Abstract
The present paper describes a shape optimization procedure for designing running shoes, focusing on two mechanical properties, namely, the shock absorption and the stability keeping the right posture. These properties are evaluated from two deformations of a sole at characteristic timings during running motion. We define approximate planes for the deformations of sole’s upper boundary by least squares method. Using the planes, we choose the tilt angle in the shoe width direction at the mid stance phase of running motion as an objective function representing the stability, and the sunk amount at the contact phase of running motion as a constraint function representing the shock absorption. We assume that the sole is a bonded structure of soft and hard hyper-elastic materials, and the bonding and side boundaries are variable. In this study, we apply the formulation of nonparametric shape optimization to the sole considering finite deformation and contact condition of the bottom of the sole with the ground. Shape derivatives of the cost (objective and constraint) functions are obtained using the adjoint method. The H1 gradient method using these shape derivatives is applied as an iterative algorithm. To solve this optimization problem, we developed a computer program combined with some commercial softwares. The validity of the optimization method is confirmed by numerical examples.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Running is one of the most popular sports in the world. However, it has been reported that the risk of injuries to lower extremities is greater than other sports, because repeated loads are applied to lower extremities during running (Matheson et al. 1987). For this reason, running shoes should not only enhance the runner’s performance but also prevent injuries during running.
Shoes have various requirement properties, such as cushioning property, shoe stability, grip property, and breathability (Cavanagh 1980; Nishiwaki 2008). Especially in the procedure designing sole of shoes, we often focus on the cushioning property and the shoe stability. The cushioning property means absorption of the impact from the ground at the contact phase. It is evaluated using time derivative of vertical ground reaction force (for instance Gard and Konz (2004), Clarke et al. (1983), Nigg et al. (1987, 1988), and Nigg (1980)). Meanwhile, the shoe stability means suppression of excessive foot joint motions called as pronation at the mid stance phase. It is evaluated using the angle between heel and lower extremity (for instance Nigg (1980), Woensel and Cavanagh (1992), Areblad et al. (1990), and Stacoff et al. (1992)). A sole that is made of soft material has good cushioning property but may not have good shoe stability. In the case of hard material, the contrary may be true.
Apparently, the sole which consists of only one material cannot provide both good properties. Some sole designs combining different hardness materials to improve both properties were reported (Nakabe and Nishiwaki 2002; Oriwol et al. 2011). These soles have a part made of hard material on the inside from the heel to the mid foot to increase rigidity of the sole. The hard part can prevent the heel to tilt towards the heel’s region and improve the stability, while it may reduce cushioning property. However, the shape of the hard material in the heel was rectangular. We have not found any cases in which both the shoe stability and the cushioning property are improved by designing the shape of the material boundary using optimization method.
Some nonlinearities must be considered to predict mechanical properties of running shoes. It is expected that 20% or more of strain occurs in the sole during running since the ground reaction force in the vertical direction is two to three times of runner’s weight (Gard and Konz 2004; Clarke et al. 1983, Nigg et al. 1987, 1988, Nigg 1980). Because of this, one has to consider geometrical nonlinearity in predicting mechanical properties of running shoes. The soles of running shoes should therefore be made of multiple materials with strong nonlinearity such as resin foams, resins, and rubbers to satisfy the various required properties. For example, it is known that resin foams have complicated behaviors under compressive load (Gibson and Ashby 1980; Mills 2007) and must be modeled as some hyper-elastic materials. In addition to this, contact condition must be taken into account because the bottom surface of the sole contacts with the ground at different parts from the moment of the first contact until complete separation from ground.
Many methods of analyzing design sensitivity of nonlinear problems have been researched over the years and applied to a variety of design optimization problems (Kaneko and Majer 1981; Ryu et al. 1985; Cardoso and Arora 1988; Tsay and Arora 1990; Vidal et al. 1991; Vidal and Haber 1993; Hisada 1995; Yamazaki and Shibuya 1998; Yuge and Kikuchi 1995). However, these methods were applicable only to parametric or sizing optimization problems. Focusing on nonparametric shape optimization methods, Ihara et al. (1999) solved a shape optimization problem minimizing the external work of an elasto-plastic body. They evaluated the shape derivative of the cost function by the adjoint method and obtained the search vector of domain variation by the traction method (Azegami and Wu 1996; Azegami and Takeuchi 2006) (H1 gradient method for domain variation (Azegami 2016, 2017). Kim et al. (2000) presented a sensitivity analysis for shape optimization problems considering the infinitesimal elasto-plasticity with a frictional contact condition. In their work, the direct differentiation method was used to compute the displacement sensitivity, and the sensitivities of various performance measures were computed from the displacement sensitivity. In updating the shape, the authors used the so-called isoparametric mapping method (Choi and Chang 1994). Furthermore, Tanaka and Noguchi (2004) presented a shape optimization method similar to the traction method but described by a discrete form, and applied to structural designs with strong nonlinearity such as a flexible micromanipulator made of hyper-elastic material. Meanwhile, Iwai et al. (2010) presented a numerical solution to shape optimization problems of contacting elastic bodies for controlling contact pressure. They used an error norm of the contact pressure to a desired distribution as an objective cost function and evaluated its shape derivative by the adjoint method and reshaped by the traction method. In these aforementioned works, the studies conducted focus on basic problems, though they include geometrical and material nonlinearities.
