Recently, the team of Academician Ding Han and Professor Wu Zhigang from Huazhong University of Science and Technology, based on machine learning and finite element analysis, proposed an inverse design method for metamaterials, with an average accuracy rate of 98.92% on the test set.

By precisely controlling the stress-strain curve of the structure, this method can address the challenges brought about by complex parameters, nonlinear geometric deformation, and nonlinear material constitutive relationships in the design of metamaterials.

It can achieve high-accuracy results that match experimental and simulation results, which is conducive to leading the development of the field of metamaterial design, as well as promoting the development of mechanical intelligent systems.

For soft mechanical metamaterials, they often possess lightweight, high strength, biocompatibility, rapid response, good shock absorption, and high environmental adaptability.

Advertisement

Therefore, they can be used in the fields of seismic-resistant buildings, lightweight structures, biomedical implants, and embodied intelligent robots.In the field of earthquake-resistant construction, flexible mechanical metamaterials can exhibit excellent shock-absorbing properties, thereby enhancing the seismic performance of buildings.

In the field of lightweight structural design, flexible mechanical metamaterials can be used to manufacture lightweight, high-strength aerospace structures, thereby improving the performance and fuel efficiency of aircraft.

In the field of biomedical implants, flexible mechanical metamaterials can be used to create more durable and biocompatible implants, such as for the manufacture of artificial joints and bone repair materials.

In the field of embodied intelligent robotics, flexible mechanical metamaterials can be used to produce robots with higher sensitivity and stronger environmental adaptability, thus better adapting to complex environments and executing delicate tasks more effectively.

Based on the application target, achieve demand-driven reverse design.Currently, many researchers are attempting to integrate their research with AI. However, in the field of flexible mechanical metamaterial structure design, the integration with AI started relatively late. Most scholars in this field also tend to innovate within the existing system.

It is reported that traditional structural design methods are a kind of forward design method, that is, a method from structure to performance. To achieve a structure with specific mechanical properties, continuous trial and error is usually required.

On the other hand, the inverse design method is a method from performance to structure, which can optimize the iterative process of structural design, shorten the design cycle, and thus accelerate product development speed.

However, in inverse design, the mathematical relationship between structural mechanical performance and its structural parameters is more complex. Therefore, the team decided to combine AI with structural design to achieve the inverse design of flexible mechanical metamaterials.

With the efficient and low-cost characteristics of inverse design, combined with the programmable characteristics of flexible mechanical metamaterial behavior, it will help to achieve mechanical intelligence faster and better.Additionally, AI can also help break through the existing challenges in the design of flexible mechanical metamaterials:

Firstly, for traditional structural design methods such as topological optimization, isogeometric optimization, and finite element analysis, they all start from mathematical logic, which not only has high computational costs but also needs to enhance their versatility. AI can reduce the cost of inverse design and improve versatility.

Secondly, in the design of flexible mechanical metamaterials, a large number of structural parameters are usually involved. AI can help people extract new features from complex data, thereby discovering some mathematical laws that are difficult to find by human efforts, and thus improve design efficiency.

Thirdly, in the existing design methods of flexible mechanical metamaterials, people rarely consider factors such as non-ideal structural configurations, large deformations of nonlinearity, and nonlinear material constitutive models.

Solving the complexity introduced by these non-ideal factors through AI methods will help improve the accuracy of the design and the effectiveness of the design methods.For traditional mechanical structure designers, they are often "the older, the more valuable." That is, the more structures they design, the richer their experience becomes.

However, this also reflects the issue that the field of structural design heavily relies on design experience. For AI that is good at analyzing large amounts of data, this is actually a stage where it can shine.

That is, by learning a large amount of data and experience, AI can summarize the laws in highly nonlinear mathematical relationships to help humans achieve rapid design.

Therefore, when traditional methods based on mathematical logic and AI technology based on data-driven are integrated, different design methods can complement each other, break through the difficulties in the design of flexible mechanical metamaterials, and achieve the design of flexible mechanical metamaterials with high efficiency and high accuracy.

