Image-Based Multiresolution Topology Optimization Using Deep Disjunctive Normal Shape Model
Introduction
Topology optimization has been extensively used as a popular computational design tool for generating optimal structural layouts. This powerful tool enables the creation of innovative structures with complex geometry that may not be realizable with traditional approaches such as shape or size optimization [1], [2]. This approach however suffers from significant computational cost especially when a detailed structural design is sought. In such cases, typically, many large scale finite element analyses need to be performed to reach to a satisfactory design which poses a significant challenge in terms of computational cost. Finding such high resolution designs in scenarios where the simulation parameters are variable is even more onerous.
In this paper, we adopt a scientific visualization/machine learning (ML) approach to overcome the computational challenges for large scale topology optimization, especially when optimal high-fidelity designs are required for variable simulation parameters. In particular, our approach utilizes multifidelity designs where mainly inexpensive low-fidelity designs in addition to judiciously chosen high-fidelity designs are used for training. Using the learning machine which encodes the relationship between the simulation parameters and the shape space we approximate optimized designs on parameters that are not previously considered in the simulations. This approach will enable the designers to efficiently traverse the simulation parameters and find the best solutions which satisfy the design objectives and criteria.
The result of this work contributes to the development of an efficient visualization tool for exploring optimized designs with variable parameters. Ideally the visualization tool is a platform that provides high resolution optimized designs interactively and this work is one stepping stone toward this objective. With such a tool, a designer could navigate through the design space interactively, assessing multiple designs qualitatively (i.e. visually) and quantitatively (through various evaluation metrics build into the tool).
In the context of engineering design, similar efforts have been devoted to embed the high-dimensional design space into semantic compact subspaces [3], [4] using a manifold clustering procedure and deep generative models. The idea of visual parameter space analysis has also been explored [5], [6] where visual interactive tools are introduced to navigate the data spaces or the space of quantities of interest as a function of reduced-dimensional parameter space. Similar to these efforts, several researchers have introduced visual analysis techniques for exploration of parameter-data space map in multi-criteria decision making [7], medical image analysis [8], [9], geoscientific simulation models [10], decision support in flood management [11], [12], inspection of fastening bolts on freight trains [13], etc. Our work in this paper can be viewed as a visual analysis technique which adopts a powerful function approximator, i.e. deep neural network to map the simulation parameters to the high-dimensional space of topology optimized designs.
Deep learning (DL) is known as a subset of ML. A large number of layers in the architecture of DL models enable them to estimate complex functions. However, there is a need for large datasets to train such complex models. In other words, DL fundamentally is neural networks that learn from large amounts of data with the use of various layers. An important class of DL, convolutional neural networks introduced in [14], are widely used in image processing. DL has also been employed in many different areas such as video processing, speech recognition, natural language processing, and computer vision [15], [16], [17], [18].
In this paper, we present a novel deep learning approach in conjunction with the disjunctive normal shape model (DNSM) for the generic topology optimization problem. We refer to this approach as Deep DNSM which is abbreviated as DDNSM. The DNSM is a shape model that first was introduced by Ramesh et al. [19] and reformulated by Mesadi et al. [20] to be used in a Bayesian framework for image segmentation tasks. The DNSM has also been used in conjunction with a convolutional neural network by Javanmardi et al. [21] for segmenting shapes with low quality and low signal-to-noise ratio. The main idea of DNSM is to model shapes as the disjunction of convex polytopes with each polytope as a conjunction of half-spaces. Inspired by [21] we have designed a new framework that uses the DNSM on top of a fully connected deep neural network as a multiresolution network with both low resolution and high resolution designs to be trained jointly. We provide more details on both DNSM and DDNSM in Section 3.
Structural topology optimization has evolved significantlysince the time it first introduced. A comprehensive review of current and future trends in this field has recently been published [22]. Among these works many researchers have addressed the problem of parametric topology optimization and topology optimization under uncertainty where the main focus is to generate robust and reliable designs [23], [24], [25], [26], [27], [28]. Such problems pose a similar challenge to the one in this paper where the simulation parameters are variable. In the same vein, several research works have also attempted to tackle the problem of data-driven topology optimization and proposed novel approaches for prediction improvement in such data-driven approximations [29], [30]. A number of researchers have proposed to use computational capabilities of a Graphics Processing Units (GPU) for large scale topology optimization where the main goal is to achieve optimal high resolution structures with minimal process time [31], [32], [33]. These efforts are in line with approaches which use machine learning techniques for large scale problems where powerful computational resources are required.