In this study, we propose a shape optimization method with respect to the desired shoe stability and cushioning property for soles of running shoes. We define the indices to evaluate both properties and then formulate a shape optimization problem using the indices as cost functions. The shape derivatives of cost functions are derived theoretically. Using the shape derivatives, shape optimization can be performed based on the standard procedure of H1 gradient method (Azegami 2016).
In the following sections, we use the notation \(W^{s, \bar {p}} \left (\varOmega ; \mathbb {R} \right )\) to represent the Sobolev space for the set of functions defined in Ω that corresponds to a value of \(\mathbb {R}\) and is \(s \in \left [ 0, \infty \right ]\) times differentiable and \(\bar {p} \in \left [ 1, \infty \right ]\)-th-order Lebesgue integrable. Furthermore, \(L^{\bar {p}} \left (\varOmega ; \mathbb {R} \right )\) and \(H^{s} \left (\varOmega ; \mathbb {R} \right )\) are denoted by \(W^{0, \bar {p}} \left (\varOmega ; \mathbb {R} \right )\) and \(W^{s, 2} \left (\varOmega ; \mathbb {R} \right )\), respectively. In addition, the notation \(C^{0, \sigma } \left (\varOmega ; \mathbb {R} \right )\) is used to represent the Hölder space with a Hölder index \(\sigma \in \left (0, 1 \right ]\). In particular, \(C^{0, 1} \left (\varOmega ; \mathbb {R} \right )\) is called the Lipschitz space. With respect to a reflexive Sobolev space X, we denote its dual space by \(X^{\prime }\) and the dual product of \(\left (x, y \right ) \in X \times X^{\prime }\) by \(\left \langle x, y \right \rangle \). Specifically, \(f^{\prime } \left (\boldsymbol {x} \right ) \left [ \boldsymbol {y} \right ]\) represents the Fréchet derivative \(\left \langle f^{\prime } \left (\boldsymbol {x} \right ), \boldsymbol {y} \right \rangle \) of \(f : X \to \mathbb {R}\) at x ∈ X with respect to an arbitrary variation y ∈ X. Additionally, \(f_{\boldsymbol {x}} \left (\boldsymbol {x}, \boldsymbol {y} \right ) \left [ \boldsymbol {z} \right ]\) represents the Fréchet partial derivative. The notation ∀ means the word “for all”, and A ⋅B represents the scalar product \({\sum }_{\left (i, j \right ) \in \left \{ 1, \ldots , m \right \}^{2}} a_{ij} b_{ij}\) with respect to \(\boldsymbol {A}=\left (a_{ij} \right )_{ij}\), \(\boldsymbol {B}=\left (b_{ij} \right )_{ij} \in \mathbb {R}^{m \times m}\).
2 Sole model
Let us consider a conceptual sole model as depicted in Fig. 1. We assume that a sole of running shoes is composed of multiple hyper-elastic bodies and contacts with the ground. In this paper, Ω10 and Ω20 denote three-dimensional bounded open domains of hyper-elastic bodies initially made of soft and hard materials for the sole, Ω30 represents a domain of hyper-elastic body for the ground, and those that do not overlap with each other. Let Ω0 denote \(\bigcup _{i \in \left \{ 1, 2, 3 \right \}} {\varOmega }_{i 0}\). The domains in Fig. 1 deformed by a domain variation, which mapping is denoted as \(\boldsymbol {i} + \boldsymbol {\phi }: {\varOmega }_{0} \to \varOmega \left (\boldsymbol {\phi } \right ) = \bigcup _{i \in \left \{ 1, 2, 3 \right \}} {\varOmega }_{i} \left (\boldsymbol {\phi } \right )\) (i is the identity mapping), and by finite hyper-elastic deformation generated by the map \(\boldsymbol {i} + \boldsymbol {u}: \varOmega \left (\boldsymbol {\phi } \right ) = \varOmega \left (\boldsymbol {\phi }, \boldsymbol {0} \right ) \to \varOmega \left (\boldsymbol {\phi }, \boldsymbol {u} \right ) = \bigcup _{i \in \left \{ 1, 2, 3 \right \}} {\varOmega }_{i} \left (\boldsymbol {\phi }, \boldsymbol {u} \right )\). The precise definitions will be introduced below.
2.1 Initial domains and boundary conditions
For the initial domains, we assume that the boundaries ∂Ωi0 (\(i \in \left \{ 1, 2, 3 \right \}\)) of Ωi0 are at least Lipschitz continuous. The domains Ω10 and Ω20 are bonded on the boundary Γ120 = ∂Ω10 ∩ ∂Ω20. The domains Ω20 and Ω30 are joined by the boundary \({\varGamma }_{\mathrm {M} 0} \cap {\varGamma }_{\mathrm {M}^{\ast } 0}\), where ΓM0 and \({\varGamma }_{\mathrm {M}^{\ast } 0}\) denote master and slave boundaries having possibility to contact on \(\partial {\varOmega }_{2 0} \setminus \bar {\varGamma }_{12 0}\) and ∂Ω30, respectively. The boundary ΓD0 on \(\partial {\varOmega }_{3 0} \setminus \bar {\varGamma }_{\mathrm {M}^{\ast } 0}\) is a Dirichlet boundary on which the hyper-elastic deformation is fixed. We use the notation \({\varGamma }_{\mathrm {N} 0} = \bigcup _{i \in \left \{ 1, 2, 3 \right \}} \partial {\varOmega }_{i 0} \setminus \left (\bar {\varGamma }_{\mathrm {D} 0} \cup \bar {\varGamma }_{12 0} \right )\) for a Neumann boundary and assume that a traction force pN is applied on \({\varGamma }_{p 0} \subset \partial {\varOmega }_{1 0} \setminus \bar {\varGamma }_{12 0} \subset {\varGamma }_{\mathrm {N} 0}\) and varies with the boundary measure during domain deformation, whose definition will be given later. Moreover, \({\varGamma }_{\mathrm {U} 0} \subset \partial {\varOmega }_{1 0} \setminus \bar {\varGamma }_{12 0}\), which includes Γp0, represents the boundary to observe the deformation of sole, which will be used to define cost functions.
