betaFeedback

WorksComputer Science

Integrating Material Selection with Design Optimization via Neural Networks

Abstract

The engineering design process often entails optimizing the underlying geometry while simultaneously selecting a suitable material. For a certain class of simple problems, the two are separable where, for example, one can first select an optimal material, and then optimize the geometry. However, in general, the two are not separable. Furthermore, the discrete nature of material selection is not compatible with gradient-based geometry optimization, making simultaneous optimization challenging. In this paper, we propose the use of variational autoencoders (VAE) for simultaneous optimization. First, a data-driven VAE is used to project the discrete material database onto a continuous and differentiable latent space. This is then coupled with a fully-connected neural network, embedded with a finite-element solver, to simultaneously optimize the material and geometry. The neural-network's built-in gradient optimizer and back-propagation are exploited during optimization. The proposed framework is demonstrated using trusses, where an optimal material needs to be chosen from a database, while simultaneously optimizing the cross-sectional areas of the truss members. Several numerical examples illustrate the efficacy of the proposed framework. The Python code used in these experiments is available at github.com/UW-ERSL/MaTruss

Introduction

In engineering design, we are often faced with the task of optimizing the underlying geometry and selecting an optimal material [1]. As a simple example, consider the truss design problem illustrated in Figure 1 where one must optimize the cross-sectional areas, while selecting, for simplicity, a single material for all the truss members. Formally, one can pose this as [2, 3]:

A truss design problem involving optimization of cross-sectional areas and material selection

Figure 1. A truss design problem involving optimizing the cross-sectional areas of truss members, and selecting an optimal material.

minimize⁡A={A1,A2,…,AN}, m∈Mϕ(A,ζm)(1)\underset{\mathbf{A}=\{A_1,A_2,\ldots,A_N\},\,m\in\mathcal{M}}{\operatorname{minimize}\quad} \phi(\mathbf{A},\zeta_m) \tag*{(1)}
subject to⁡g(A,ζm)≤0(2)\operatorname{subject\ to}\quad g(\mathbf{A},\zeta_m)\leq0 \tag*{(2)}
K(A,Em)u=f(3)K(\mathbf{A},E_m)\mathbf{u}=\mathbf{f} \tag*{(3)}
Amin⁡≤A≤Amax⁡(4)A_{\min}\leq\mathbf{A}\leq A_{\max} \tag*{(4)}

where ϕ\phi is the objective (for example, compliance), gg denotes a set of constraints (such as yield stress and buckling), f\mathbf{f} is the applied force, KK is the truss stiffness matrix, and u\mathbf{u} is the nodal displacements. The design variables are the cross-sectional areas of the members A={A1,A2,…,AN}\mathbf{A}=\{A_1,A_2,\ldots,A_N\}, with limits Amin⁡A_{\min} and Amax⁡A_{\max}, and the material choice m∈Mm\in\mathcal{M}. We denote the collection of relevant material properties by ζm\zeta_m; these may include, for example, the Young’s modulus (Em)(E_m), cost per unit mass (Cm)(C_m), mass density (ρm)(\rho_m) and yield strength (Ym)(Y_m).

Literature Review

Observe that the two sets of design variables are tightly coupled. In other words, one cannot, for example, pick a material, and then optimize the cross-sectional areas (or, vice-versa). To quote [4] "... either approach does not guarantee the optimal combination of geometry and material". Furthermore, while the cross-sectional areas are continuously varying, the material choice is discrete, making the problem difficult to solve using classic gradient-based optimization.

If the geometry is fixed, Ashby’s method [5] that relies on a performance metric, is the most popular strategy for material selection. Furthermore, for simple design problems where the loads, geometry and material functions are separable [6], a material index can be used to select the best material. This is simple, efficient and reliable. However, when there are multiple objectives or constraints, a weighted approach is recommended [7]. In practice, computing these weights is not trivial, and the material choice may be far from optimal. Several non-gradient methods have been proposed for material selection [8, 9, 10], but these cannot be integrated with gradient-based optimization.

An alternate concept of a design index was proposed in Ananthasuresh and Ashby [11]. It was shown that the design index can be used to construct a smooth material function, enabling simultaneous optimization of the geometry and selection of the material via gradient based optimization [4]. However, design indices were limited to simple determinate trusses. In [12], for the special case of a single stress constraint and a single load, the authors concluded that utmost two materials are sufficient and that the truss design problem can be cast as a linear programming problem. Unfortunately, this method does not apply when there are multiple constraints. Recently, the authors of [13] considered two materials (glue-laminated timber and steel) to design truss structures incorporating stress constraints as a mixed-integer quadratic problem (MIQPs). However the authors note that "[MIQPs]... require algorithmic tuning and considerable computational power". Another hybrid approach was proposed in [14] where a combination of the gradient-based methods such as sequential quadratic programming (SQP) and an evolution method such as genetic algorithm (GA) are used to solve MDNLPs. But these methods can be prohibitively expensive.

Mathematically, the problem of (discrete) material selection and (continuous) area optimization can be cast as mixed-discrete nonlinear programming problems (MDNLPs). Such problems are fairly common in engineering, and several solution strategies methods have been proposed; see [15, 16, 17]. However, regardless of the strategy, these methods entail repeated solution of a sequence of nonlinear programming problems with careful relaxations and approximations, making them sensitive to assumptions and underlying models [18]. Furthermore, the popular branch and bound algorithm used in solving MDNLPs [16] does not apply here since the optimization problem depends indirectly on the material index through the database.

