Method for designing multi-layer optical structure and electronic device for performing the same

Patent metadata

Field	Value
Publication number	US-12164104-B2
Application number	US-202017112556-A
Country	US
Kind code	B2
Filing date	Dec 4, 2020
Priority date	Dec 4, 2020
Publication date	Dec 10, 2024
Grant date	Dec 10, 2024

Abstract

A method for designing a multi-layer optical structure includes: obtaining candidate multi-layer optical structures through a sequence generator; obtaining a candidate spectrum for each candidate multi-layer optical structure; obtaining a difference between the candidate spectrum and a target spectrum; updating sequence generator parameters through reinforcement learning training and iteratively performing the obtainings and the updating in response to a first termination condition being not met; and selecting one of all obtained candidate structures to be a target multi-layer optical structure in response to the first termination condition being met. The difference between a spectrum of the target multi-layer optical structure and the target spectrum is minimized through the process. The method of the present application can perform the designs robustly and effectively.

First claim

What is claimed is:

1. A method for designing a multi-layer optical structure, comprising:
obtaining a plurality of candidate multi-layer optical structures through a sequence generator based on at least one parameter of the sequence generator, wherein each of the plurality of candidate multi-layer optical structures has a candidate sequence of materials and a candidate thickness for each of the materials;
obtaining a candidate spectrum for each of the plurality of candidate multi-layer optical structures;
obtaining a difference between the candidate spectrum and a target spectrum;
determining whether a first termination condition being met;
updating the at least one parameter of the sequence generator through reinforcement learning training and reperforming the obtainings and the updating if the first termination condition is not met; and
selecting one candidate multi-layer optical structure whose spectrum has a minimal difference from the target spectrum among all obtained candidate multi-layer optical structures to be a target multi-layer optical structure if the first termination condition is met; and
wherein the sequence generator comprises a first unit, a second unit, and a third unit; and
the obtaining a plurality of candidate multi-layer optical structures through a sequence generator based on at least one parameter of the sequence generator, comprises:
obtaining, by the first unit, a hidden state for a current layer of one of the plurality of candidate multi-layer optical structures based on a hidden state, a material, and a thickness of a previous layer;
obtaining, by the second unit, a material for the current layer based on the hidden state for the current layer;
obtaining, by the third unit, a thickness for the current layer based on the hidden state for the current layer and the material for the current layer;
reperforming the obtainings to obtain the one of the plurality of candidate multi-layer optical structures until a second termination condition being met; and

the obtaining, by the second unit, a material for the current layer based on the hidden state for the current layer, comprises: removing a material of the previous layer from a plurality of materials of the current layer; and
the removing a material of the previous layer from the plurality of materials of the current layer, comprises:
obtaining, by the second unit, probability distributions of a plurality of materials for the current layer based on the hidden state for the current layer;
multiplying the probability distributions by a non-repetitive gating function, which sets a probability of the material for the previous layer to be 0;
selecting a material from the remaining of the plurality of materials to be the material for the current layer.

2. The method according to claim 1 , wherein
the number of times for reperforming the obtainings and the updating is predefined; and
the first termination condition is the number of times being reached or a difference between a candidate spectrum of one of the obtained plurality of candidate multi-layer optical structures and the target spectrum being less than a predefined threshold.

3. The method according to claim 1 , wherein the obtaining, by the second unit, a material for the current layer based on the hidden state for the current layer, comprises:
obtaining probability distributions of a plurality of materials for the current layer of one of the plurality of candidate multi-layer optical structures; and
obtaining the material for the current layer based on the probability distributions of the plurality of materials.

4. The method according to claim 1 , wherein the obtaining, by the third unit, a thickness for the current layer based on the hidden state for the current layer and the material for the current layer, comprises:
obtaining probability distributions of a plurality of thicknesses for the current layer of one of the plurality of candidate multi-layer optical structures; and
obtaining the thickness for the current layer based on the probability distributions of the plurality of thicknesses.

5. The method according to claim 1 , wherein
the number of layers for each of the plurality of candidate multi-layer optical structures is predefined; and
the second termination condition is the number of layers being reached, or an end-of-sequence (EOS) token being obtained by the second unit.

6. The method according to claim 1 , wherein the updating the at least one parameter of the sequence generator through reinforcement learning training, comprises:
obtaining a gradient for updating the at least one parameter through a proximal policy optimization (PPO) algorithm.

7. The method according to claim 1 , wherein the reinforcement learning training is performed based on a reward value of each of the plurality of candidate multi-layer optical structures, and the reward value is obtained by subtracting the difference between the candidate spectrum and the target spectrum from 1 .

8. The method according to claim 1 , wherein the selecting one of all obtained candidate multi-layer optical structures to be a target multi-layer optical structure, comprises:
finetuning the selected candidate multi-layer optical structure to obtain the target multi-layer optical structure through a quasi-Newton method.

9. An electronic device, comprising a processor and a non-transitory memory, wherein computer programs are stored in the non-transitory memory, and the computer programs are executed by the processor to perform operations of:
obtaining a plurality of candidate multi-layer optical structures through a sequence generator based on at least one parameter of the sequence generator, wherein each of the plurality of candidate multi-layer optical structures has a candidate sequence of materials and a candidate thickness for each of the materials;
obtaining a candidate spectrum for each of the plurality of candidate multi-layer optical structures;
obtaining a difference between the candidate spectrum and a target spectrum;
determining whether a first termination condition being met;
updating the at least one parameter of the sequence generator through reinforcement learning training and reperforming the obtainings and the updating if the first termination condition is not met; and
selecting one candidate multi-layer optical structure whose spectrum has a minimal difference from the target spectrum among all obtained candidate multi-layer optical structures to be a target multi-layer optical structure if the first termination condition is met;
wherein the sequence generator comprises a first unit, a second unit, and a third unit; and
when obtaining a plurality of candidate multi-layer optical structures through a sequence generator based on at least one parameter of the sequence generator, the computer programs are further executed by the processor to perform operations of:
obtaining, by the first unit, a hidden state for a current layer of one of the plurality of candidate multi-layer optical structures based on a hidden state, a material, and a thickness of a previous layer;
obtaining, by the second unit, a material for the current layer based on the hidden state for the current layer;
obtaining, by the third unit, a thickness for the current layer based on the hidden state for the current layer and the material for the current layer;
reperforming the obtainings to obtain the one of the plurality of candidate multi-layer optical structures until a second termination condition being met; and

when obtaining, by the second unit, a material for the current layer based on the hidden state for the current layer, the computer programs are further executed by the processor to perform and op

Assignees

• Ningbo Inlight Tech Co Ltd
• The Regents Of The Univ Of Michicgan
• Univ Michigan Regents

Inventors

• Wang Haozhu
• Ji Chengang
• Guo Lingjie Jay

Classifications

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F30/10

  Geometric CAD · CPC title

• G06F30/27

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title