Probability-Density-Based Deep Learning Paradigm for the Fuzzy Design of Functional Metastructures

Yingtao Luo, Peng-Qi Li, Dongting Li, Yu‐Gui Peng, Zhi-Guo Geng, Shuhuan Xie, Yong Li, Andrea Alù, Jie Zhu, Xue‐Feng Zhu

TL;DR

This paper introduces a novel "Probability-Density-Based Deep Learning" (PDN) paradigm for the inverse design of functional metastructures. Unlike traditional methods that aim for a single solution, PDN leverages probability distributions to capture the inherent "fuzziness" of inverse design problems, enabling the efficient discovery of multiple plausible designs and providing a confidence measure for each.

What Problem Does This Paper Address?

Designing functional metastructures, which are engineered materials with properties not found in nature (like invisibility cloaks or super-resolution lenses), is a complex challenge. Traditional "forward design" methods involve manually creating structures based on physical rules. More recently, "inverse design" aims to find structures that achieve a desired function. However, inverse design is often a "one-to-many" problem: a single desired function can be realized by many different structures. Existing deep learning methods struggle with this multivalued nature, often producing only one design or suffering from unstable training and hyperparameter sensitivity. This paper addresses the need for an inverse design method that can efficiently and accurately explore the vast parameter space of metastructures, identify multiple viable designs, and provide a measure of confidence for each potential solution.

Methods

The core innovation is the Probability-Density-Based Deep Learning (PDN) architecture. It consists of two main modules:

  1. Front-end Neural Network: This network takes a target transmission spectrum (a desired functional property) as input and outputs parameters for a mixture Gaussian distribution. Crucially, it doesn't directly output the metastructure's geometry.
  2. Rear-end Mixture Gaussian Sampling: The output parameters (mixing coefficients, means, and standard deviations) are used to construct a mixture Gaussian. This distribution represents the probability of different metastructure configurations achieving the target function.

The training objective is to maximize the likelihood of the actual metastructure labels in the dataset, effectively learning to place high probability density around known good designs. This probabilistic output allows the identification of multiple "local maxima" within the Gaussian distribution, each corresponding to a plausible and potentially unique metastructure design. The amplitude of the mixture Gaussian at any point serves as a confidence score for that particular design. This approach is inspired by quantum mechanics' use of probability density to describe particle states.

Datasets

The authors generated a dataset of acoustic metastructures using COMSOL Multiphysics and MATLAB. The metastructures consist of 5 cylindrical layers, with radii sampled from 8 discrete values for each layer, resulting in 32,768 training data points. The input to the PDN is a transmission spectrum across 250 frequencies (20 Hz to 5000 Hz), and the output is the 5-dimensional radii of the cylindrical layers. 1,000 test data points were also generated.

Key Insights and Ideas

  • Embracing "Fuzziness": The paper argues that inverse design problems are inherently "fuzzy" or multivalued. Instead of forcing a one-to-one mapping, PDN embraces this by modeling the probability distribution of solutions.
  • Confidence Scoring: The probability density output of the mixture Gaussian provides a natural way to assess the confidence in a predicted design, allowing users to prioritize more likely or robust solutions.
  • Beyond Single Solutions: Unlike methods that yield only one design, PDN can generate multiple distinct metastructures for the same target function, offering greater design flexibility and potentially leading to more innovative or manufacturable solutions.
  • Generalization: The PDN demonstrated good generalization capabilities, predicting structures with radii that fell within interpolated ranges of the training data, suggesting it learns a robust underlying relationship.

Key Results

The PDN was successfully applied to inverse design problems in acoustics, demonstrating its effectiveness through experimental verification.

  • Accurate Transmission Spectra: In two examples, the predicted transmission spectra of the designed metastructures closely matched the target spectra and were experimentally validated.
  • Multiple Plausible Designs: For a given target transmission spectrum, PDN identified multiple distinct metastructures (e.g., A1, A2, A3 in Figure 2; B1, B2, B3, B4 in Figure 3) that all achieved similar desired functionalities.
  • Balanced Performance: Compared to other deep learning models (ANN, GAN, VAE, TNN), PDN showed a superior balance of accuracy, efficiency (time cost), and output variety. While TNN achieved high accuracy, it lacked variety. GAN and VAE offered variety but struggled with accuracy or time costs, especially without extensive hyperparameter tuning. PDN achieved good accuracy with multiple outputs and a reasonable training time.

Limitations and Future Work

  • Fixed Output Dimension: The current PDN architecture is designed for a fixed number of layers (5 in this study). The authors suggest incorporating recurrent neural networks (RNNs) or long short-term memory (LSTM) modules to allow for flexible output dimensions, enabling designs with varying numbers of layers or structural parameters.
  • Adaptive Input: Extending the input to be adaptive, perhaps by embedding acoustic functionalities or allowing variable input sizes for transmission/reflection spectra, could further enhance flexibility.
  • Data-Driven Physics Discovery: The authors envision PDN as a powerful tool for data-driven discovery of underlying physical equations governing inverse design, potentially leading to more fundamental understanding and universally applicable models.

Summary updated

Abstract

In quantum mechanics, a norm-squared wave function can be interpreted as the probability density that describes the likelihood of a particle to be measured in a given position or momentum. This statistical property is at the core of the fuzzy structure of microcosmos. Recently, hybrid neural structures raised intense attention, resulting in various intelligent systems with far-reaching influence. Here, we propose a probability-density-based deep learning paradigm for the fuzzy design of functional metastructures. In contrast to other inverse design methods, our probability-density-based neural network can efficiently evaluate and accurately capture all plausible metastructures in a high-dimensional parameter space. Local maxima in probability density distribution correspond to the most likely candidates to meet the desired performances. We verify this universally adaptive approach in but not limited to acoustics by designing multiple metastructures for each targeted transmission spectrum, with experiments unequivocally demonstrating the effectiveness and generalization of the inverse design.
BibTeX
@inproceedings{luo2020probability,
  author = {Yingtao Luo and Peng-Qi Li and Dongting Li and Yu‐Gui Peng and Zhi-Guo Geng and Shuhuan Xie and Yong Li and Andrea Alù and Jie Zhu and Xue‐Feng Zhu},
  title = {{Probability-Density-Based Deep Learning Paradigm for the Fuzzy Design of Functional Metastructures}},
  booktitle = {Research},
  year = {2020},
  doi = {10.34133/2020/8757403},
}
Figure