2.2 Domain variations
In this study, we assume that Ω0 is a variable domain. As previously stated, the varied domain is defined as
where ϕ represents the displacement in the domain variation. Similarly, with respect to an initial domain or boundary (⋅)0, \((\cdot ) \left (\boldsymbol {\phi } \right )\) represents \(\left \{ \left . \left (\boldsymbol {i} + \boldsymbol {\phi } \right ) \left (\boldsymbol {x} \right ) \right | \boldsymbol {x} \in (\cdot )_{0} \right \}\).
When the design variable ϕ is selected as above, the domain of the solution to a state determination problem (hyper-elastic deformation problem) varies with the domain variation. Such a situation makes it difficult to apply a general formulation of function optimization problem. Hence, we will expand the domain of ϕ from Ω0 to \(\mathbb {R}^{3}\) and assume \(\boldsymbol {\phi } : \mathbb {R}^{3} \to \mathbb {R}^{3}\). Furthermore, since we will be considering the gradient method on a Hilbert space later, a linear space and an admissible set for ϕ are defined as
In the definition of X, the boundary conditions for domain variation were added from the situation of the present study. The additional condition for \(\mathcal {D}\) was added to guarantee that \(\varOmega \left (\boldsymbol {\phi } \right )\) has Lipschitz regularity.
3 Hyper-elastic deformation problem
In the shape optimization problem formulated later, the solution \(\boldsymbol {u} : \varOmega \left (\boldsymbol {\phi } \right ) \to \mathbb {R}^{3}\) of hyper-elastic deformation problem will be used in cost functions. In this section, we will formulate this problem according to a standard procedure for hyper-elastic continuum.
We define a linear space and an admissible set for u as
for qR > 3. The additional condition for \(\mathcal {S}\) was added to guarantee the domain variation obtained by the H1 gradient method introduced later being in \(\mathcal {D}\).
As explained in Section 2, we consider that the traction pN acting on Γp0 deforms \(\varOmega \left (\boldsymbol {\phi } \right ) = \varOmega \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\) as
Similarly, with respect to any other domain or boundary \((\cdot ) \left (\boldsymbol {\phi } \right )\), \(\left \{ \left . \left (\boldsymbol {i} + \boldsymbol {u} \right ) \left (\boldsymbol {x} \right ) \right | \boldsymbol {x} \in (\cdot ) \left (\boldsymbol {\phi } \right ) \right \}\) is denoted as \((\cdot ) \left (\boldsymbol {\phi }, \boldsymbol {u} \right )\).
On the boundary ΓM0, having the possibility to contact with ΓS0, we define the shortest vector from \(\boldsymbol {x} \in {\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) to \({\varGamma }_{\mathrm {M}^{\ast }} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) by \(\boldsymbol {d} \left (\boldsymbol {u} \right ): {\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right ) \to \mathbb {R}^{3}\) and a penetration distance by
where \(\boldsymbol {\nu } \left (\boldsymbol {u} \right )\) is the outward unit normal vector on \({\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\). We introduce a Lagrange multiplier \(p : {\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right ) \to \mathbb {R}\) to the nonpenetrating condition \(g \left (\boldsymbol {u} \right ) \le 0 \). The physical meaning of p ≥ 0 is the absolute value of contact pressure. For p, a linear space and an admissible set are defined as
A strain used in hyper-elastic deformation problem is defined according to the standard procedure. With respect to the mapping y = i + u : Ω0 →Ω, let the deformation gradient tensor be
and the Green-Lagrange strain be
where I denotes the third-order unit matrix. \(\boldsymbol {E}_{\mathrm {L}} \left (\boldsymbol {u} \right )\) and \(\boldsymbol {E}_{\mathrm {B}} \left (\boldsymbol {u}, \boldsymbol {v} \right )\) are defined as
The definition of constitutive equation for hyper-elastic material is started by assuming the existence of a nonlinear elastic potential \(\pi : \mathbb {R}^{3 \times 3} \to \mathbb {R}\) which gives the second Piola-Kirchhoff stress tensor as
In this paper, we will use the Neo-Hookean model and the hyper foam model (Dassault Systèmes 2018), in which π are given as
respectively. Here, ei, ki, li, and μi denote material parameters. nH represents the order for the hyper foam model. The first and third invariants \(i_{1} \left (\boldsymbol {u} \right )\) and \(i_{3} \left (\boldsymbol {u} \right )\) are defined by
where \(m_{1} \left (\boldsymbol {u}\right ), m_{2}\left (\boldsymbol {u}\right )\), and \(m_{3}\left (\boldsymbol {u}\right )\) are the principal values of the right Cauchy-Green deformation tensor defined by
The first Piola-Kirchhoff stress \(\boldsymbol {\Pi } \left (\boldsymbol {u} \right )\) and Cauchy stress \(\boldsymbol {\Sigma } \left (\boldsymbol {u} \right )\) can be obtained by \(\boldsymbol {S} \left (\boldsymbol {u} \right )\) as
where \(\omega \left (\boldsymbol {u} \right )\) represents \(\det \left | \boldsymbol {F} \left (\boldsymbol {u} \right ) \right |\).