Paper Overview

The primary contribution of this work is the use of variational autoencoders (VAEs) to solve such problems. VAEs are a special form of neural networks that can convert discrete data, such as a material database, into a continuous and differentiable representation, making the design problem amenable to gradient-based optimization. In the proposed method (see Figure 2), a data-driven VAE is trained using a material database, and the trained decoder is then integrated with neural networks to simultaneously optimize the geometry and material.

Figure 2. An overview of the proposed method.

The proposed VAE framework is discussed in § 2. In § 3, the VAE is integrated with two additional and simple neural networks to solve the truss design problem. Specifically, we leverage the differentiable material representation to simultaneously optimize the geometry and material. § 4 demonstrates the proposed framework through several numerical examples. In § 5, limitations and future work are discussed.

A Differentiable Material Representation

In this section, we discuss how variational autoencoders (VAEs) can be used to obtain a continuous and differentiable representation of a discrete material database. This will serve as a foundation for the next section on design optimization.

VAE Architecture and Training

VAEs are popular generative models that have wide applicability in data compression, semi-supervised learning and data interpolation; please see [19]. In essence, a VAE captures the data in a form that can be used to synthesize new samples similar to the input data. For example, VAEs have been used to generate new microstructures from image databases [20], in designing photonic crystals [21], and heat-conduction materials [22].

However, in this paper, we do not use VAEs to synthesize new data, instead we simply use the VAE’s ability to map uncorrelated data onto an abstract latent space. In that sense, VAEs are similar to principal component analysis (PCA) in its ability to extract useful information from the data. However, the nonlinear nature of VAEs allows for far greater generalization than PCA [23]. In particular, the proposed VAE architecture for capturing material properties is illustrated in Figure 3, and consists of the following components:

Figure 3. Architecture of the variational autoencoder.

  1. A four-dimensional input module corresponding to the four properties in table 1, namely the Young’s modulus (E)(E), cost (C)(C), mass density (ρ)(\rho) and yield strength (Y)(Y). The input set is denoted by ζ\zeta.

  2. An encoder FF consisting of a fully-connected network of 250 neurons, where each neuron is associated with an ReLU activation function and weights [24].

  3. A two-dimensional latent space, denoted by z0,z1z_0, z_1 that lies at the heart of the VAE.

  4. A decoder DD, which is similar to the encoder, consists of a fully-connected network of 250 neurons.

  5. A four-dimensional output corresponding to the same four properties; the output set is denoted by ζ^\hat{\zeta}.

In this work, the VAE was trained on a material database consisting of 92 materials [25], where table 1 represents a small sample.

MaterialClassEE [Pa][Pa]cost CC [$/kg][\$/kg] | ρ\rho [kg/m3][kg/m^3] | YY [Pa][Pa]
A286 IronSteel2.01E+115.18E+007.92E+036.20E+08
AISI 304Steel1.90E+112.40E+008.00E+035.17E+08
Gray Cast IronSteel6.62E+106.48E-017.20E+031.52E+08
3003-H16Al Alloy6.90E+102.18E+002.73E+031.80E+08
5052-OAl Alloy7.00E+102.23E+002.68E+031.95E+08
7050-T7651Al Alloy7.20E+102.33E+002.83E+035.50E+08
AcrylicPlastic3.00E+092.80E+001.20E+037.30E+07
ABSPlastic2.00E+092.91E+001.02E+033.00E+07
PE HDPlastic1.07E+092.21E+009.52E+022.21E+07

Table 1. A curated subset of materials and their properties used in the training.

As mentioned earlier, the VAE’s primary task is to match the output to the input as closely as possible. This is done through an optimization process (also referred to a training), using the weights associated with the encoder and decoder as optimization parameters. In other words, we minimize ∥ζ−ζ^∥\lVert\zeta- \hat{\zeta} \rVert. Additionally, a KL divergence loss is imposed to ensure that latent space resembles a standard Gaussian distribution z∼N(μ=0,σ=1)z \sim\mathcal{N}(\mu= 0, \sigma= 1) [19]. Thus, the net loss can then expressed as:

L=∥ζ−ζ^∥+βKL(z∥N)(5)L = \lVert\zeta- \hat{\zeta} \rVert+ \beta KL(z\mathbin{\|}N) \tag*{(5)}

where β\beta is set to a recommended value of 5∗10−55 * 10^{-5}. Further the input is scaled between (0,1)(0, 1) for all attributes, and the output is re-scaled back after training. This ensures that all material properties are weighted equally. In this work, we used PyTorch [26] to model the VAE, and the gradient-based Adam optimizer [27] to minimize eq. (5) with a learning rate of 0.002 and 50, 000 epochs. The convergence is illustrated in Figure 4; the training took approximately 51 seconds on a Macbook M1 pro.

Figure 4. Convergence plot during training of the VAE.

