Scientists create a deep learning framework for material design, achieving mater

Recently, Professor Hu Run's team from Huazhong University of Science and Technology has successfully built a universal design framework for emissivity engineering based on deep reinforcement learning.

The framework addresses two key issues:

First, it solves the problem of low design efficiency in large optimization space involving material selection. The framework can simultaneously optimize material selection and structural parameters within a large optimization space (~10^10), with a considerably high design efficiency (requiring less than 15% of the total number of structures to be calculated).

Second, it solves the universality issue under different target spectral emissivity. Based on a self-built material candidate library, the framework can autonomously select materials and design structures according to different target spectral emissivity, without the need for the designer's prior knowledge.

In summary, this work has significant value in the interdisciplinary fields of photonics and machine learning, providing a useful technology for the field of nanophotonics and promoting further development of emissivity engineering.The general deep learning framework proposed in this paper for emissivity engineering can achieve efficient optimization design of wavelength-selective emitters for different applications, such as infrared camouflage, passive radiative cooling, gas detection, and thermophotovoltaic technology, among others.

Advertisement

Due to the framework's powerful material selection and structural optimization capabilities, it can provide a variety of structural parameters for a specific application goal, allowing users to choose the most suitable parameters based on actual needs (material type, cost, reliability, etc.) to prepare wavelength-selective emitters.

In addition, passive radiative cooling technology has attracted a lot of attention in recent years, with multiple papers published in Science and Nature. It is believed that this work can provide some inspiration for the design of passive radiative cooling structures, such as the Xiaomi SU7 car window glass that has been popular in the technology circle recently. By regulating the spectral characteristics of ultraviolet, visible light, near-infrared, and mid-infrared, it can achieve passive cooling of vehicles.

So, what is thermal radiation? What is the significance of utilizing thermal radiation?

In nature, all objects with a temperature greater than absolute zero will spontaneously emit thermal radiation. How to utilize this radiation energy is of great importance for energy management and application.However, the surface thermal radiation of most objects generally exhibits a broadband characteristic. Uncontrollable thermal radiation at unwanted wavelengths can lead to energy waste and inefficient utilization.

Therefore, it is necessary to modulate the thermal radiation of objects to achieve selective thermal radiation at specific wavelengths, thereby realizing efficient energy utilization.

In addition, modulating the thermal radiation wavelength of objects can also give rise to new applications, such as infrared camouflage and passive radiative cooling technologies.

From the knowledge of heat transfer, modulating the thermal radiation wavelength of an object can be considered from two aspects:

Firstly, the temperature of the object can be modulated so that its radiation peak falls within the desired wavelength range. However, the precision of such modulation is limited, and the modulation of temperature is also affected by factors such as materials and the environment in which the device is used.The second approach is to modulate the surface emissivity of an object. This method not only has better feasibility but also offers finer modulation precision, enabling effects that temperature modulation cannot achieve.

Therefore, emissivity engineering, which modulates the surface emissivity of an object, is the mainstream method for achieving selective thermal radiation wavelengths.

Emissivity engineering relies on the modulation of electromagnetic waves by subwavelength structures. By designing structures or modifying surfaces, the spectral emissivity of an object can be altered.

The object after modulation is called a wavelength-selective thermal emitter, which can be realized in various structural forms, including one-dimensional multilayer structures, photonic crystals, nano-gratings, and optical cavities.

Among them, the one-dimensional multilayer structure has a high degree of control flexibility due to the diversity of material types and the wide range of tunability of layer numbers and thicknesses. Moreover, compared to complex high-dimensional structures, the preparation process of multilayer thin films is simpler and has the potential for mass production.Therefore, the one-dimensional multilayer structure has become the main carrier for realizing emissivity engineering. And how to design the multilayer structure has become the key issue that needs to be resolved.

However, most of the research is based on the manual design optimization based on the optical resonance conditions. This design method relies on optical knowledge and previous design experience, which not only has a certain design threshold, but also is difficult to achieve the efficiency and optimality of the design.

Optimization design based on machine learning algorithms, such as Bayesian optimization, genetic algorithms, stochastic gradient descent algorithms, etc., can significantly improve design efficiency.

But when the design space is very large, especially when the types of materials, which are discrete and non-derivable parameters, are added to the design space, it will make the above methods unfeasible or the required computational resources significantly increase, and the efficiency is greatly reduced.