Based on the definitions above, we formulate the hyper-elastic deformation problem of a sole including contact as follows.
Problem 1 (Hyper-elastic deformation)
For \(\boldsymbol {\phi } \in \mathcal {D}\) and a given pN having proper regularity, find \(\left (\boldsymbol {u}, p \right ) \in \mathcal {S} \times \mathcal {Q}\) such that
Equation (15) gives the KKT (Karush-Kuhn-Tucker) conditions for the contact on \({\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\). \(\boldsymbol {p}_{\mathrm {N}} \left (\boldsymbol {u} \right )\) in (11) is defined as varying with the boundary measure in a domain variation, which is defined by
where dγ and \(\mathrm {d}\gamma \left (\boldsymbol {u} \right )\) denote the respective infinitesimal boundary measures before and after deformations.
For later use, we define the Lagrange function with respect to Problem 1 as
Here, v ∈ U was introduced as the Lagrange multiplier and ν1 is the outward unit normal to Ω10 on \({\varGamma }_{12} \left (\boldsymbol {\phi }, \boldsymbol {u} \right )\) (see Fig. 1). To obtain (17), we multiplied (10) by v, integrated it over \(\varOmega \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\), and used (8) and the boundary conditions. In this process, the notations
were used. The operator \(\boldsymbol {E}^{\prime } \left (\boldsymbol {u} \right ) \left [ \boldsymbol {v} \right ]\) denotes \({\sum }_{i \in \left \{ 1, 2, 3 \right \}}\left (\partial \boldsymbol {E} /\right .\)\( \left .\partial u_{i} \right ) v_{i}\). Moreover, ui and vi (\(i \in \left \{ 1, 2, \mathrm {M}, \mathrm {M}^{\ast } \right \}\)) denote the vectors u and v in \({\varOmega }_{i} \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\) or on \({\varGamma }_{i} \left (\boldsymbol {0}, \boldsymbol {u} \right )\), respectively. In the right-hand side of (17), the integral on the internal boundary \({\varGamma }_{12} \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\) was added to later evaluate the shape derivative on the boundary. Moreover, \(g^{\prime } \left (\boldsymbol {u} \right ) \left [ \boldsymbol {v}_{\mathrm {M}} \left (\boldsymbol {u} \right ) - \boldsymbol {v}_{\mathrm {M}^{\ast }} \left (\boldsymbol {u} \right ) \right ] = \boldsymbol {\nu } \left (\boldsymbol {u} \right ) \cdot \left (\boldsymbol {v}_{\mathrm {M}} \left (\boldsymbol {u} \right ) - \boldsymbol {v}_{\mathrm {M}^{\ast }} \left (\boldsymbol {u} \right ) \right )\) was used in the last term. Using \({\mathcal{L}}_{\mathrm {P}}\), the weak form of Problem 1 can be written as
combined with (15). This condition can be replaced by a variational inequality as follows. When \({\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right ) \) and \({\varGamma }_{\mathrm {M}^{\ast }} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) contact-impact one another, the conditions p > 0 and \(g \left (\boldsymbol {u} \right ) = 0\) hold. To satisfy the inequality \(g \left (\boldsymbol {u} \right ) \le 0\), we require the condition
to hold. Therefore, defining
we can rewrite (18) combined with (15) to obtain the variational inequality given by
4 Approximate planes for sole deformation
The objective of this paper is to improve the properties of cushioning and shoe stability. In this section, we define how to evaluate those properties. To do so, we introduce approximate planes of sole’s deformations and show the way to calculate them.
4.1 Definition of approximate planes
Figure 2 shows a solemodel, coordinate system \(\left (x_{1}, x_{2}, x_{3} \right )\)\(\in \mathbb {R}^{3}\), and one of the approximate planes of deformation at time tr with cost functions fr (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) which will be defined later.
Before showing the definition of the planes, we will explain conventional methods to evaluate the shoe stability and cushioning property. Shoe stability is evaluated by the angle between the heel and lower extremity or between the heel and the ground at the time of minimum of ground reaction force in the foot length direction. Cushioning property, on the other hand, is evaluated using the derivative of the vertical ground reaction force at the time of the first peak of vertical ground reaction force. According to experimental researches, it is thought that these values depend on the deformation of the sole. In particular, the displacement in x3-direction on top surface of the sole is dominant.
Based on experience and information, we propose to evaluate the cushioning property and shoe stability using the parameters of the least squares approximate planes of the displacements in x3-direction at characteristic timings during running motion.
The foot pressure distribution on each timing is measured experimentally using F-Scan (Tekscan, Inc.). The results are shown in Fig. 3. Among them, we use the pressure distributions at the times tS (Fig. 3c) and tC (Fig. 3a) for shoe stability and cushioning property, respectively, for pN in the static deformation problem (Problem 1). Appropriate contact condition between the bottom surface \({\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) of the sole and the ground \({\varGamma }_{\mathrm {M}^{\ast }} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) is considered at each time. We will write the two displacements obtained as the solutions of Problem 1 as ur (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) and components in x3-direction as ur3.