Thus, the VAE captures each material uniquely and unambiguously using a non-dimensional latent space, i.e., z0z_0, z1z_1 in our case; this is visualized in Figure 5. For example, annealed AISI 4340 is represented by the pair (−1.2,1.0)(-1.2, 1.0) while Acrylic is represented by (0.6,−0.4)(0.6, -0.4). The VAE also clusters similar materials together in the latent space, as can be observed in Figure 5. In an abstract sense, the latent space in Figure 5 is similar to the popular Ashby charts [5].

Encoding of the materials in the latent space

Figure 5. Encoding of the materials in the latent space; only a few materials are annotated for clarity.

Representational Accuracy

One can expect the values reconstructed using the decoder to deviate from the true material data; table 2 summarizes the errors. The following observations are worth noting:

  1. Despite the lack of correlation between material properties, and two orders of magnitude difference in values, the VAE captures the entire database of 92 materials reasonably well, using a simple two-dimensional latent space.

  1. The error can be further reduced by either increasing the dimension of the latent space, or limiting the type of materials considered (see § 4).

  1. Finally, as will be discussed in the next section, the original material data will be used, as part of a post-processing step, at the end of optimization.

MaterialΔE%\Delta E\%ΔC%\Delta C\%Δρ%\Delta\rho\%ΔY%\Delta Y\%
A286 Iron0.80.70.80.0
ABS1.40.00.71.7
AISI 3044.33.10.41.0
Gray Cast Fe1.31.80.20.2
3003-H163.26.81.31.1
5052-O1.42.40.61.2
7050-T76510.91.91.01.9
Acrylic2.30.50.20.5
PE HD0.70.91.11.8
Max error5.06.83.37.1

Table 2. Percentage error between actual and decoded data.

Differentiability of Latent Space

A crucial aspect of the latent space is its differentiability. In particular, we note that using the decoder, each of the output material properties is represented via analytic activation functions; for example, E^=DE∗(z0,z1)\hat{E}=D^{*}_{E}(z_{0},z_{1}) where DE∗(⋅)D^{*}_{E}(\cdot) is a combination of activation functions and weights. It means that one can back propagate through the decoder to compute analytical sensitivities; for example, the sensitivity (∂E^∂z0)\left(\frac{\partial\hat{E}}{\partial z_{0}}\right). This plays an important role in gradient-driven design optimization.

Material Similarity

While the accuracy and differentiability of the VAE lends itself to design optimization, the latent space also provides key insights into material characteristics. For instance, one can compute the Euclidean distance between materials in the latent space. These distances are reported in Figure 6. As one can expect, steels are closer to aluminum alloys, than they are to plastics.

Figure 6. Symmetric distance in the latent space between materials

Furthermore, one can overlay the latent space in Figure 5 with specific material properties to gain further insight. This is illustrated in Figure 7 where the contour plots of the Young’s moduli are illustrated in the latent space. This allows designers to visualize similarities between materials, specific to a material property.

Contours of Young’s moduli in the latent space

Figure 7. Contours of the Young’s moduli in the latent space.

Design Optimization

Having constructed a differentiable representation of material properties, one can now pose the optimization problem discussed in section 1, now using {A,z0,z1}\{A,z_{0},z_{1}\} as continuous design variables. Without a loss in generalization, we consider a specific instance of the truss optimization problem, where the objective is the compliance with three sets of constraints: cost constraint, buckling constraint, and tensile yield constraint. We assume here that members will fail due to buckling, before failing due to compressive yield [4]. The resulting optimization problem for circular truss members

can be posed as [28]:

minimize⁡A={A1,A2,…,AN},z0,z1J=fTu(A,E^)subject to[K(A,E^)]u=fgc:=(ρ^C^C∗∑k=1NAkLk)−1≤0gb:=max⁡k(−4PkLk2π2E^Ak2)−1Fs≤0gy:=max⁡k(PkY^Ak)−1Fs≤0Amin⁡≤A≤Amax⁡(6)\begin{aligned} \underset{A=\{A_1,A_2,\ldots,A_N\},z_0,z_1}{\operatorname{minimize}}\quad& J=f^{T}u(A,\hat{E}) \\ \text{subject to}\quad& [K(A,\hat{E})]u=f \\ & g_c:=\left(\frac{\hat{\rho}\hat{C}}{C^{*}}\sum_{k=1}^{N} A_kL_k\right)-1\leq0 \\ & g_b:=\max_k\left(\frac{-4P_kL_k^2}{\pi^2\hat{E}A_k^2}\right)-\frac{1}{F_s}\leq0 \\ & g_y:=\max_k\left(\frac{P_k}{\hat{Y}A_k}\right)-\frac{1}{F_s}\leq0 \\ & A_{\min}\leq A\leq A_{\max} \tag*{(6)} \end{aligned}

where PkP_k is the internal force in member kk, LkL_k is its length, C∗C^{*} is the allowable cost, FsF_s is the safety factor, and the material properties (E^,ρ^,Y^,C^)(\hat{E},\hat{\rho},\hat{Y},\hat{C}) are decoded from the latent space coordinates (z0,z1)(z_0,z_1). Further, to facilitate gradient driven optimization, the max⁡\max operator is relaxed here using a p-norm, i.e., max⁡i(xi)≈∥x∥p\max_i(x_i)\approx\lVert x\rVert_p, with p=6p=6.