Machine learning techniques have achieved compelling results in data-based approximations such as image processing, pattern recognition, finance, etc. [34], [35], [36]. Complementing these efforts, novel techniques have been proposed for prediction/approximation with physics-based or simulation-based learning machines. The main goal in these works is to predict physics-based simulations which are typically governed by partial differential equations (PDE) [37], [38]. This type of approaches has close connection to the topology optimization problem where the solution of the optimization problem is sought for a physical systems described with PDEs.
Topology optimization using machine learning is a relatively new research direction. As examples, authors in [39], [40], [41] use generative modeling techniques for topology optimization where they use variational auto encoders (VAE) and generative adversarial networks (GANs) to predict the optimized topology designs. A similar work [30] proposes the use of Principal Component Analysis (PCA) and a fully connected neural network to learn the mapping between loading configurations and optimal topologies. Authors in [42] use convolutional encoder–decoder architecture to solve the topology optimization problem by posing it as an image segmentation task. Finally authors in [43] introduce a theory-driven learning mechanism based on the optimality criteria in topology optimization for generation of near-optimal topology designs.
It is also noteworthy to mention approaches based on multiresolution analysis in the context of shape and topology optimization [44], [45], [46], [47], [48], [49], [50], [51], [52]. The main theme among these works is to leverage the computational efficiency of low resolution models to obtain more expensive high resolution designs. Our multiresolution approach in this paper is indeed similarly motivated by reducing the computational cost; however unlike the aforementioned references which consider a deterministic setting, our strategy is devised to incorporate parametric variations in design such as variations in load and boundary conditions [28], [53]. Another important line of research which is relevant to this paper is the incorporation/representation of geometric features in topology optimization which is done via combining free-form topology optimization with embedded components/holes or representing the structure via union of geometric primitives, e.g. rectangles with straight or semicircular ends [54], [55], [56], [57], [58], [59], [60], [61], [62], [63], [64], [65], [66], [67], [68], [69]. We however remark that our approach in this paper, DDNSM which is an effective way to represent binary images is solely used to learn from and subsequently approximate the topology optimized designs as binary images. We note that we do not use this approach for the task of topology optimization (or representing the geometry) itself; we instead use a well-known density-based approach to generate the optimized designs [70].
In this paper, we mainly focus on the computational challenges associated with the high resolution designs with variable parameters. To this end we propose an algorithm which utilizes a limited number of high resolution images corresponding to previously found optimized designs in conjunction with a large number of inexpensive low resolution optimized topology images for network training and generate a learning machine which predicts the near-optimal high resolution images for previously unseen simulation parameters. Our work contributes to the existing literature on topology optimization using machine learning by providing a novel approach which connects topology optimization with scientific visualization. In particular our work can be considered among the few recent studies which leverage deep learning for predicting near optimal designs [41], [43]. These approaches are particularly useful when there is variation in design parameters and could be effective for fast exploration of optimal design manifold. We remark that in this paper we provide a systematic computational cost analysis which serves as a practical guideline to assess the learning cost in comparison with the cost of direct high resolution designs. We also remark that our deep learning approach is, in essence, different from the zero-order optimization approaches (or population-based approaches) such as genetic programming. Our approach similarly to any deep learning approach is based on the minimization of a loss function which is performed via a gradient-descent approach. As mentioned earlier we also use a gradient-based approach to generate the training data which renders our approach entirely gradient-based [71].
The organization of this paper is as follows. In Section 2 we briefly discuss the topology optimization problem, in particular the compliance minimization. We also discuss the details of data generation for network training in this section. Section 3 discusses the details of DNSM and DDNSM approach in conjunction with an algorithm for selection of important high resolution samples. Section 4 presents numerical results on topology design of an L-bracket elastic structure, a heat sink structure, and a 3D linear elastic structure. Finally Section 5 discusses the concluding remarks.
Section snippets
Notation
The following notations are frequently used in this article:
- •
Bold faced characters denote vectors and matrices e.g. denotes the vector of displacements.
- •
We use superscript to denote the resolution e.g. denotes low resolution shape, and use subscripts to denote parametric samples e.g. is a low resolution shape with sample index . We also use subscripts to denote the dimension in the parameter space e.g. is the first parametric variable. Different uses of subscripts are clear from the
Deep disjunctive normal shape model
In this section, we explain the Disjunctive Normal Shape Model (DNSM) and its incorporation on top of a fully connected deep neural network, termed as Deep Disjunctive Normal Shape Model (DDNSM). We also discuss details of our importance sampling strategy in this section.
