Achieving efficient inverse design of low-dimensional heterostructures based on a vigorous scalable multi-task learning network

Abstract

A scalable multi-task learning (SMTL) model is proposed for the efficient inverse design of low-dimensional heterostructures and the prediction of their optical response. Specifically, several types of nanostructures, including single and periodic graphene-Si heterostructures consisting of n×n graphene squares (n=1∼9), 1D periodic graphene ribbons, 2D arrays of graphene squares, pure Si cubes and their periodic array counterparts, are investigated using both traditional finite element method and SMTL network, with the former providing training data (optical absorption) for the latter. There are two important algorithms implemented in SMTL model: one is the normalization mechanism that makes different parameters of different structures on the same scale, ensuring that SMTL network can deal with tasks with different dataset impartially and without bias; the other one is used to capture the impact of nanostructures’ dimensions on their optical absorption and thus improve the generalization ability of SMTL. Utilizing SMTL model, we first study the absorption property of the multiple shaped nanostructures and look deeper into the impacts of n×n graphene squares and Si cuboid on the optical absorption of their heterostructures. Equally important, the multi-structure inverse design functionality of SMTL is confirmed in this context, which not only owns high accuracy, fast computational speed, and excellent generalizable ability, but also can be applied to contrive new structures with desired optical response. This work adds to the rapidly expanding field of inverse design in nanophotonics and establishes a multi-task learning framework for heterostructures and more complicated nanoparticles.

Automatic Segmentation of Novel Coronavirus Pneumonia Lesions in CT Images Utilizing Deep-Supervised Ensemble Learning Network

“…In general, the designed deep learning models are trained with small datasets, which may lead to poor generalization ability and serious overfitting performance. To cope with the shortcomings, many technical means, such as data augmentation ( 31 ), multi-task learning ( 33 ), transfer learning ( 34 ), and attention mechanism ( 35 ) can be used to achieve the segmentation task. In this section, a transfer learning strategy is applied to solve the shortcomings.…”

Data-Enhanced Deep Greedy Optimization Algorithm for the On-Demand Inverse Design of TMDC-Cavity Heterojunctions

“…Therefore, a normalization method is highly desired to make the above parameters be on the same scale. Thus, the parameter normalization method can be represented as [ 45 ]: where x represents the actual value, x min stands for the minimum value, and δ is the distance of discrete attributes. Notably, the dataset contains only reflectance spectra with non-negative incident angles are included in the dataset due to the symmetry feature, with the incident angle resolution being 1°.…”

“…Here, “optimization” means reducing the discrepancies between the current spectrum and the target spectrum to obtain more precise used parameters by changing the incident angle and structural variable, whose adjustment process is defined as “action”. Moreover, “greedy” indicates that the model selects actions that better reduce the spectral differences, characterized by the usage of mean absolute percentage error (MAPE) [ 45 ]: …”

“…The DL algorithms are employed in various fields, such as nature language processing [ 24 , 25 , 26 ], image recognition [ 27 ], finance [ 28 , 29 , 30 ], and medicine [ 31 , 32 , 33 ]. Especially for the aspect of nanophotonics, the DL approach has proved to be one of the most advanced tools to study many nonintuitive and nonlinear physics issues [ 34 , 35 , 36 , 37 ], including the problem of the inverse design of photonics devices [ 38 , 39 , 40 , 41 , 42 , 43 , 44 , 45 ]. Peurifpy et al [ 46 ] utilized 50,000 samples to train neural networks to design a multi-layer dielectric spherical nanoparticle in 2018, which was regarded as the landmark in this field.…”

Dual Optical Frequency Comb Neuron: Co‐Developing Hardware and Algorithm

“…[41] Additionally, we abstract the DOFCN as a composite vector convolution (CVC) operator and build CVC neural networks (CVCNNs). Subsequently, we select signal modulation format identification (MFI) and optical metamaterial structure inverse design (SID) [42,43] tasks to test the classification and regression capability of CVCNNs, respectively. The ablation experiments reveal that CVCNNs have improved performance than that of the Baseline (32-bit CNNs) reported in refs.…”