To solve the above optimization we use two additional neural networks (NNs) (see Figure 8a): a truss network NNTNN_T, parameterized by weights wTw_T, and a material network NNMNN_M, parameterized by weights wMw_M. The truss network NNTNN_T is a simple feed-forward NN with two hidden layers with a width of 20 neurons, each containing an ReLU activation function. The output layer of NNTNN_T consists of NN neurons where NN is the number of truss members, activated by a Sigmoid function, generates a vector OTO_T of size NN whose values are in [0,1][0,1]. The output is then scaled as A←Amin⁡+OT(Amax⁡−Amin⁡)A\leftarrow A_{\min}+O_T(A_{\max}-A_{\min}) to satisfy the area bounds eq. (3f). The material network NNMNN_M (Figure 8b) is similar to NNTNN_T, except that the output layer consists of two output neurons activated by Sigmoid functions. The outputs OMO_M are scaled as z←−3+6OMz\leftarrow-3+6O_M, resulting in zi∈[−3,3]z_i\in[-3,3] corresponding to six Gaussian deviations. These outputs now interface with the trained decoder D∗D^{*} from Figure 3. Thus, by varying the weights wMw_M of the material network one can create points in the latent space, that then feeds to the trained decoder resulting in values of material constants.

Figure 8. The truss and material neural networks.

Loss Function

With the introduction of the two NNs, the weights wTw_T now control the areas AA, while the weights wMw_M control the material constants ζ^\hat{\zeta}. In other words, the weights wTw_T and wMw_M now form the design variables. Further, since NNs are designed to minimize an unconstrained loss function, we convert the constrained minimization problem in eq. (3) into an unconstrained minimization by employing a log-barrier scheme as proposed in [29]:

L(wT,wM)=J+ψ(gc)+ψ(gb)+ψ(gy)(7)L(w_T,w_M)=J+\psi(g_c)+\psi(g_b)+\psi(g_y) \tag*{(7)}

where,

ψt(g)={−1tlog⁡(−g),g≤−1t2,tg−1tlog⁡(1t2)+1t,otherwise(8)\psi_t(g)= \begin{cases} -\frac{1}{t}\log(-g), & g\leq-\frac{1}{t^2},\\ tg-\frac{1}{t}\log\left(\frac{1}{t^2}\right)+\frac{1}{t}, & \text{otherwise} \end{cases} \tag*{(8)}

with the parameter tt updated during each iteration (as described in the next section). Thus the optimization problem reduces to a simple form:

minimize⁡wT,wML(wT,wM)subject to[K(A(wT),E^(wM))]u=f(9)\begin{aligned} \underset{w_T,w_M}{\operatorname{minimize}}\quad&L(w_T,w_M) \\ \text{subject to}\quad&[K(A(w_T),\widehat{E}(w_M))]u=f \tag*{(9)} \end{aligned}

A schematic of the proposed framework is presented in Figure 9.

Figure 9. Forward computation

Structural Analysis

We rely on classical structural analysis to evaluate the performance of the truss structure during each iteration. The solver compute the stiffness matrix for each element and upon assembling the global stiffness matrix, the nodal displacement vector uu are computed using a sparse solver in PyTorch [26]. This allows us to exploit backward propagation for automatic differentiation [30], resulting in an end-to-end differentiable solver with automated sensitivity analysis as described next.

Sensitivity Analysis

In the proposed framework, the loss function in eq. (7) is minimized using gradient-based Adagrad optimizer [31]. Further, the sensitivities are computed automatically using back propagation. A schematic representation of the backward computation graph is shown in Figure 10. For instance the sensitivity of the loss function with respect to the weights wTw_T can be expressed as:

Figure 10. Backward computation

∂L∂wT=[∂L∂J⏟I∂J∂u⏟II∂u∂A⏟III+⋯ ]∂A∂wT⏟IV(10)\frac{\partial L}{\partial w_T} = \left[ \underbrace{\frac{\partial L}{\partial J}}_{I} \underbrace{\frac{\partial J}{\partial u}}_{II} \underbrace{\frac{\partial u}{\partial A}}_{III} +\cdots \right] \underbrace{\frac{\partial A}{\partial w_T}}_{IV} \tag*{(10)}

This is illustrated in Figure 10 where each term corresponds to an edge in the graph. Similarly, the sensitivity with respect to wMw_M is given by:

∂L∂wM=[∂L∂J⏟I∂J∂u⏟II∂u∂ζ^⏟V+⋯ ]∂ζ^∂wD∗⏟VI∂wD∗∂z⏟VII∂z∂wM⏟VIII(11)\frac{\partial L}{\partial w_M} = \left[ \underbrace{\frac{\partial L}{\partial J}}_{I} \underbrace{\frac{\partial J}{\partial u}}_{II} \underbrace{\frac{\partial u}{\partial\widehat{\zeta}}}_{V} +\cdots \right] \underbrace{\frac{\partial\widehat{\zeta}}{\partial w_{D^*}}}_{VI} \underbrace{\frac{\partial w_{D^*}}{\partial z}}_{VII} \underbrace{\frac{\partial z}{\partial w_M}}_{VIII} \tag*{(11)}

These are once again illustrated in Figure 10. Note that the backward computation is carried out automatically in the proposed framework.