We define the least squares approximate plane of ur3 at each time by
for \(\left (x_{1}, x_{2} \right ) \in \mathbb {R}^{2}\). Here, αr, βr, and δr are constants to express the gradients in x1- and x2-directions and the average of sinking in x3-direction of the approximate plane of ur3, respectively. Those parameters are functions of ur. So, we express them as \({\alpha }_{r} \left (\boldsymbol {u}_{r} \right )\), \({\beta }_{r} \left (\boldsymbol {u}_{r} \right )\), and \({\delta }_{r} \left (\boldsymbol {u}_{r} \right )\), respectively.
In this study, we will use \({\alpha }_{\mathrm {S}} \left (\boldsymbol {u}_{\mathrm {S}} \right )\) and \({\delta }_{\mathrm {C}} \left (\boldsymbol {u}_{\mathrm {C}} \right )\) for cost functions with respect to shoe stability and cushioning property, respectively.
4.2 Calculation of approximate planes
The parameters αr, βr, and δr can be obtained as a solution of a least square problem between ur3 and wr. Here, the r is omitted for simplicity. We define the objective function for this problem as
where \({\varGamma }_{\mathrm {U}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) is the boundary to observe sole deformation (refer to Figs. 1 and 2). The least square problem can then be written as follows.
Problem 2 (Approximate plane)
For a given u3, find α, β and δ satisfying
Using the stationary conditions with partial differentiations of fS by those parameters, the solution of Problem 2 can be obtained as
Solving the above system of equations, the parameters α, β, and δ of the least squares planes can be expressed as
where
and
From here, we will denote α, β, and δ of (23), (24), and (25) at time tr (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) by \(\alpha \left (\boldsymbol {u}_{r} \right )\), \(\beta \left (\boldsymbol {u}_{r} \right )\), and \(\delta \left (\boldsymbol {u}_{r} \right )\), respectively.
5 Shape optimization problem
Using the definitions given in the previous section, we will now formulate the shape optimization problem that we will examine in this section. Considering the design process of shoes, we assume that there is an ideal value δI for the cushioning property \(\delta \left (\boldsymbol {u}_{\mathrm {C}} \right )\). Therefore, we define the objective and constraint cost functions as
respectively. Using these cost functions, we construct a shape optimization problem as follows.
Problem 3 (Shape optimization)
For fS and fC of (26) and (27), respectively, find \(\varOmega \left (\boldsymbol {\phi } \right )\) such that
6 Shape derivatives of cost functions
In order to solve the shape optimization problem by a gradient-based method, the Fréchet derivatives of the cost functions with respect to variation of the design variable are required. We will derive them with the Lagrange multiplier method (the adjoint method). Using the Lagrange function \({\mathcal{L}}_{\mathrm {P}}\) with respect to Problem 1 defined in (17), the Lagrange function of fr (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) can be defined as
where vr was introduced as the Lagrange multiplier with respect to Problem 1 for fr. In the Lagrange multiplier method, the shape derivative (Fréchet derivative with respect to domain variation) of fr is given by \(\tilde {f}_{r}^{\prime } \left (\boldsymbol {\phi } \right ) \left [ \boldsymbol {\varphi } \right ]\) using the notation \(\tilde {f}_{r} \left (\boldsymbol {\phi } \right ) = f_{r} \left (\boldsymbol {u}_{r} \left (\boldsymbol {\phi } \right ) \right )\), which is referred to as the reduced cost function, and calculated via the functional \({\mathcal{L}}_{r \boldsymbol {\phi }} \left (\boldsymbol {\phi }, \boldsymbol {u}_{r}, p_{r}, \boldsymbol {v}_{r} \right ) \left [ \boldsymbol {\varphi } \right ]\).
The shape derivative of \({\mathcal{L}}_{r}\) with respect to an arbitrary variation \(\left (\boldsymbol {\varphi }, \hat {\boldsymbol {u}}, \hat {p}, \hat {\boldsymbol {v}} \right ) \in X \times U \times P \times U\) of \(\left (\boldsymbol {\phi }, \boldsymbol {u}_{r}, p_{r}, \boldsymbol {v}_{r} \right )\) can be written as
Here, the third term of the right-hand side of (28) becomes
Then, if \(\left (\boldsymbol {u}_{r}, p_{r} \right )\) is the solution of Problem 1, this term becomes zero. The second term of the right-hand side of (28) becomes
where we used the notation
and the coefficients aj and dj defined in (23) and (25). Noticing that \({\mathcal{L}}_{r \boldsymbol {u}_{r} p_{r}} = 0\), for an arbitrary \(\left (\hat {\boldsymbol {u}}, \hat {p} \right ) \in U \times P\), is the weak form of the adjoint problem below, the second term on the right-hand side of (28) becomes zero when vr is the solution of the following problem.