L-Bracket
The L-shape structure is frequently used as a benchmark example in the structural topology optimization studies [77], [78]. In this example, we use this type of structure to generate an image dataset. Fig. 7 shows the geometry, boundary conditions and the loading parameters that are used in this example. The domain is discretized using 64 × 64, 32 × 32 and 16 × 16 standard square finite elements for high, medium and low resolution optimizations. We assume a plane stress condition for the
Conclusion
We presented a machine learning approach to generate high resolution topology designs from multiresolution images. Our approach used a novel scientific visualization method based on deep disjunctive normal shape model (DDNSM). The learning process involves simulation parameters and multiresolution images associated with each simulation parameter. Learning the map between the parameters and multiresolution images, the DDNSM generates near optimal high resolution designs on unknown simulation
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work was sponsored by The Defense Advanced Research Projects Agency, USA TRADES Award HR0011-17-2-0016.
References (81)
- et al.
Generating optimal topologies in structural design using a homogenization method
Comput Methods Appl Mech Engrg
(1988) - et al.
Automated tracing of microtubules in electron tomograms of plastic embedded samples of Caenorhabditis elegans embryos
J Struct Biol
(2012) Deep learning in neural networks: An overview
Neural Netw
(2015)- et al.
Risk-averse structural topology optimization under random fields using stochastic expansion methods
Comput Methods Appl Mech Engrg
(2018) - et al.
Topology optimization under uncertainty via non-intrusive polynomial chaos expansion
Comput Methods Appl Mech Engrg
(2017) - et al.
Gradient based design optimization under uncertainty via stochastic expansion methods
Comput Methods Appl Mech Engrg
(2016) - et al.
Efficient topology optimization using gpu computing with multilevel granularity
Adv Eng Softw
(2017) - et al.
Large-scale robust topology optimization using multi-gpu systems
Comput Methods Appl Mech Engrg
(2016) - et al.
Machine learning of linear differential equations using gaussian processes
J Comput Phys
(2017) - et al.
Non-iterative structural topology optimization using deep learning
Comput Aided Des
(2019)
One-shot generation of near-optimal topology through theory-driven machine learning
Comput Aided Des
Shape optimisation with multiresolution subdivision surfaces and immersed finite elements
Comput Methods Appl Mech Engrg
Boundary element based multiresolution shape optimisation in electrostatics
J Comput Phys
Multi-resolution multi-scale topology optimization — a new paradigm
Int J Solids Struct
A multi-resolution approach for multi-material topology optimization based on isogeometric analysis
Comput Methods Appl Mech Engrg
Topology optimization with closed b-splines and boolean operations
Comput Methods Appl Mech Engrg
Sensitivity analysis with the modified heaviside function for the optimal layout design of multi-component systems
Comput Methods Appl Mech Engrg
Explicit layout control in optimal design of structural systems with multiple embedding components
Comput Methods Appl Mech Engrg
Engineering feature design for level set based structural optimization
Comput Aided Des
Explicit structural topology optimization based on moving morphable components (mmc) with curved skeletons
Comput Methods Appl Mech Engrg
A geometry projection method for continuum-based topology optimization with discrete elements
Comput Methods Appl Mech Engrg
Topology optimization of non-linear elastic structures and compliant mechanisms
Comput Methods Appl Mech Engrg
Failure mitigation in optimal topology design using a coupled nonlinear continuum damage model
Comput Methods Appl Mech Engrg
Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
J Comput Phys
Topology optimization: Theory, methods and applications
Design manifolds capture the intrinsic complexity and dimension of design spaces
J Mech Des
Tuner: Principled parameter finding for image segmentation algorithms using visual response surface exploration
IEEE Trans Vis Comput Graphics
Visual parameter space analysis: A conceptual framework
IEEE Trans Vis Comput Graphics
Weightlifter: Visual weight space exploration for multi-criteria decision making
IEEE Trans Vis Comput Graphics
Open-box spectral clustering: Applications to medical image analysis
IEEE Trans Vis Comput Graphics
A visual analysis concept for the validation of geoscientific simulation models
IEEE Trans Vis Comput Graphics
Many plans: Multidimensional ensembles for visual decision support in flood management
Comput Graph Forum
Run watchers: Automatic simulation-based decision support in flood management
IEEE Trans Vis Comput Graphics
Automated status inspection of fastening bolts on freight trains using a machine vision approach
Proc Inst Mech Eng F
Convolutional networks for images, speech, and time series
Handb Brain Theory Neural Netw
Deep learning
Nature
Imagenet classification with deep convolutional neural networks
Very deep convolutional networks for large-scale image recognition
Disjunctive normal shape models
Cited by (0)
- 1
These authors contributed equally to this work.