood combustion and compare our model with a simplified combustion model for branching structures.

8.1.1 Comparison to Real Wood Combustion. We conducted experiments where we burnt two wood samples fixated in a vertical and a skewed orientation. A flame was produced by a Bunsen burner at the lower ends of the sticks. Fig. 12 shows the setup of the experiment. We observed that pyrolysis progresses more rapidly when the sample is oriented vertically compared to when it is fixated in a skewed orientation. As the flames follow the inverse gravity direction, they quickly cover the entire stick, which causes a rapid pyrolysis of the wood (Fig. 12, top, left).

We quantify this observation by measuring the time of the combustion process, starting with the ignition of the material until no flame is visible anymore. An important quantitative measure of wood combustion in material sciences is the mass loss rate under a constant heat flux over time. Fig. 13 (left) shows that the mass loss is increasing very fast in the first few seconds of the experiment until reaching a well pronounced tipping point (red line). This tipping point coincides with the rapid formation of the char layer, which has an insulating effect for wood (Sec. 6.1.3) and contributes to a reduction of the mass loss rate [Weng and Fan 2007].

We compare these observations with our model by defining virtual wood samples in our particle-based representation (Sec. 4). The samples and a fire emitter are positioned in a scene to match the laboratory setup. We manually set up the parameters for this simulation to represent dry wood. Fig. 12 (bottom) shows that our model produces similar visual results for both scene setups. Please note,that our current model cannot handle the bending of wood samples due to more advanced effects such as the cracking of charcoal. Also,we do not simulate the release of resins responsible for the typical flickering of flames. The pyrolysis of the vertical sample took 87 sec. in the burning chamber and 81 sec. in our simulation. For the skewed sample we measured 165 sec. (burning chamber) and 235 sec. (simulation). Furthermore, the graph in Fig. 13 (left) shows that the mass loss curve in our model has similar characteristics compared to the mass loss curve of the real wood sample. The simulated mass loss curve (green line) expresses the exponential increase in the beginning, the tipping point, and the subsequent decrease in mass loss. A simpler model, without taking the effects of charring into account, but combusting at a similar rate (blue line), is missing the characteristic tipping point.


8.1.2 Simplified Model for Wood Combustion. Our mathematical model captures the essential phenomena in wood combustion, such as the formation of a pyrolyzing front and wood charring. It is necessary to capture these phenomena to obtain a realistic simulation of wood combustion. A naïve model of solid combustion, where the insulating effects of charring are ignored and mass loss is proportional to the total mass of particles, yields unrealistic combustion patterns.In Fig. 6, we illustrate a comparison of a naïve approach with our method. In the simple model, thick and thin branch segments combust at similar rates, leading to unreasonable changes of branch thickness, whereas, in our approach thicker branches realistically combust proportionally much slower compared to thinner branches.

Furthermore, in our model, the rate of mass loss can be controlled by the parameter cmin representing the insulation effect of a char layer. Fig. 13 (right) shows the mass loss rate as a result for different values ofcmin, where low values indicate a high insulation and high values less insulation.

8.2 Discussion and Limitations
Our focus was on defining a model to interactively explore combustion for botanical tree models. To compute this complex phenomenon in real-time, we carefully abstract and simplify components in our model. This includes the simulation of the temperature gradient inside the wood, the representation of char contraction, the fluid boundary conditions, and the explicit modeling of mass transfer from the branch surface to the surrounding air.

We do not use trees as a boundary condition for our fluid simulation, but instead use the heat and smoke generation terms sT (Eq. 2) and sS , (Eq. 3), provided by our combustion model. However,including the tree geometry as a boundary condition would allow to more precisely model the convection of temperature currents around branches. A more accurate mass transfer would enable us to introduce a time delay between the heat transfer from air to surface and vice versa, which would allow a more realistic simulation of temperature changes on the branches. Moreover, we abstained from calculating an internal temperature gradient for branches (e.g. with a volumetric method) and instead express the movement of a pyrolyzing front associated with individual surface elements directly.Furthermore, we assume that the char contraction is constant for different types of wood. However, among other factors, the char contraction factor is known to be dependent on temperature and a more plausible hypothesis would yield more accurate results. Dynamic changes of the formation of charcoal could then be used to include the procedural formation of char cracking patterns or the bending of wood as part of our surface deformation approach.