Post Processing

Once the optimization process is complete, we obtain an optimal set of latent coordinates z∗z^*. However, there might not exist a material in the database corresponding precisely to z∗z^*. We therefore define a confidence metric γm\gamma_m for each material mm by using the distance from z∗z^* to the material mm in the latent space:

γ(zm,z∗)=1−∥z∗−zm∥max⁡∀k∈M(∥z∗−zk∥)(12)\gamma(z_m,z^*) = 1 - \frac{\lVert z^* - z_m\rVert}{\max_{\forall k\in M}(\lVert z^* - z_k\rVert)} \tag*{(12)}

The metric serves to rank the materials based on their distance from z∗z^*. After finding the closest material, we repeat the geometry optimization to compute the optimal areas A∗A^*. This is analogous to the rounding method in discrete optimization [16]: "After rounding, it is usually best to repeat the optimization once more, allowing only the continuous design variables to vary. This process may not lead to the exact optimum, and sometimes may not even lead to a feasible solution, but for many problems this is an effective approach." This is explored further through numerical experiments in § 4.

Algorithm

The proposed algorithms are summarized in this section. As mentioned earlier, we assume that a material database has been provided. The procedure begins with training the VAE as described in algorithm 1. The encoder is then discarded, and the decoder retained. In algorithm 2, starting with the given truss, specified loads, restraints, area bounds and other constraints, the areas and material are optimized. In our experiments t0=3t_0 = 3 and μ=1.01\mu= 1.01, and the learning rate for the Adagrad optimizer was set to the value of 2×10−32 \times10^{-3}. Once optimized, we find the optimal areas using the nearest material in the latent space.

Numerical Experiments

In this section, we present results from several experiments that illustrate the method and algorithm.

Simultaneous Optimization of Area and Material

The central hypothesis of this paper is that simultaneous optimization of material and geometry leads to superior performance compared to sequential optimization. To validate this hypothesis, we consider the 6-bar truss system illustrated in Figure 11, and perform the following three experiments:

Image not extracted

Figure 11. Loading of 6-bar mid-cantilever truss.

  1. Material optimization: Here the areas are fixed at the initial value of Ak=2×10−3m2A_k = 2 \times10^{-3}m^2, ∀k\forall k, and only the material is optimized using the decoder.

Algorithm 1 Encode Materials
1: procedure MATENCODE($\zeta$, $F$, $D$) ▷ Input: Training data, encoder and decoder
2:     epoch = 0 ▷ iteration counter
3:     repeat ▷ VAE training
4:         $F(\zeta) \to \{\mu,\sigma\}$ ▷ Forward prop. encoder
5:         $\{\mu,\sigma\} \to z$ ▷ Reparameterization {citep}`[19]`
6:         $\{\mu,\sigma\} \to KL(z\|\mathcal{N})$ ▷ KL loss
7:         $D(z) \to \hat{\zeta}$ ▷ Forward prop. decoder
8:         $\{\zeta,\hat{\zeta},KL\} \to L$ ▷ VAE Loss
9:         $w+\Delta w(\nabla L) \to w$ ▷ Update VAE weights, Adam Optimizer
10:        epoch ++
11:    until error is acceptable ▷ Iterate
12:    return $D$ ▷ Trained decoder
Algorithm 2 Truss Optimization
1: procedure TRUSSOPT(truss, loads, restraints, $D^*$, ...) ▷ Input: Initial truss, trained decoder, max cost
2:     $k=0$ ▷ iteration counter
3:     repeat ▷ Optimization (Training)
4:         $NN_T(w_T) \to A$ ▷ Fwd prop $NN_T$; compute truss areas
5:         $NN_M(w_M) \to z$ ▷ Fwd prop $NN_M$; compute latent coordinate
6:         $D^*(z) \to \hat{\zeta}$ ▷ Fwd prop $D^*$; compute material properties
7:         $\{A,\hat{E}\} \to [K]$ ▷ compute stiffness matrix
8:         $\{[K],f\} \to u$ ▷ State Eq., FEA eq. (3b)
9:         $\{u,f\} \to J$ ▷ Compliance, design objective eq. (3a)
10:        $\{A,\hat{\rho},\hat{C}\} \to g_c$ ▷ cost constraint eq. (3c)
11:        $\{A,\hat{E}\} \to g_b$ ▷ Buckling constraint eq. (3d)
12:        $\{A,\hat{Y}\} \to g_y$ ▷ Yielding constraint eq. (3e)
13:        $\{J,g_c,g_b,g_y\} \to L$ ▷ Loss eq. (4)
14:        $AD(L \gets w_T,g_c \gets w_T,g_b \gets w_T,g_y \gets w_T) \to \nabla_{w_T}L$ ▷ auto diff. for sens. w.r.t. $NN_T$ eq. (7)
15:        $AD(L \gets w_M,g_c \gets w_M,g_b \gets w_M,g_y \gets w_M) \to \nabla_{w_M}L$ ▷ auto diff. for sens. w.r.t. $NN_M$ eq. (8)
16:        $w_j+\Delta w_j(\nabla_{w_j}L) \to w_j$; $j=\{T,M\}$ ▷ AdaGrad optimizer step
17:        $t \gets t_0\mu^k$ ▷ Increment $t$; log-barrier term
18:        $k++$
19:    until $\|\Delta w\|<\epsilon^*$ ▷ Check for convergence
20:    return $\{z,A\}$
21:    Find nearest material and optimize for area ▷ section 3.4
  1. Area optimization: Using the optimal material from the first experiment, the areas are optimized by disregarding the decoder.

