Skip to main content

Table 1 A comparison of intelligent design methods as illustrated on Fig. 1

From: Intelligent designs in nanophotonics: from optimization towards inverse creation

Category

Methods

Refs. on meta-atoms

Design logic1

Model training phase

Structure design phase

    

Typical OF2

Training method

Training time

Typical OF3

Optimization method

Design time

Simulators4

Traditional optimizer

GA

[29–33]

Forward

N.A.

N.A.

N.A.

MSE

Non-gradient

Hours to days

Yes

 

PSO

[34]

Forward

N.A.

N.A.

N.A.

MSE

Non-gradient

Hours to days

Yes

 

TO

[35]

Inverse

N.A.

N.A.

N.A.

FoM

Gradient/non-gradient

Hours to days

Yes

Deep learning

MLP

[11, 27, 36–39]

Bidirection

MSE

Gradient descent

Hours to days

N.A.

N.A.

Seconds

No

 

AE

[40]

Inverse

MSE

Gradient descent

Hours to days

N.A.

N.A.

Seconds

No

 

VAE

[41]

Inverse

MSE

Gradient descent

Hours to days

N.A.

N.A.

Seconds

No

 

GAN

[42, 43]

Inverse

MSE

Gradient descent

Hours to days

N.A.

N.A.

Seconds

No

 

DQN

[44]

Inverse

MSE

Gradient descent

Hours to days

N.A.

N.A.

Seconds

No

  1. By case studies of a meta-atom, the algorithm workflow takes its structure distributions as input and its optical response (e.g., spectra, EM fields, etc.) as output. Here we generally divides the meta-atom development process into the model training phase and structure design phase. The DL methods frequently contain both phases, while only the latter exist in traditional optimizations. Aberrations: GA genetic algorithm, PSO particle swarm optimization, TO topological optimization, MLP multilayer perception, GAN generative adversarial network, VAE variational autoencoder, DQN double deep Q-learning, OF objective function, MSE mean square error, FoM Figure of Merit, N.A. Not applicable
  2. 1Design logic: the forward design starts from the input to the output, while the inverse design indicates the backwards direction, i.e., from the output towards the input. Note that the meta-atom geometries are commonly adopted as the input, while the demanded optical response is designated the output
  3. 2Typical OF over model training phase: the MSE is provided here for reference only and more detailed evaluation formulas should be consulted in corresponding references
  4. 3Typical OF over structure design phase: the terms (MSE and FoM) present here are for reference only and the specific objective function used for structure designs could be varied under different cases
  5. 4Simulators: this term indicates whether additional EM solvers are employed over the entire structure design phase