Problem 4 (Adjoint problem forfr)
Let ur be the solution of Problem 1 for \(\boldsymbol {\phi } \in \mathcal {D}\), find \(\boldsymbol {v}_{r} \in \mathcal {S}\) such that
If \(\left (\boldsymbol {u}_{r}, p_{r} \right )\) and vr are the solutions of Problem 1 and Problem 4, using the formula for shape derivative of \({\mathcal{L}}_{r}\) (see, for instance Delfour and Zolésio (2011, equations (4.7) and (4.16)), we obtain the shape derivative of \( \tilde {f}_{r} \left (\boldsymbol {\phi } \right )\) by
where ΓC0 was defined in (1) and dγ1 denotes the infinitesimal boundary measure on \(\partial {\varOmega }_{1} \left (\boldsymbol {\phi }, \boldsymbol {0}_{\mathbb {R}^{3}} \right )\). Here, the shape gradient for \(\tilde {f}_{r} \left (\boldsymbol {\phi } \right )\) (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) is given by
where ν1 denotes the outward unit normal on \(\partial {\varOmega }_{1} \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\) and uri (\(i \in \left \{ 1, 2 \right \}\)) and vr1 denote ur and vr on \({\varOmega }_{i} \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\) and \({\varOmega }_{1} \left (\boldsymbol {\phi }, \boldsymbol {0} \right )\), respectively. In the process to obtain the result of (28) with (29) and (30), the integrals defined over the boundary \({\varGamma }_{\mathrm {C} 0} = {\varGamma }_{\mathrm {D} 0} \cup {\varGamma }_{\mathrm {U} 0} \cup {\varGamma }_{\mathrm {M} 0} \cup {\varGamma }_{\mathrm {M}^{\ast } 0}\) in (28) vanished since these boundaries are held fixed during domain deformation φ ∈ X. Moreover, the terms including the mean curvature on \({\varGamma }_{12} \left (\boldsymbol {\phi }, \boldsymbol {0}_{\mathbb {R}^{3}} \right )\) also disappears due to the fact that the terms on \(\partial {\varOmega }_{1} \left (\boldsymbol {\phi }, \boldsymbol {0}_{\mathbb {R}^{3}} \right )\) and \(\partial {\varOmega }_{2} \left (\boldsymbol {\phi }, \boldsymbol {0}_{\mathbb {R}^{3}} \right )\) which have the unit normal vectors are oppositely directed.
7 Solution to shape optimization problem
Finally, we show the solution to the shape optimization problem given in Problem 3. Using the shape gradient gr (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)) in (30), we can write the Lagrange function for Problem 3 and its shape derivative as
with respect to an arbitrary φ ∈ X. Here, \(\lambda _{\mathrm {C}} \in \mathbb {R}\) is the Lagrange multiplier for the constraint \(f_{\mathrm {C}} \left (\boldsymbol {u}_{\mathrm {C}} \right ) = 0\). The H1 gradient method of domain variation type is formulated by seeking φgr ∈ X that decreases fr with respect to iterations \(k \in \left \{0, 1, {\ldots } \right \}\) by the following methods.
Problem 5 (H1gradient method forfr)
Let \(a_{X} : {X} \times {X} \to \mathbb {R}\) be a bounded and coercive bilinear form in X, and ca be a positive constant to control the magnitude of φgr (\(r \in \left \{ \mathrm {S}, \mathrm {C} \right \}\)). For \(\boldsymbol {g}_{r} \left (\boldsymbol {\phi }_{k} \right ) \in {X}^{\prime }\), find φgr ∈ X such that
In this paper, we use
where cΩ is a positive constant to guarantee the coerciveness of the bilinear form aX. A simple algorithm for solving Problem 3 by the H1 gradient method is shown next.
We developed a computer program based on Algorithm 7. In the program, a commercial finite element program Abaqus 2018 (Dassault Systèmes) is used to solve state determination and adjoint problems (Problems 1 and 4). Moreover, OPTISHAPE-TS 2018 (Quint Corporation) is used to solve the boundary value problem (Problem 5) using the H1 gradient method.
8 Numerical example
To confirm the solvability of the shape optimization problem (Problem 3) by the method shown in the previous section, we will show a numerical result using a cubic model. After that, a simple sole model will be used to demonstrate the validity of the presented method in the design of shoe sole such that the parameter for stability is maximized while the parameter for cushioning property keeps an ideal value.
In the following sections, we will show the necessity of considering the material nonlinearity in the analysis of the cubic model, as well as the necessity of using the contact condition in the case of a simple sole model.
8.1 Cubic model
Figure 4 shows a finite element model of a cubic body consisting of Ω1 = (0,0.05) × (0,0.05) × (0.025,0.05) [m3] and Ω2 = (0,0.05) × (0,0.05) × (0,0.025) [m3] defined in Fig. 1. The loading boundary Γp0 is assumed on the upper boundary of Ω1. The pressure distributions at the times tS and tC are assumed as follows:
where e3 denotes the unit vector on x3-coordinate. The magnitude of the pressure is decided as much as equivalent to that of the actual pressure in the running motion. The material parameters used in ABAQUS are shown in Table 1 assuming that Ω1 and Ω2 are made of hyper foams. The Young’s modulus eY used in the linear elastic analysis is calculated using the material parameters in Table 1 and Poisson’s ratio νP by \(2 \left (1 + \nu _{\mathrm {P}} \right ) \left (\mu _{1} + \mu _{2} + \mu _{3} \right )\) (eY = 0.78 MPa and νP = 0.3 in Ω1, and eY = 2.60 MPa and νP = 0.3 in Ω2). The finite element model is constructed with 5304 ten-node tetrahedron elements and 8485 nodes. In this analysis, we assumed that \({\varGamma }_{\mathrm {M} 0} = {\varGamma }_{\mathrm {M}^{\ast } 0}\) is set as ΓD0 (i.e., fixed), and only the material boundary Γ120 = ∂Ω10 ∩ ∂Ω20 varies.
Figure 5 shows the optimized shapes of the cubic model. Figure 5a is the result obtained by considering the finite deformation and material nonlinearity, while (b) is the result obtained through linear elastic analysis. The iteration histories of cost functions are shown in Fig. 6. The notation fSinit represents the value of fS at the initial shape. From the graphs, it can be confirmed that both the objective cost functions fS decreased monotonically while keeping the constraint condition fC ≤ 0. However, the amount of reduction is greater in the case considering the nonlinearities than the case of linear analysis. This result demonstrates that the linear analysis cannot follow the deformation sufficiently (too stiff) and cannot extend to the range of the deformation which we consider for the shoe design.