  1. Simultaneous material and area optimization: Here the material and areas are optimized simultaneously.

The results are reported in Table 3 for a cost constraint of C∗=$60C^*=\$60 and a safety factor FS=4F_S=4. Observe that scenario-1 (material optimization) leads to the lowest performance since the areas are not optimized. For this scenario, we also performed a brute-force search wherein the compliance was evaluated for all the materials; the brute-force search also resulted in ‘AISI 4130 norm’ as the best choice. In scenario-2, when the areas are optimized using ‘AISI 4130 norm’ as the material, the performance improved significantly. Finally, the best performance is achieved in scenario-3 when both material and areas are optimized simultaneously. Observe that the optimal material in scenario-3 differs from the one found in scenario-1.

ScenarioJ∗J^*Closest Material (confidence)Area (10−3m210^{-3}m^2)
13.79AISI 4130 norm (89.76%)[2, 2, 2, 2, 2, 2]
22.95AISI 4130 norm (89.76%)[2.2,2.6,2.6,3.2,1,1]
32.58AISI 1010 (90.69%)[3.6,3.1,3.1,3.5,1.3,1.2]

Table 3. Cost constrained optimization of the truss in Figure 11

For scenario-3, Table 4 lists the three closest materials and their confidence values.

Closest MaterialConfidence (γ\gamma)
1. AISI 101090.69%
2. 4032-T686.64%
3. AISI 102080.01%

Table 4. Three closest materials under cost constraint.

As an additional experiment, we replaced the cost constraint (eq. (3c)) in the above problem with a mass constraint:

gm:=(ρM∗∑k=1NAkLk)−1≤0(13)g_m := \left(\frac{\rho}{M^*}\sum_{k=1}^{N} A_k L_k\right)-1 \le0 \tag*{(13)}

with M∗=40kgM^* = 40\mathrm{kg}; all other parameters being the same. The results are presented in Table 5. Once again we observe that simultaneous optimization results in the best performance. Furthermore, when cost constrained was imposed, a steel alloy was chosen as the best material, whereas when a mass constrained is imposed, an aluminium alloy was chosen.

ScenarioJ∗J^*Closest Material (confidence)Area (10−3m210^{-3}\mathrm{m}^2)
110.5Al 380F die (91.63%)[2, 2, 2, 2, 2, 2]
28.75Al 380F die (91.63%)[3 2.5,2.4,3.3,1,1]
38.56Al 2018 (94.16%)[3.1,2.4,2.5,3.1,1,1]

Table 5. Mass constrained optimization of the truss in Figure 11

For scenario-3 under mass constraint, Table 6 lists the three closest materials and the confidence values.

Closest MaterialConfidence (γ\gamma)
1. Al 201894.16%
2. Al 380F die90.64%
3. Al 2014-T490.13%

Table 6. Three closest materials under mass constraint.

Material Database Subsets

While we have used the entire material database in the previous experiment, we now consider subsets of the database. Subsets of materials can often be justified based on experience. The VAE reconstruction error for the full database and three sample subsets is reported in table 7. As one expect, the VAE error reduces for the subsets.

We now carry out a few experiments using these subsets, and report the results. First, for the truss problem in Figure 11, we carry out a cost-constrained simultaneous optimization for the full database, as in the previous section. However, we

Material classNo. materialsΔE%\Delta E\%ΔC%\Delta C\%Δρ%\Delta\rho\%ΔY%\Delta Y\%
All Matls.925.06.83.37.1
Steel140.582.60.071.4
Aluminum520.090.220.042.9
Plastic122.30.360.50.7

caption: VAE reconstruction error trained with the full and various material subsets.

now report additional results in table 8. When all materials are considered (row-2), the compliance prior to snapping to the closest material is reported as J=2.25J = 2.25. The closest material is AISI 1010, resulting in a final compliance of J∗=2.58J^{*} = 2.58. As one can expect, J∗>JJ^{*} > J.

Next, since AISI 1010 is a steel alloy, we now consider only the Steel subset, and carry out simultaneous optimization. The results are reported in row-3. Observe that the raw compliance J=2.6J = 2.6 is surprisingly larger than J=2.25J = 2.25 in the previous row. This can be attributed to the larger VAE reconstruction error when all materials are considered, i.e., J=2.25J = 2.25 is not as reliable as the estimate J=2.6J = 2.6. Further, after the closest material is selected (AISI 1020), the final compliance is J∗=2.6J^{*} = 2.6. This is slightly larger than the value in the previous row. This discrepancy is analogous to rounding errors in integer optimization [16].

Material classJJOptimal materialEEρ\rhoYYJ∗J^{*}Area (10−3m210^{-3}m^{2})
All Matls.2.25AISI 10102E117.9E33.3E82.58[3.6,3.1,3.1,3.5,1.3,1.2]
Steel2.6AISI 10202E117.9E34.2E82.6[3.7,2.9,3.1,3.8,1.1,1.2]

caption: Material refinement with cost constraint.