Existing models for the combustion of trees and wood either focus on computationally evolved methods to analyze the physical processes of pyrolysis in detail [Thi et al. 2016] or neglected pyrolysis altogether to study the effects of wood combustion at a higher level of abstraction, e.g. for fire spread in forests [Seidl et al. 2012]. In material sciences, efforts are made to improve structural design by taking realistic fire scenarios into account. However, current methods, e.g. based on finite elements, are too slow to handle complex geometry. Here, real-time approaches can provide a way towards generating new insights. Methods that consider the tree geometry can significantly improve simulations for forest fire modeling.

8 评估、讨论和限制

树木的热解是一种复杂的自然现象，难以观察和充分量化。尽管许多研究调查了不同树种的燃烧以及在实验室等密闭环境中木质材料的热解，但没有数据集为植物树几何模型的燃烧提供基线。

8.1评估

为了评估我们的计算模型，我们提供了真实木材燃烧实验数据的视觉和定量比较，并将我们的模型与分支结构的简化燃烧模型进行了比较。

8.1.1与真实木材燃烧的比较。我们进行了实验，燃烧了两个垂直和倾斜方向固定的木材样品。燃烧棒下端的本生燃烧器产生了火焰。图12显示了实验的设置。我们观察到，与以倾斜方向固定样品相比，当样品垂直定向时，热解进展更快。当火焰沿着相反的重力方向移动时，它们会迅速覆盖整个木棒，从而导致木材快速热解（图12，顶部，左侧）。

我们通过测量燃烧过程的时间来量化这一观察结果，从点燃材料开始，直到再也看不到火焰。在材料科学中，木材燃烧的一个重要的定量指标是随着时间的推移在恒定热通量下的质量损失率。图13（左）显示，在实验的前几秒，质量损失快速增加，直到达到明显的临界点（红线）。该临界点与炭层的快速形成相吻合，炭层对木材具有绝缘作用（第6.1.3节），有助于降低质量损失率[Weng和Fan 2007]。

我们通过在基于粒子的表示中定义虚拟木材样本，将这些观察结果与我们的模型进行比较（第4节）。样本和火发射器放置在场景中，以匹配实验室设置。我们手动设置此模拟的参数以表示干燥木材。图12（底部）显示了我们的模型为两种场景设置产生了类似的视觉结果。请注意，由于木炭开裂等更先进的影响，我们当前的模型无法处理木材样品的弯曲。此外，我们没有模拟导致典型火焰闪烁的树脂释放。垂直样品的热解在燃烧室中耗时87秒，在我们的模拟中耗时81秒。对于倾斜样品，我们测量了165秒（燃烧室）和235秒（模拟）。此外，图13（左）中的图表显示，与真实木材样品的质量损失曲线相比，我们模型中的质量损失特性相似。模拟质量损失曲线（绿线）表示开始时的指数增加、临界点以及随后的质量损失减少。一个更简单的模型，没有考虑碳化的影响，但以类似的速度燃烧（蓝线），缺少了特征临界点。

8.1.2木材燃烧的简化模型。我们的数学模型捕捉了木材燃烧中的基本现象，如热解前缘的形成和木材烧焦。有必要捕捉这些现象，以获得木材燃烧的真实模拟。固体燃烧的天真模型，其中忽略了碳化的绝缘效应，质量损失与颗粒的总质量成正比，产生了不切实际的燃烧模式。在图6中，我们说明了天真方法与我们的方法的比较。在简单的模型中，厚的和薄的分支段以相似的速率燃烧，导致分支厚度的不合理变化，而在我们的方法中，较厚的分支实际上比较薄的分支按比例燃烧得慢得多。