For reference, we examined a finite deformation analysis using the linear material. However, we could not continue the analysis by mesh distortion. In addition, using the optimized model obtained by the linear analysis shown in Fig. 5b, we analyzed the finite deformation with the nonlinear material and obtained the cost values fS/fSinit = 0.64 and 1 + fC/δI = 1.03. From the result, it can be confirmed that the inequality constraint fC ≤ 0 is violated. This violation is unacceptable in the design region of shoes.
8.2 Sole model
A finite element model of a sole is constructed with eight-node hexahedral elements as shown in Fig. 2. The model consists of three domains colored in yellow, light blue, and gray. The domains in yellow and light blue were modeled by hyper foam with different hardness. For the domain in gray, the Neo-Hookean model was used. We assumed that Ω20 consists of the domains in light blue and gray in which the gray domain is fixed in the domain variation. The parameters of these models were identified based on experimental results. We used the foot pressure distributions as shown in Fig. 3a and c for pN. In this analysis, we assumed that the ground Ω30 is a rigid body and fixed in the domain variation, and contact condition between the bottom surface \({\varGamma }_{\mathrm {M}} \left (\boldsymbol {0}, \boldsymbol {u} \right )\) of the sole and the ground \({\varGamma }_{\mathrm {M}^{\ast } 0} = {\varGamma }_{\mathrm {D} 0}\) with friction free is considered. For contact type in Abaqus 2018, the option ‘surface-surface’ was used. In the case using a fixed condition, we assumed that \({\varGamma }_{\mathrm {M} 0} = {\varGamma }_{\mathrm {M}^{\ast } 0}\) is the fixed boundary ΓD0, and only the material boundary Γ120 = ∂Ω10 ∩ ∂Ω20 varies.
Figure 7a shows the optimized shapes of Ω10 and Ω20 obtained by the developed program using the nonlinear material and aforementioned contact condition. In contrast, Fig. 7b shows the result in which the contact condition is changed to the fixed condition on \({\varGamma }_{\mathrm {M} 0} = {\varGamma }_{\mathrm {M}^{\ast } 0} (= {\varGamma }_{\mathrm {D} 0})\). The iteration histories of cost functions with respect to the number of reshaping are shown in Fig. 8. Here, fSinit denotes the fS at the initial shape, too. From the graphs, it can be confirmed that the objective cost function fS decreased monotonically while keeping the constraint condition fC ≤ 0. This variation means that maintaining the average sinking δI at the time tC in the contact phase, the tilt angle \(\alpha \left (\boldsymbol {u}_{\mathrm {S}} \right )\) at the time tS in the mid stance phase is minimized.
Comparing the results between the cases with contact condition and with fixed condition, we obtained a large reduction of fS in the case with contact condition as evident in Fig. 8. A remarkable difference is found at the shape of mid top surface in \({\varOmega }_{2} \left (\boldsymbol {\phi } \right )\). In the contact case, a wavy variation in the thickness from the heel to toe is generated in the process of optimization, while only a small variation is created in the fixed case. Those differences can be considered as an effect of the contact and fixed conditions. When the contact condition is used with respect to the pressure distributions at the times tS (Fig. 3c), the bottom surface of \({\varOmega }_{2} \left (\boldsymbol {\phi } \right )\) contacts with the mid foot from the heel and does not contact at the remaining part (Fig. 9b). This loading condition makes intensely focused variation in the mid top surface in \({\varOmega }_{2} \left (\boldsymbol {\phi } \right )\). On the other hand, when the fixed condition is used, all the bottom surface supports the pressure by tensile and compressible forces which are distributed widely. As a result, it can be considered that the shape variation becomes small and the reduction of fS becomes small, too.
From the result that a great difference is observed between the contact and fixed conditions, it can be concluded that the contact condition is required in the design of sole.
9 Conclusion
In this paper, an optimized shape design problem for shoe sole improving the shoe stability and cushioning property was formulated using parameters of planes that approximate the deformations analyzed by the finite element method using the pressures measured during the running motion as external boundary forces. In the finite element analysis, finite deformation, nonlinear constitutive equations for hyper-elastic materials, and contact condition between the bottom of sole and the ground were taken into account. Using the approximate planes, the tilt angle in the mid stance phase and the average sinking in the contact phase were chosen as objective and constraint cost functions, respectively. The target boundaries to optimize were the bonded boundary of soft and hard materials and side boundary. Numerical results using a realistic finite element model demonstrated that the optimized shape of the formulated problem can be obtained by the presented method. Regarding the selection of the parameters for objective and constraint cost functions, we have to hear desires of running shoe developers based on evaluation by runners.
