Mark Messner, chief mechanical engineer at Argonne National Laboratory in the United States, is one of the professionals who can predict how materials behave at high temperatures and pressures. Current forecasting methods work well, but are time consuming and often require supercomputers, especially if you already have a set of specific material properties — such as stiffness, density, or strength — and want to find out what type of structure a material would need to meet these properties.
“Usually, a lot of physics-based simulations need to be run to solve such a problem,” Messner said. Therefore, he sought a shorter path and realized that neural networks — a type of artificial intelligence (AI) that discovers patterns in vast datasets — could predict exactly what would happen to a substance under extreme conditions. And they can do this much faster and easier than traditional simulations.
Messner’s new method found the properties of a material more than 2,000 times faster than the traditional approach – read the Journal of Mechanical Design 2019 October article. Messner recognized that much of the computation could be done on a standard GPU laptop instead of supercomputers, which are often inaccessible to most businesses.
This was the first time someone had called a convolutional neural network. used to accurately identify the structural properties of a material. (The simpler type of convolution differs from any other neural network, but is ideal for recognizing patterns in photographs, for example.) This is one of the researchers’ first steps in accelerating the design and characterization of materials in the transition to a completely clean energy economy.
The role of cats on the Internet
Messner began his scientific career at the Lawrence Livermore National Laboratory, where a team was working with a 3D printer to make a micron, or a millionth of a meter structures. Although the research was at the forefront, it was slow. Therefore, the question arose as to whether artificial intelligence leads to results faster?
At that time, the technology giants of Silicon Valley had already begun to use convolutional neural networks to recognize faces and animals in images. This inspired Messner.
“My idea was that the structure of a material is no different from a 3D image. By implication, the 3D version of this neural network will do a good job of recognizing the properties of the structure. Just as a neural network learns that an image is a cat or something else, “the researcher said.
To test his theory, Messner took four steps:
- Designed and defined a square of bricks (quasi-pixels);
- Random samples were taken from this design, and created 2 million data points with a physics-based simulation. These points linked the design to the properties of the desired density and stiffness;
- It fed 2 million data points into the convolutional neural network. This taught the network to look for the right results;
- The genetic algorithm, another type of artificial intelligence designed to optimize results, along with the learned convolutional neural network used to find the overall structure corresponding to the desired properties.
The result was immediate. The new AI method found the appropriate structure 2760 times faster than the traditional physics-based model (0.00075 seconds versus 0.207 seconds).