Next, we repeated the above experiment for a mass constraint (instead of a cost constraint) and report the results in table 9. When all materials are considered (row-2), the algorithm converges to the aluminum class, with an optimal material Al 2018, as before. Further, J∗>JJ^{*} > J as expected. The results using the aluminum subset is reported in row-3. Now, the optimal material is AA356.0-F, and it differs from Al 2018. Now the performance J∗=8.38J^{*} = 8.38 has significantly improved compared to J∗=8.53J^{*} = 8.53.

Material classJJOptimal materialEEρ\rhoYYJ∗J^{*}Area (10−3m210^{-3}m^{2})
All Matls.8.21Al 20187.40E102.80E34.21E88.53[3.1,2.4,2.5,3.1,1,1]
Aluminum8.23AA356.0-F7.24E102.68E31.45E88.38[3.2,2.7,2.6,3.2,0.9,1]

caption: Material refinement with mass constraint.

Convergence

To study the convergence characteristics of simultaneous optimization, we consider the antenna tower illustrated in Figure 12. We impose a cost constraint of $50, factor of safety of FS=4FS = 4, and area bounds [10−9,10−2]m2[10^{-9}, 10^{-2}]m^{2}. Then the areas and material are simultaneously optimized, resulting in an optimal material of ‘AISI 4130 norm’ and J∗=32.5J^{*} = 32.5. The convergence in compliance is illustrated in Figure 13. We observe that the material choice converges early while the areas continue to be optimized; this was typical. The optimization took 1.89 seconds, with forward propagation and FEA, accounting for 16% each, while back-propagation consumed 54%.

Figure 12. Loading of 47-bar antenna tower.

Convergence of compliance for the material and area for loading in fig. 12

Figure 13. Convergence of compliance for the material and area for loading in fig. 12.

Conclusion

Engineers today have access to over 150, 000 materials [32], and this number is growing as new materials are being discovered [33, 34]. This presents a significant challenge to design engineers. In this paper, using variational auto encoders (VAEs), we proposed a generic method to simultaneously select the optimal material, while optimizing the geometry. The proposed method was demonstrated using trusses.

There are several limitations to the proposed method. A heuristic confidence-metric was used, as a final step, to snap from the continuous latent space to the nearest material. While this is a well known method in integer optimization, and was found to be effective in out study, it might lead to discrepancies in performance, as was noted in the numerical experiments. On the other hand, since the proposed method suggests an optimal set of material properties, it could drive material innovation in targeted applications. A very small material database with about 100 materials, with limited number of attributes, was considered here. A larger database and additional attributes needs to be explored. The method could benefit from better tuning of the networks, optimal choice of the NN architecture, and increased latent space.

Despite these limitations, we believe that the method holds promise. For example, it could be extended to topology optimization with material selection, specifically using the NN framework proposed in [35]. Furthermore, inclusion of thermo-elastic properties [36], auxetic properties [37], coating properties [38], global warming indices [13] can also be considered. Inclusion of uncertainty in material properties is also of significant interest. Finally, the method can be extended to the selection of discrete components such as springs, bolts, etc., and to the selection of discrete microstructures [39].

Compliance with ethical standards

The authors declare that they have no conflict of interest.

Reproduction of Results

The Python code pertinent to this paper is available at https://github.com/UW-ERSL/MaTruss. The material data was sourced from SolidWorks.

Acknowledgements

The authors would like to thank the support of National Science Foundation through grant CMMI 1561899.