References
Areblad M, Nigg BM, Ekstrand J, Olsson KO (1990) Three-dimensional measurement of rearfoot motion during running. J Biomech 23:933–940
Azegami H (2016) Shape optimization problems (in Japanese). Morikita Publishing, Tokyo
Azegami H (2017) Solution of shape optimization problem and its application to product design. In: Itou H, Kimura M, Chalupecký V, Ohtsuka K, Tagami D, Takada A (eds) Proceedings of the international conference CoMFoS15 Mathematical analysis of continuum mechanics and industrial applications, Springer Singapore, Mathematics for industry, 26: 83–98. https://doi.org/10.1007/978-981-10-2633-1
Azegami H, Takeuchi K (2006) A smoothing method for shape optimization: Traction method using the Robin condition. Int J Comput Methods 3(1):21–33
Azegami H, Wu ZQ (1996) Domain optimization analysis in linear elastic problems: Approach using traction method. JSME Int J Ser A 39(2):272–278
Cardoso JB, Arora JS (1988) Variational method for design sensitivity analysis in nonlinear structural mechanics. AIAA J 26:595– 603
Cavanagh P (1980) The running shoe book. Anderson World, California
Choi KK, Chang KH (1994) A study on the velocity computation for shape design optimization. J Finite Elem Anal Des 15:317–347
Clarke TE, Frederick EC, Cooper LB (1983) Effects of running shoes on ground reaction forces. Int J Sports Med 4:247–251
Dassault Systèmes (2018) Abaqus Analysis User’s Guide 2018. Vélizy-Villacoublay
Delfour MC, Zolésio JP (2011) Shapes and geometries: Metrics, analysis, differential calculus and optimization, 2nd edn. Society for Industrial and Applied Mathematics, Philadelphia
Gard SA, Konz RJ (2004) The effect of a shock-absorbing pylon on the gait of persons with unilateral transtibial amputation. J Rehabil Res Dev 40:109–124
Gibson LJ, Ashby MF (1980) Cellular solids structure and properties, Cambridge University Press, Cambridge, England
Hisada T (1995) Recent progress in nonlinear fem-based sensitivity analysis. JSME Int J 38:301–310
Ihara H, Azegami H, Shimoda M (1999) Solution to shape optimization problems considering material nonlinearity. Comput Aided Opt Des Struct 40:87–95
Iwai T, Sugimoto A, Aoyama T, Azegami H (2010) Shape optimization problem of elastic bodies for controlling contact pressure. JSIAM Lett 2:1–4
Kaneko I, Majer G (1981) Optimal design of plastic structures under displacement constraints. Comput Methods Appl Mech Engrg 27:369–391
Kim NH, Choi KK, Chen JS (2000) Shape design sensitivity analysis and optimization of elasto-plasticity with frictional contact. AIAA J 398:1742–1753
Matheson GO, Clement DB, Makenzie DC, Taunton J, Lioyd-Smith D, Maclntyre JG (1987) Stress fractures in athletes: A study of 320 cases. Am J Sports Med 15:46–68
Mills NJ (2007) Polymer foams handbook. Butterworth-Heinemann, Oxford
Nakabe N, Nishiwaki T (2002) Development of simplified numerical foot model for sole stability design. Eng Sport 4:824–830
Nigg BM (1980) Biomechanics of running shoes. Human Kinetics, Champaign, Illinois
Nigg B M, Bahlsen HA, Luethi SM, Stokes S (1987) The influence of running velocity and midsole hardness on external impact forces in heel-toe running. J Biomech 20:951–959
Nigg BM, Herzog W, Read LJ (1988) Effect of viscoelastic shoe insoles on vertical impact forces in heel-toe running. Am J Sports Med 16:70–76
Nishiwaki T (2008) Running shoe sole stiffness evaluation method based on eigen vibration analysis. Sports Technol 1:76–82
Oriwol D, Sterzing T, Milani TL (2011) The position of medial dual density midsole elements in running shoes does not influence biomechanical variables. Footwear Sci 3:107–116
Ryu YS, Haririan M, Wu CC, Arora JS (1985) Structural design sensitivity analysis of nonlinear response. Comput Struct 21:245–255
Stacoff A, Reinschmidt C, Stüssi E (1992) The movement of the heel within a running shoe. Med Sci Sports Exerc 24:695–701
Tanaka M, Noguchi H (2004) Structural shape optimization of hyperelastic material by discrete force method. Theor Appl Mech Japan 53:83–91
Tsay JJ, Arora JS (1990) Nonlinear structural design sensitivity analysis for path dependent problems, part1: General theory. Comput Methods Appl Mech Engrg 81:183–208
Vidal CA, Haber RB (1993) Design sensitivity analysis for rate-independent elastoplasticity. Comput Methods Appl Mech Engrg 107:393–431
Vidal CA, Lee HS, Haber RB (1991) The consistent tangent operator for design sensitivity analysis of history-dependent response. Comput Syst Engrg 2:509–523
Woensel WV, Cavanagh PR (1992) A perturbation study of lower extremity motion during running. Int J Sport Biomech 8:30–47
Yamazaki K, Shibuya K (1998) Sensitivity analysis of nonlinear material and its application to shape optimization. In: Collect Tech Pap AIAA/ASME/ASCE/AHS/ASC Struct Struct Dyn Mater Conf, pp 1706–1713
Yuge K, Kikuchi N (1995) Optimization of a frame structure subjected to a plastic deformation. Struct Optim 10:197–208
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Responsible Editor: Seonho Cho
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Replication of results
The results described in this paper can be replicated by implementing the formulas and algorithms described in this paper. Regarding the foot pressure distributions and material parameters for the sole model used in the numerical examples, the authors want to withhold its replication for commercial purposes.
Rights and permissions
About this article
Cite this article
Nonogawa, M., Takeuchi, K. & Azegami, H. Shape optimization of running shoes with desired deformation properties. Struct Multidisc Optim 62, 1535–1546 (2020). https://doi.org/10.1007/s00158-020-02560-0
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00158-020-02560-0