Table 3 The AI algorithms used for the prediction of mechanical metamaterials properties and their inverse design

From: Mechanical metamaterials and beyond

AI algorithms

Advantages

Disadvantages

Prediction

Inverse design

Type

Ref.

Type

Ref.

Artificial neural networks

• Flexibility and scalability

• Computationally efficient

• Parallel processing

• Powerful ability to extract features in data

• Accuracy relies on amount of data

• Prone to overfitting

• Black box nature

Nonlinear mechanical metamaterials and fractal metamaterials

157,193

Inflatable soft membranes

158

Deep learninga

• Powerful ability to extract features in data

• Handling complex Data

• Parallel processing

• Flexibility and scalability

• Computationally intensive

• Accuracy relies on large amount of image data

• Prone to overfitting

• Black box nature

Copper spheres embedded in polylactide matrices

194

2D and elastic mechanical metamaterials

153,195

Tetra-chiral auxetics, cellular metamaterials

196,197

Gradient mechanical metamaterials

155,198

Magneto-mechanical metamaterials and auxetic kirigami metamaterials

199,200

Metasurfaces, magneto-mechanical metamaterials and auxetic mechanical metamaterials

154,199,201

Evolutionary strategyb

• Exceling at global optimization

• Scalability and invariance

• Built-in feature selection

• High interpretability

• Robustness to noise

• Less susceptible to overfitting

• Computationally intensive

• Fixed standard deviation parameter of noise

• Slow search speed

Nanoscale corrugated plates

202

2D and 3D mechanical metamaterial with nonlinear response, fractal metamaterials

157,193,203

Genetic programming

• Global search ability

• Scalability

• Simple process

• Built-in feature selection

• High interpretability

• Robustness to noise

• Computationally intensive

• Complicated programming implementation

• Slow search speed

Graphene origami metamaterials

204,205

Auxetic mechanical metamaterials with zero Poisson’s ratio

206

Bayesian network classifiers

• High learning efficiency

• Small time and space overhead in classification

• Computational complexity

• Dimensional challenges in computing probability

• Low interpretability

--

--

Mechanical metamaterials with negative stiffness

151

Decision trees

• Simple data preparation

• High interpretability

• High efficiency and accuracy

• Prone to overfitting

• Bias toward features with more levels

• Difficulty in handling missing data

Non-rigid square-twist origami

207

--

--

  1. aDeep learning is a specialization of artificial neural networks with multiple hidden layers.
  2. bEvolutionary strategy and genetic programming are branches of evolutionary computation.