Therefore, existing work either fixes the type of material to optimize the structure, or fixes the structure to optimize the material arrangement to reduce the design space.Additionally, when facing different applications (i.e., different target spectral emissivities), existing research has not yet formed a universal design framework, resulting in researchers still needing to review a large amount of literature before optimizing the design to determine the materials and initial structural parameters of the wavelength-selective emitters.

Previously, Hu Run's research team has conducted many studies on metamaterial design and emissivity engineering, establishing various mature design frameworks based on machine learning algorithms, such as Monte Carlo Tree Search, Bayesian Optimization, and Genetic Algorithms.

However, these methods all have more or less some shortcomings. For example, the Monte Carlo Tree Search algorithm is good at optimizing the arrangement of materials, but it is quite limited in optimizing structural parameters.

Bayesian Optimization and Genetic Algorithms have a wide range of applications, but when the design space is large, especially when it involves the selection of materials in different application backgrounds, both algorithms require a significant amount of computational resources, and the design efficiency will be greatly reduced, taking several days or even months to obtain the optimal structure.

In addition, before the optimization is executed, it is still necessary to manually determine the material selection and initial structural parameters, and these manual tasks will affect the optimization results.Thanks to the rapid development of deep learning, they sought to draw some inspiration from it, hence they set their research topic on the efficient design of wavelength-selective thermal radiators under a large optimization space (in a cross-application context) using deep learning.

Later, the research team surveyed the related work on the inverse design of photonic devices using deep learning and found that there are many types of neural networks, and most of the work's target tasks are relatively simple, such as perfect solar absorbers.

In addition, the team found that both supervised neural networks and unsupervised neural networks require the collection of datasets in advance, which is a rather time-consuming but very important task. The size and quality of the dataset can have a serious impact on the neural network model being built.

Later, they also attempted to construct datasets for different target spectral emissivities (involving different bands) and tried several neural networks, finding that when targeting a specific spectral emissivity (the same band), certain effects could be achieved.

However, when combining datasets with different target emissivities to achieve cross-application design, the results are still not satisfactory.During a chance research survey, they focused on deep reinforcement learning and observed its outstanding performance in games. It is suitable for problems with large optimization spaces, and due to the mechanism of reinforcement learning, there is no need to collect datasets in advance. Therefore, the research team attempted to achieve the goal using deep reinforcement learning.

Then, they began to build a deep reinforcement learning design framework, including code writing, testing, and optimizing wavelength-selective thermal radiators under different application backgrounds.

When designing the deep learning framework, how to define the initialization method of materials and structures for wavelength-selective emitters was a challenge they encountered.

Different initialization methods can significantly affect the results and efficiency of the optimization design. To this end, they tried many different approaches, such as random initialization.

That is, at the beginning of each iteration, materials and structure parameters are randomly selected. The advantage of this initialization method is that it can explore a wider design space, but the design efficiency is very low, and sometimes it may not even converge.Additionally, they also attempted a fixed initialization method, where the initial structure is fixed for each iteration. Although this significantly improved design efficiency, the exploration range was limited, and the optimization design results were greatly influenced by the initial structure.

They also tried several other methods, but to balance design efficiency and the breadth of exploration, the team ultimately decided to use the "random initialization plus optimal selection" approach.

Furthermore, finding relatively optimal parameters requires continuous attempts and summarizing patterns. The focus during this period is on how to ensure the consistency between the prepared thickness and the designed thickness, which is also the key to verifying the effectiveness of the wavelength selective radiator.

The growth patterns of different materials during the coating process and the required coating time need to be continuously explored in order to master them. For this reason, they spent several months before obtaining relatively satisfactory samples.

Although these preparation works do not occupy many figures and text in the paper, they also fill the gaps in their previous work experience.Ultimately, the relevant paper was published under the title "General deep learning framework for emissivity engineering" in Light-Science & Applications (IF 19.4).

The first author is Yu Shilu, a master's student at Huazhong University of Science and Technology. Professor Hu Run from Huazhong University of Science and Technology, Professor Junichiro Shiomi from the University of Tokyo in Japan, and Professor Li Wangnan from Hubei University of Arts and Science serve as co-corresponding authors.

Next, they will use the deep learning design framework to conduct some application research involving multispectral control, including multispectral camouflage technology, color or transparent radiative cooling technology, etc.

Because these applications have higher requirements for spectral control and involve multiple bands such as ultraviolet, visible light, near-infrared, mid-infrared, and even microwaves, how to efficiently achieve multispectral control or the decoupling design of the spectrum is what they need to consider.

Of course, they will also explore the direction of spectral control and dynamic control, striving to achieve some refreshing new functions and new applications.

POST A COMMENT