Neural Networks For Inverse Design Research Articles

Harnessing artificial neural networks for inverse design and analysis of MEMS-based Fabry–Pérot filters

MEMS-based Fabry–Pérot (MEMS-FP) filters have become increasingly important in optical applications such as frequency selection spectroscopy and spectral analysis. Prior to the manufacturing of MEMS-FP devices, conducting a thorough theoretical analysis of their properties is crucial. Currently, Finite Element Analysis (FEA) is the primary method for investigating the theoretical performance of MEMS-FP during the design process. However, the complexity of modeling and prolonged simulation times associated with FEA software pose significant challenges, impeding the development timeline for MEMS-FP devices. To address these challenges, we proposed a machine learning-based method to computer-aided design of the MEMS-FP devices. Our approach involves using an Artificial Neural Network (ANN) to predict their physical properties and an inverse design method, utilizing optimization algorithms such as PSO, NSGA-II, and SA, to derive the optimal design scheme for MEMS-FP devices. Our method accelerates the prediction of physical properties of MEMS-FP devices and enables the design automation of optimal schemes in just minutes within a limited design space. This approach not only reduces the research time and complexity associated with MEMS-FP devices but also shortens their development cycle. Furthermore, one specific MEMS-FP device is fabricated and tested for proof-of-concept of our proposed method.

Artificial neural networks for inverse design of a semi-auxetic metamaterial

This study introduces an artificial neural network approach for the inverse design of a novel semi-auxetic mechanical metamaterial to achieve a specified stress-strain curve and/or Poisson's ratio-strain curve. To accomplish this, after presenting the metamaterial and assessing its characteristics, 1500 structures of the same metamaterial with various parameters are generated using a parametric model. The metamaterials are then gone through a compression test simulation using Finite Element (FE) analysis; accordingly, each metamaterial's stress-strain and Poisson's ratio curves are derived. The results of FE simulations are validated using mesh convergence check and experimental compression tests on a 3D printed specimen of the proposed metamaterial. In the next step, 80 % of the data are randomly selected to be used as training data for the artificial neural networks (ANN), while the remaining 20 % is employed to evaluate the performance of the ANNs using different metrics. The capability of the ANNs to predict the design parameters of the proposed metamaterial is assessed by providing different kinds of inputs, including the stress-strain curve, Poisson's ratio curve, and both. The observations reveal that the ANNs achieve more accurate results when both the stress-strain and Poisson's ratio-strain curves are provided as the inputs. The presented ANN in this study serves as a robust tool for precisely designing the parameters of the proposed metamaterial, allowing for the attainment of the desired stress-strain and/or Poisson's ratio-strain behavior. It is shown that the proposed metamaterial owns potential applications in crawling soft robotics, automotive, and construction industries.

Physics-model-based neural networks for inverse design of binary phase planar diffractive lenses.

The inverse design approach has enabled the customized design of photonic devices with engineered functionalities through adopting various optimization algorithms. However, conventional optimization algorithms for inverse design encounter difficulties in multi-constrained problems due to the substantial time consumed in the random searching process. Here, we report an efficient inverse design method, based on physics-model-based neural networks (PMNNs) and Rayleigh-Sommerfeld diffraction theory, for engineering the focusing behavior of binary phase planar diffractive lenses (BPPDLs). We adopt the proposed PMNN to design BPPDLs with designable functionalities, including realizing a single focal spot, multiple foci, and an optical needle with size approaching the diffraction limit. We show that the time for designing single device is dramatically reduced to several minutes. This study provides an efficient inverse method for designing photonic devices with customized functionalities, overcoming the challenges based on traditional data-driven deep learning.

Deep Neural Networks for Inverse Design of Nanophotonic Devices

Deep learning is now playing a major role in designing photonic devices, including nanostructured photonics. In this article, we investigate three models for designing nanophonic power splitters with multiple splitting ratios. The first model is a forward regression model, wherein the trained deep neural network (DNN) is used within the optimization loop. The second is an inverse regression model, in which the trained DNN constructs a structure with the desired target performance given as input. The third model is a generative network, which can randomly produce a series of optimized designs for a target performance. Focusing on the nanophotonic power splitters, we show how the devices can be designed by these three types of DNN models.

Artificial neural networks for inverse design of resonant nanophotonic components with oscillatory loss landscapes

AbstractMachine learning offers the potential to revolutionize the inverse design of complex nanophotonic components. Here, we propose a novel variant of this formalism specifically suited for the design of resonant nanophotonic components. Typically, the first step of an inverse design process based on machine learning is training a neural network to approximate the non-linear mapping from a set of input parameters to a given optical system’s features. The second step starts from the desired features, e.g. a transmission spectrum, and propagates back through the trained network to find the optimal input parameters. For resonant systems, this second step corresponds to a gradient descent in a highly oscillatory loss landscape. As a result, the algorithm often converges into a local minimum. We significantly improve this method’s efficiency by adding the Fourier transform of the desired spectrum to the optimization procedure. We demonstrate our method by retrieving the optimal design parameters for desired transmission and reflection spectra of Fabry–Pérot resonators and Bragg reflectors, two canonical optical components whose functionality is based on wave interference. Our results can be extended to the optimization of more complex nanophotonic components interacting with structured incident fields.

Neural networks for inverse design of phononic crystals

Intelligent design of one-dimensional (1D) phononic crystals (PCs) by neural networks (NNs) is proposed. Two neural network models, supervised neural network (S-NN) and unsupervised neural network (U-NN), are used to realize the inverse design of PCs, concerning both geometric and physical parameter designs. Performances of the two models are compared and discussed. The results show that the bandgaps of the designed PCs by the two NNs are highly consistent with the target bandgaps. For the design of single or two parameters, the performances of the two NNs are excellent; while for the case of three-parameter design, U-NN works much better than S-NN due to the impact of non-uniqueness on S-NN. The present work confirms the feasibility of inverse design of PCs by NNs, and provides a useful reference for the application of NNs in the intelligent inverse design of 2D or 3D PCs.