References

  1. [1]Rudy Eggert. Engineering design. Pearson/Prentice Hall, 2005.
  2. [2]G. I. N. Rozvany, Bendsoe M.P., and U. Kirsch. Layout optimization of structures. Applied Mechanics Reviews, 48(2):41–119, 02 1995.DOI
  3. [3]W Achtziger. Truss topology optimization including bar properties different for tension and compression. Structural optimization, 12(1):63–74, 1996.DOI
  4. [4]Sourav Rakshit and GK Ananthasuresh. Simultaneous material selection and geometry design of statically determinate trusses using continuous optimization. Structural and Multidisciplinary Optimization, 35(1):55–68, 2008.DOI
  5. [5]Michael F Ashby and David CEBON. Materials selection in mechanical design. Le Journal de Physique IV, 3(C7):C7–1, 1993.
  6. [6]Michael F Ashby and Kara Johnson. Materials and design: the art and science of material selection in product design. Butterworth-Heinemann, 2013.
  7. [7]MF Ashby. Multi-objective optimization in material design and selection. Acta materialia, 48(1):359–369, 2000.DOI
  8. [8]A Jahan, MY Ismail, SM Sapuan, and F Mustapha. Material screening and choosing methods–a review. Materials & Design, 31(2):696–705, 2010.DOI
  9. [9]R Venkata Rao. A material selection model using graph theory and matrix approach. Materials Science and Engineering: A, 431(1-2):248–255, 2006.DOI
  10. [10]Chang-Chun Zhou, Guo-Fu Yin, and Xiao-Bing Hu. Multi-objective optimization of material selection for sustainable products: artificial neural networks and genetic algorithm approach. Materials & Design, 30(4):1209–1215, 2009.DOI
  11. [11]GK Ananthasuresh and MF Ashby. Concurrent design and material selection for trusses. 2003.
  12. [12]Mathias Stolpe and Krister Svanberg. A stress-constrained truss-topology and material-selection problem that can be solved by linear programming. Structural and Multidisciplinary Optimization, 27(1):126–129, 2004.DOI
  13. [13]Ernest Ching and Josephine V Carstensen. Truss topology optimization of timber–steel structures for reduced embodied carbon design. Engineering Structures, page 113540, 2021.DOI
  14. [14]Satadru Roy, William A Crossley, and Samarth Jain. A hybrid approach for solving constrained multi-objective mixed-discrete nonlinear programming engineering problems. 2021.DOI
  15. [15]JS Arora, MW Huang, and CC Hsieh. Methods for optimization of nonlinear problems with discrete variables: a review. Structural optimization, 8(2):69–85, 1994.DOI
  16. [16]Joaquim RRA Martins and Andrew Ning. Engineering design optimization. Cambridge University Press, 2021.DOI
  17. [17]J. Lee and S. Leyffer. Mixed Integer Nonlinear Programming. The IMA Volumes in Mathematics and its Applications. Springer New York, 2011.
  18. [18]Matthias Köppe. On the complexity of nonlinear mixed-integer optimization. In Mixed Integer Nonlinear Programming, pages 533–557. Springer, 2012.
  19. [19]Diederik P Kingma and Max Welling. An introduction to variational autoencoders. arXiv preprint arXiv:1906.02691, 2019.DOI
  20. [20]Liwei Wang, Yu-Chin Chan, Faez Ahmed, Zhao Liu, Ping Zhu, and Wei Chen. Deep generative modeling for mechanistic-based learning and design of metamaterial systems. Computer Methods in Applied Mechanics and Engineering, 372:113377, 2020.DOI
  21. [21]Xiang Li, Shaowu Ning, Zhanli Liu, Ziming Yan, Chengcheng Luo, and Zhuo Zhuang. Designing phononic crystal with anticipated band gap through a deep learning based data-driven method. Computer Methods in Applied Mechanics and Engineering, 361:112737, 2020.DOI
  22. [22]Tinghao Guo, Danny J Lohan, Ruijin Cang, Max Yi Ren, and James T Allison. An indirect design representation for topology optimization using variational autoencoder and style transfer. In 2018 AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, page 0804, 2018.
  23. [23]Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
  24. [24]Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural networks, 61:85–117, 2015.
  25. [25]Dassault Systèmes. Solidworks. http://www.solidworks.com, May 2021.
  26. [26]Adam Paszke, Sam Gross, Francisco Massa, Adam Lehrer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc., 2019.arxiv.org/abs/1912.01703
  27. [27]Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR, dec 2015.arxiv.org/abs/1412.6980
  28. [28]Krishnan Suresh. Design Optimization using MATLAB and SOLIDWORKS. Cambridge University Press, 2021.DOI
  29. [29]Hoel Kervadec, Jose Dolz, Jing Yuan, Christian Desrosiers, Eric Granger, and I Ben Ayed. Constrained deep networks: Lagrangian optimization via log-barrier extensions. CoRR, abs/1904.04205, 2(3):4, 2019.DOI
  30. [30]Aaditya Chandrasekhar, Saketh Sridhara, and Krishnan Suresh. Auto: a framework for automatic differentiation in topology optimization. Structural and Multidisciplinary Optimization, 64(6):4355–4365, Dec 2021.DOI
  31. [31]John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011.
  32. [32]Michael F. Ashby. Chapter 1 - introduction. In Michael F. Ashby, editor, Materials Selection in Mechanical Design (Fourth Edition), pages 1–13. Butterworth-Heinemann, Oxford, fourth edition edition, 2011.
  33. [33]Logan Ward, Ankit Agrawal, Alok Choudhary, and Christopher Wolverton. A general-purpose machine learning framework for predicting properties of inorganic materials. npj Computational Materials, 2(1):1–7, 2016.
  34. [34]Xiou Ge, Richard T Goodwin, Jeremy R Gregory, Randolph E Kirchain, Joana Maria, and Lav R Varshney. Accelerated discovery of sustainable building materials. arXiv preprint arXiv:1905.08222, 2019.arxiv.org/abs/1905.08222
  35. [35]Aaditya Chandrasekhar and Krishnan Suresh. Tourn: Topology optimization using neural networks. Structural and Multidisciplinary Optimization, 63(3):1135–1149, Mar 2021.DOI
  36. [36]Oliver Giraldo-Londoño, Lucia Mirabella, Livio Dalloro, and Glaucio H Paulino. Multi-material thermomechanical topology optimization with applications to additive manufacturing: Design of main composite part and its support structure. Computer Methods in Applied Mechanics and Engineering, 363:112812, 2020.
  37. [37]Koshi Takenaka. Negative thermal expansion materials: technological key for control of thermal expansion. Science and technology of advanced materials, 13(1), feb 2012.pubmed.ncbi.nlm.nih.gov/27877465
  38. [38]Anders Clausen, Niels Aage, and Ole Sigmund. Topology optimization of coated structures and material interface problems. Computer Methods in Applied Mechanics and Engineering, 290:524–541, 2015.
  39. [39]Yu-Chin Chan, Daicong Da, Liwei Wang, and Wei Chen. Remixing functionally graded structures: Data-driven topology optimization with multiclass shape blending. arXiv preprint arXiv:2112.00648, 2021.DOI

Paper details

vlnc:W_83msem6c
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents