To improve the utilization of optical data sets, the Tianda team proposed an AI model to enhance spectral prediction effects

Editor | Dead Leaf Butterfly

Recently, the team of Associate Professor Wu Liang and Academician Yao Jianquan of the Institute of Laser and Optoelectronics of Tianjin University and the team of Professor Xiong Deyi of the Natural Language Processing Laboratory reported A scheme to enhance spectral predictions using deep learning models with multi-frequency supplementary inputs. This scheme can improve the accuracy of spectral prediction by using multi-frequency input data. In addition, this solution can also reduce noise interference in the spectrum prediction process, thereby improving the prediction effect.

This solution can improve the utilization of existing optical data sets and enhance the prediction effect of the spectral response corresponding to the metasurface structure without increasing the training cost.

The relevant research results were titled "Enhanced spectrum prediction using deep learning models with multi-frequency supplementary inputs" and were published in "APL Machine Learning" on May 16, 2024 》.

Paper link: https://doi.org/10.1063/5.0203931

Research background

In recent years, the rapid development of deep learning technology has brought unprecedented changes and innovations to various fields, and has become an effective tool for processing complex and huge data in multiple disciplines.

Methods based on neural networks can effectively detect relevant features and potential patterns of target data, but there are still certain challenges if the deep learning model directly learns these related data from different fields and different formats. To solve this problem, feature extraction techniques can be used. Feature extraction techniques can transform raw data into a representation suitable for a specific task. Different feature extraction methods can be used, such as FFT based on frequency domain analysis, WT based on wavelet transform, etc. By applying these technologies, different fields can be combined.

In recent years, research fields that combine deep learning technology have generally faced problems such as the small size and low quality of existing data sets, which has affected the model's ability to perform target tasks. learning result.

In the entire "AI for Science" research process, the most expensive part is mainly the construction of data sets. Therefore, how to use existing data sets more effectively is crucial.

Research by the Tianjin University team has proven that adding supplementary multi-frequency input information to the existing data set during the target spectrum prediction process can significantly improve the prediction accuracy of the network. This approach provides new ideas for using data sets for interdisciplinary research and applications in deep learning and other fields such as photonics, composite material design, and biomedicine.

Research Highlights

The innovative point of the research is to propose the idea of ​​splitting spectral information in the full frequency range, which is manifested in combining the actual design requirements and dividing the full-frequency spectral information according to The learning tasks are split into the working frequency part and the non-working frequency part.

In order to demonstrate the universality of this solution, the target operating frequency band was refined into a low-frequency information (0-1 THz) part and a high-frequency information (1-2 THz) part to demonstrate the enhancement of model learning. Effect.

Compared with direct prediction of the working frequency range data, after supplementing other frequency information, the overall transmission spectrum data prediction error dropped by about 80%. Among them, the Transformer-based model after supplementing the low-frequency information , the prediction error is only about 40% of the direct prediction. The designed metasurface structure and model architecture are shown in Figure 1:

Figure 1 (a)-(b) Schematic diagram of the metasurface structure, in which the brown part represents "1" pixels and the yellow part represents "0" pixels. (c) Schematic diagram of CNN, LSTM, GRU and Transformer network. In the forward network, the input of the model is a 25*25 pixel metasurface matrix, and the output is the optical response, while the reverse network is the opposite. The "H" and "L" marked in the figure represent the relevant amplitude and phase data of high frequency and low frequency respectively.

In order to more intuitively demonstrate the prediction effect of optimized amplitude and phase parameters at different operating frequencies, some metasurface structures are randomly selected for simulation demonstration in CST Studio Suite software, as shown in Figure 2:

Figure 2 Schematic diagram of the prediction effect of optimized high-frequency and low-frequency data. (a)-(f) Demonstrate the different prediction performance of the optimized network model in different frequency ranges by comparing real data (purple solid line) with predicted data (black dashed line). Green areas represent frequency information data used as supplementary input, while yellow areas represent areas used to validate optimized prediction performance. where a and b represent the prediction results of the high-frequency and low-frequency amplitudes of the x-polarization state. (c)-(d) Prediction results of high-frequency and low-frequency amplitudes of y-polarization state. (e)-(f) Prediction results of high-frequency and low-frequency phases.

Summary and Outlook

This research effectively improves the utilization efficiency of existing data sets by conducting targeted data set splitting for learning tasks of different optical problems. This further improves the learning effect of the deep learning model.

This optimization solution effectively alleviates the problem of small existing optical data sets (especially related data sets in the terahertz band), and also provides more research areas that combine deep learning technology but have expensive data, such as Composites design, medical image analysis, financial data prediction, and more provide a new perspective on optimizing data sets.

First author: Xing Xiaohua, Ren Yuqi Instructors: Wu Liang, Xiong Deyi, Yao Jianquan

Paper collaborators: Zou Die, Zhang Qiankun, Mao Bingxuan

Acknowledgments: Professor Zhang Shuang (University of Hong Kong) and Professor Han Jiaguang for their help during the thesis work. Relevant research is supported by projects such as the National Key R&D Program and the National Natural Science Foundation.

Correspondent: Zhang Qiankun Shi Senfang

The above is the detailed content of To improve the utilization of optical data sets, the Tianda team proposed an AI model to enhance spectral prediction effects. For more information, please follow other related articles on the PHP Chinese website!