Inverse binary optimization of convolutional neural network in active learning efficiently designs nanophotonic...

Area of Science:

• Computational materials science
• Nanophotonics engineering
• Machine learning applications

Background:

• Active learning schemes are used for optimizing designs in nanophotonics and materials.
• Factorization machines (FM) with quantum annealing (QA) are current methods, but struggle with complex design spaces due to their second-order nature.

Purpose of the Study:

• To introduce an Inverse Binary Optimization (IBO) scheme using convolutional neural networks (CNNs) as a more effective surrogate model for active learning in binary optimization.
• To evaluate the performance of the CNN-IBO framework in optimizing complex nanophotonic structures.

Main Methods:

• Developed an Inverse Binary Optimization (IBO) scheme utilizing a pre-trained Convolutional Neural Network (CNN) as a surrogate model.
• Employed backward error propagation to optimize input binary vectors within the CNN-IBO framework.
• Benchmarked the CNN-IBO method against FM-QA using nanophotonic designs (planar multilayer, stratified grating structures).

Main Results:

• The CNN-IBO framework successfully optimized nanophotonic designs.
• CNN-IBO achieved optimal designs using significantly fewer actively accumulated training data compared to FM-QA.
• Demonstrated the superior efficiency and effectiveness of CNN-based surrogate functions in active learning for binary optimization.

Conclusions:

• The proposed CNN-IBO method offers a powerful and efficient approach for binary optimization in complex design spaces.
• CNN-IBO shows significant potential for accelerating the discovery of optimal nanophotonic and material designs.
• This framework advances active learning strategies by leveraging the representational power of deep learning models.