However, as mentioned earlier, flexible mechanical metamaterials themselves have problems such as not ideal configuration, large nonlinear deformation, and nonlinear constitutive models.Translate the following passage into English:

By using finite element analysis to simulate non-ideal and non-linear constraints with high precision, combined with machine learning to fit highly non-linear mathematical relationships, the team was able to achieve on-demand inverse design in this study according to the application object.

Start with complex materials

Specifically, the research group combined computer simulation technology with machine learning algorithms, using data generated by simulation based on mathematical logic to train the machine learning model, thereby establishing the idea of inverse design.

The first step in generating accurate simulation data is to accurately depict the material's constitutive model.

In fact, for any material, its behavior during deformation is a combination of elastic behavior, viscous behavior, and plastic behavior.After trying different materials, the team decided to use thermoplastic polyurethane as the base material.

The reason is: it has good elasticity and plasticity. When subjected to external forces applied at different rates, it can also show different deformation viscosities, and the mechanical behavior during deformation is more complex.

Since thermoplastic polyurethane material is so "complex," if precise design can be achieved for its structure, then for other materials with similar complexity or lower complexity, it will definitely bring some reference.

To ensure that the simulation results can accurately reflect the experimental results in reality, they continuously adjust the experimental assumptions and simulation settings.On this basis, they utilized Python scripts to carry out large-scale computational processing and data processing, generating and analyzing a vast amount of simulation data.

After completing the collection and processing of data, the research team began to construct the inverse network and the forward network to achieve the inverse design of flexible mechanical metamaterials.

For the inverse network, it can learn the mapping relationship from target performance to material structure, thereby being able to directly generate the corresponding material structure based on the given target performance parameters.

For the forward network, it can learn the mapping relationship from material structure to material performance.

At this point, for the design scheme generated by the inverse network, the forward network can verify whether this scheme can meet the expected performance and carry out validation and optimization for the design results, thereby ensuring the feasibility and reliability of the design scheme.In this way, the combination of inverse networks and forward networks can quickly generate new material design schemes to meet specific performance requirements.

Recently, the relevant paper titled "Tailoring Stress-strain Curves of Flexible Snapping Mechanical Metamaterial for On-demand Mechanical Responses via Data-driven Inverse Design" was published in Advanced Materials (IF 27.4).

Huazhong University of Science and Technology doctoral students Zhi-Ping Chai and Zong-Zhi Sheng are the first authors, and Zhi-Gang Wu serves as the corresponding author [1].

Let real material engineers collaborate with AI

In general, this project is the first inverse design of flexible structures achieved by the team based on simulation data and machine learning methods.However, it still has some imperfections: for example, the structural configuration and deformation behavior are still relatively simple.

Therefore, they will continue to conduct research on the inverse design of deformable flexible structures.

Taking human joints as an example, although we cannot precisely control the movement of each joint, when active joint movements are combined with passive joint movements, amazing effects can be achieved.

This is precisely the reason why humans can run and jump flexibly and operate tools. In fact, this is also the charm of mechanical intelligence.

Therefore, the team hopes to design flexible structures with various motion behaviors and combine them to form a complete motion mechanism, which can operate flexibly like human joints.In order to expand the design space and achieve a variety of flexible structural designs for action behaviors, they will further explore compressive behavior, shear behavior, bending behavior, and torsional behavior.

It is hoped to perfect the reverse design system of flexible structures, providing a new perspective for the traditional mechanical structure design field, and ultimately achieving the design of flexible structures with mechanical intelligence.

In addition, the generative AI model used in this work is a relatively basic model, which does not have high requirements for computing power.

For complex generative AI models such as encoder-decoders and diffusion models, they have been tested to some extent.

Although their computing power requirements are higher, the generated design schemes are also richer and more reliable. Therefore, the team is also considering using more advanced generative AI models.Ultimately, the research team hopes to combine the data analysis capabilities and logical reasoning abilities of AI to deeply understand design requirements, allowing real human materials engineers to collaborate side by side with AI, creating mechanical structures with outstanding performance.

In addition to studying the design of flexible structures, the research team is also exploring directions such as AI intelligent plant systems, AI-based tactile perception systems, AI-based adaptive grasping methods, and AI-based flexible robot motion planning [2].

Overall, from the structure of flexible robots, to sensing, and then to control methods, the team hopes to fully integrate AI technology to develop a series of complementary embodied intelligence technologies.