此外，在我们的模型中，质量损失率可以由表示炭层绝缘效果的参数cmin控制。图13（右）显示了不同最小值的质量损失率，其中低值表示高绝缘，高值表示低绝缘。

8.2讨论和限制

我们的重点是定义一个模型，以交互方式探索植物树模型的燃烧。为了实时计算这一复杂现象，我们仔细地抽象和简化了模型中的组件。这包括木材内部温度梯度的模拟、炭收缩的表示、流体边界条件以及从树枝表面到周围空气的传质的显式建模。

我们不使用树作为流体模拟的边界条件，而是使用燃烧模型提供的热量和烟雾生成项$s_T$（方程式2）和$s_S$（方程式3）。然而，将树的几何结构作为边界条件将允许更精确地模拟树枝周围温度流的对流。更精确的质量传递将使我们能够在从空气到表面的热传递之间引入时间延迟，反之亦然，这将允许对分支上的温度变化进行更真实的模拟。此外，我们没有计算分支的内部温度梯度（例如，使用体积法），而是直接表示与单个表面元素相关的热解前沿的运动。此外，我们假设不同类型的木材的炭收缩是恒定的。然而，除其他因素外，已知炭收缩因子取决于温度，更合理的假设将产生更准确的结果。然后，木炭形成的动态变化可用于包括木炭开裂图案的程序形成或木材弯曲，作为我们表面变形方法的一部分。

现有的树木和木材燃烧模型要么侧重于计算进化方法，以详细分析热解的物理过程[Thi等人，2016]，要么完全忽略热解，以在更高的抽象层次上研究木材燃烧的影响，例如森林火灾蔓延[Seidl等人，2012]。在材料科学中，通过考虑实际火灾场景，努力改进结构设计。然而，当前的方法，例如基于有限元的方法，对于处理复杂的几何结构来说太慢了。在这里，实时方法可以提供一种产生新见解的方法。考虑树木几何形状的方法可以显著改善森林火灾模型的模拟。

# 9 CONCLUSION AND FUTURE WORK
We have presented a computationally efficient method for the combustion of botanical tree models. We simulate realistic fires and provide a novel method for the pyrolysis of wood, that allows us to express the mass loss of virgin wood, the insulating effect of char, and the temperature changes due to moisture evaporation.Our approach is based on a biologically-plausible mathematical formulation and supports modeling a variety of effects, including the deformation of lateral branch geometry, the bending and breaking of branches as well as the plausible animation of wood combustion. Our method is the first attempt to jointly simulate botanical tree models and wood combustion in a biologically plausible way.Thereby, our model can be viewed as a link between low- and high level methods with the potential to provide insights in a variety of fields, including material science and forestry research.

There are a number of open issues that appear as avenues for future work. It would be interesting to explore through further analysis, if our method can help to uncover how geometric, physical,and biological properties of trees contribute to fire spread across different plants. Moreover, we are currently approximating different phenomena, such as heat conduction on the surface and char contraction, as constant values in our mathematical framework. Increasing the complexity of these components would allow to improve the realism of combustion for botanical tree models.

9 结论和未来工作

我们提出了一种计算有效的植物树模型燃烧方法。我们模拟了真实的火灾，并为木材的热解提供了一种新的方法，这使我们能够表达原始木材的质量损失、木炭的隔热效果以及水分蒸发引起的温度变化。我们的方法基于生物学上合理的数学公式，并支持对各种影响进行建模，包括侧枝几何结构的变形、树枝的弯曲和断裂以及木材燃烧的合理动画。我们的方法是首次尝试以生物学上合理的方式联合模拟植物树模型和木材燃烧。因此，我们的模型可以被视为低水平和高水平方法之间的联系，有可能在各种领域提供见解，包括材料科学和林业研究。

有一些公开的问题似乎是未来工作的途径。如果我们的方法能够帮助揭示树木的几何、物理和生物特性如何导致火灾在不同植物之间蔓延，那么通过进一步的分析进行探索将是非常有趣的。此外，我们目前正在将不同的现象，例如表面上的热传导和炭收缩，近似为数学框架中的常数值。增加这些组件的复杂性可以提高植物树模型燃烧的真实性。