Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

• Mechanical metamaterials are used to solve inverse problems.
• Inverse problems involve finding microarchitectures that give desired properties.
• Additive manufacturing techniques have limited resolution.
• Multi-objective inverse design problem is difficult to solve.
• Deep-DRAM combines four decoupled models to solve the problem.
• Deep-DRAM finds many solutions to the multi-objective inverse design problem.
• Filtering step is used to identify designs with minimum peak stresses.

Paper Content

Introduction

• Emergence of architected materials with bespoke properties in 21st century
• Computational models used to establish design-property relationships
• Additive manufacturing techniques enable realization of complex designs
• Metamaterials used in medical devices, soft robotics, and other advanced areas of research
• Inverse design problem of finding microarchitecture resulting in specific set of elastic properties is challenging
• Homogenized solutions only valid at convergence limits and may deviate from actual elastic properties
• Deep learning models and deep generative models used to tackle size-agnostic inverse design problem
• Nonlinear inverse problems relevant to microarchitecture design of mechanical metamaterials
• Non-uniqueness of solution, need to find multiple solutions
• Deep generative models used to enable investigations of structure-response relationship
• Modular approach (Deep-DRAM) used to solve size-agnostic design problem with additional requirement of minimizing maximum stresses
• Data generated using finite element models, actual mechanical metamaterials 3D printed and tested

Results and discussion

Elastic properties of rn unit cells

• Considered RN unit cells composed of 16 nodes
• Allowed for a broader range of elastic properties
• Higher chance of extreme negative and extreme positive Poisson’s ratios
• Possible to generate unit cells with average nodal connectivity values of 2.5, 3, 3.5, 4, and 4.5
• Number of beam-like elements in the RN unit cells varied between 20 and 42
• Estimated number of unit cells that can be generated ≈ 9.22 × 10 11
• Average simulation time per design ≈ 5.42 × 10 −4 s
• Takes ≈ 5761 days to perform FE analysis on all these RN unit cells
• Performed FE analysis on 6 million randomly generated RN unit cells
• Elastic properties cover a cone-like region with a range of (0 , 0.25) and (−1.5 , 1.2)
• Trained a DL model to predict the four elastic properties of any RN unit cell
• Model performance had a coefficient of determination ( 2 ) of >0.993 and >0.999
• Evaluation of the trained model had an overall coefficient of determination ( 2 ) of 0.997
• Evaluation of the DL model takes ≈ 2.44 × 10 −5 s per design

Generative inverse design framework

• Trained a deep generative model based on CVAE
• Latent space size chosen to be 8
• Selected two hidden layers with 512 and 260 neurons
• ReLU and Sigmoid activation functions for output layer
• Overall coefficient of determination (2) of 0.865
• Overall 2 of 0.977, MAE of 1.2 × 10−2, and MSE of 3.0 × 10−4
• DL-predicted elastic properties of generated unit cells compared with FE results
• 2 of 0.98, 0.98, 0.99, and 0.99 for 12, 21, 11, and 22 respectively

Unit cells with requested rare elastic properties

• Generative model can create grid-sampled values of double-auxetic elastic properties
• Model can generate 100 designs for each input
• Generative model expands envelope of achievable elastic properties
• Generative model can generate unit cells with predefined elastic properties, including double-auxeticity
• Case studies show high accuracy of generative model

Elastic properties of combinatorial designs

• Studied combinatorial designs composed of 1x2 repetitions of RN unit cells
• Assumed 1 and 2 are values varying in a range of (2, 20)
• Studied a total of 100 combinatorial designs from each RN unit cell
• Used undersampled dataset of RN unit cells to perform numerical simulations
• Trained a forward predictor called “size-agnostic model”
• Model had an overall R2 of 0.995 for the test dataset
• Combinatorial designs showed a wide range of elastic properties
• Nonlinear relationship between elastic properties and dimensions of combinatorial designs
• Selected 1=2=20 as maximum size of combinatorial designs
• Validated results of simulations with mechanical tests on specimens

Inverse design of lattice structures with requested elastic properties and dimensions

• Developed a framework called Deep-DRAM to inversely design lattice structures with given elastic properties and dimensions
• Used a deep generative model to generate 10 5 RN unit cells
• Unit cells have natural variations in elastic properties
• Size-agnostic model used to predict elastic properties of generated unit cells
• MSE values calculated to find designs that best match target elastic properties
• Constant value for elastic modulus and range of Poisson’s ratios used
• 1.96 × 10 7 combinatorial designs generated and filtered
• Heat maps of MSE values used to quantify expected error values for different elastic properties and dimensions
• Gray regions represent designs with MSE values exceeding acceptance threshold
• Envelopes of successful design generations converge for large enough values of 1 and 2

Stress distribution

• Deep-DRAM provides solutions to the design problem of finding RN lattice structures with pre-defined dimensions and elastic properties.
• It is possible to apply additional design requirements, such as criteria regarding the stress distributions observed within the generated structures.
• Three groups of combinatorial designs with predefined specifications were studied and the first 1000 designs with MSE < 0.1 were further analyzed using FE simulations.

Discussions and future outlook

• Deep-DRAM framework is a combination of four modules
• Modular approach allows modules to be combined with other tools
• Modularity and probabilistic nature of CVAE allows multi-objective design optimization
• Generating and evaluating individual designs is in the range of micro-seconds
• Modular design and ad hoc combination with direct FE modeling affords flexibility

Conclusions

• Developed size-agnostic inverse design framework, Deep-DRAM
• Generates RN lattice structures with predefined elastic properties and dimensions
• Combines deep generative models with forward predictors
• Generates bespoke mechanical metamaterials
• Satisfies additional design requirements, such as minimum peak stresses
• FE models created using MATLAB
• Design structures by randomly connecting each node to its surrounding nodes
• Graph-based search method used to filter and discard invalid unit cell designs
• 900 times faster than image-based filtering method
• RN unit cells incorporated into combinatorial designs
• Three-node quadratic beam elements with rectangular cross-sections
• Elastic material with Young’s modulus of 0.6 MPa and Poisson’s ratio of 0.3
• Two FE models created to apply strain of 5% along 1- and 2-directions
• Normal stresses divided by strain to calculate elastic moduli and Poisson’s ratio
• Unit cell elastic properties model trained to learn mapping from RN unit cell designs to elastic properties
• Pipeline training used to select options and parameters for data analysis and hyperparameter tuning
• MSE used as loss function to ensure regression line changes only slightly for modest change in data point
• Coefficient of determination used to assess overall performance of model
• Deep generative model based on CVAE
• Recognition and reconstruction networks coupled in encoder-decoder format
• Loss function of CVAE modified to incorporate conditional information
• Pre-trained unit cell elastic properties model incorporated in overall loss function
• Adam optimizer preselected
• t-SNE used to reduce latent space dimension to two
• Elastic properties manually classified into three classes
• Confusion matrix used to assess classification accuracy of model
• 1 score used to evaluate RN reconstruction
• 2 used to evaluate predicted elastic responses