Generative machine learning algorithm for lattice structures with superior mechanical properties

Abstract

Lattice structures are typically made up of a crisscross pattern of beam elements, allowing engineers to distribute material in a more structurally effective way. However, a main challenge in the...

Overview of Methods and Software for the Design of Functionally Graded Lattice Structures

“…A typical FGLS design workflow using ML is presented in ref. [179] Following are some examples of the use of ML in the design or optimization of lattice structures.…”

Inverse design of truss lattice materials with superior buckling resistance

“…From the computational side, deep-learning models 20 , specifically, represent a promising alternative to physics-based models for materials design, providing much faster (several orders of magnitude) yet accurate structure-property relationship predictions, thus allowing efficient design space exploration [21][22][23][24][25][26][27][28] . From composites [29][30][31][32][33] , through complex symmetric architectured materials 34 , stretchable kirigami-inspired-cut materials 35,36 , spinodoid metamaterials 37 , up to polycrystalline solids 38 , several studies have attempted to solve the inverse design problem exploiting the powerful computational and predictive capabilities provided by deep-learning techniques, mainly deep neural networks, used either as generative 31,36,39,40 or surrogate forward models coupled with other optimization methods 32,41 (e.g., evolutionary algorithms). Yet, only a few studies have provided solutions for the inverse design of truss lattice materials, mainly focusing on (i) pre-existing architectures conveniently modified to obtain lattices with desired properties, such as tunable stiffness anisotropy 42 and stronger micro-lattices with arranged defects 43 ; (ii) single complex 3D novel unit cells with load carrying applications, such as stronger lattice cores for sandwich structures 44,45 ; (iii) basic architectures (such as a square lattice), on which reinforcements are non-uniformly added in order to match the desired mechanical response from a database 46 ; (iv) targeted linear properties (in solid mechanics terms), such as lattice' stiffness 39,41,42,47 and Poisson's ratio 48 .…”

“…From composites [29][30][31][32][33] , through complex symmetric architectured materials 34 , stretchable kirigami-inspired-cut materials 35,36 , spinodoid metamaterials 37 , up to polycrystalline solids 38 , several studies have attempted to solve the inverse design problem exploiting the powerful computational and predictive capabilities provided by deep-learning techniques, mainly deep neural networks, used either as generative 31,36,39,40 or surrogate forward models coupled with other optimization methods 32,41 (e.g., evolutionary algorithms). Yet, only a few studies have provided solutions for the inverse design of truss lattice materials, mainly focusing on (i) pre-existing architectures conveniently modified to obtain lattices with desired properties, such as tunable stiffness anisotropy 42 and stronger micro-lattices with arranged defects 43 ; (ii) single complex 3D novel unit cells with load carrying applications, such as stronger lattice cores for sandwich structures 44,45 ; (iii) basic architectures (such as a square lattice), on which reinforcements are non-uniformly added in order to match the desired mechanical response from a database 46 ; (iv) targeted linear properties (in solid mechanics terms), such as lattice' stiffness 39,41,42,47 and Poisson's ratio 48 . To solve the inverse-design problem of lattice materials, we propose a bottom-up approach using a fully connected deep neural network (DNN) (as a decider) in conjunction with a genetic algorithm (as a sampler) to search for optimal architecture candidates, whose engineering applications are later verified by finite element (FE) simulations and experiments on 3D-printed lattices.…”

“…To solve the inverse-design problem of lattice materials, we propose a bottom-up approach using a fully connected deep neural network (DNN) (as a decider) in conjunction with a genetic algorithm (as a sampler) to search for optimal architecture candidates, whose engineering applications are later verified by finite element (FE) simulations and experiments on 3D-printed lattices. It is worth underlying that the idea of combining an ML model with a genetic algorithm to solve inverse problems is not new in general 24,32,41,49 . In solid mechanics, specifically, strong yet tough polycrystalline materials subject to complex fracture mechanisms 38,50 (e.g., crack branching) as well as stiff and strong digital composites 32 have been inverse-designed by exploiting deep learning models as efficient surrogate solvers combined with genetic algorithms.…”

Inverse Design of Digital Materials Using Corrected Generative Deep Neural Network and Generative Deep Convolutional Neural Network

“…Inspired by generative ML models such as autoencoders and generative adversarial networks, researchers at the University of California, Berkeley, developed generative ML algorithms including generative inverse design networks [ 18 ] and hybrid neural network and genetic optimization [ 22 ] for inverse material design. Generative deep neural networks (GDNN) effectively avoid problems associated with topology optimization: the local minima are mitigated by random initialization of the input; gradients are calculated rapidly by back‐propagation; and the optimization target is the output of the neural network.…”