Then there is glass, a strange intermediate substance that has puzzled physicists for decades. Take a snapshot of the molecules in the glass. They look like fluid disorders. But most molecules hardly move, making this material as hard as a solid.
Glass is formed by cooling some liquid. However, why do the molecules in the liquid drop sharply at a certain temperature, but the structural arrangement has not changed significantly (this phenomenon is called glass transition)? So far, people haven't figured out what caused the glass to be so hard.
Now, researchers at DeepMind, an artificial intelligence company owned by Google, have used AI to study the changes of molecules in glass with hardening. DeepMind's artificial neural network can only use the "snapshot" of its physical structure at a time to predict how molecules will move in a very long time scale. Victor Bapst of DeepMind believes that even if the microstructure of glass seems to have no features, "this structure may be more predictive of dynamics than people think."
Peter Hallowell, who studies glass transition at the University of Sydney, agrees. He said that this paper is "more telling than previous papers on glass hardness" and "the structure is dynamically coded in some way", so glass is not as chaotic as liquid after all.
In order to understand the microscopic changes leading to glass transition, physicists need to relate two kinds of data: how the molecules in glass are arranged in space and how they move (slowly) with time. One way to relate these substances is to relate them to a quantity called dynamic trend: given the current position of a group of molecules, how many molecules they may move at a specific time in the future. This developing quantity comes from using Newton's law to calculate molecular trajectories, starting with many different random initial velocities, and then averaging the results together.
By simulating these molecular dynamics, computers can generate "trend maps" of thousands of glass molecules, but only on the time scale of one trillionth of a second. By definition, molecules in glass move very slowly. Giulio Biroli, a condensed matter physicist at Paris Normal University, said, "It is impossible for ordinary computers to calculate their trends to a few seconds or more, because it takes too much time".
More importantly, Biroli said, understanding what structural features (if any) may lead to molecular tendency in glass only through these simulations does not give physicists much insight.
DeepMind researchers set out to train an AI system to predict the characteristics of glass without actually running the simulation, and tried to understand the sources of these characteristics. They use a special artificial neural network, which takes a graph (a group of nodes connected by lines) as input. Each node in the graph represents the three-dimensional position of a molecule in glass. Lines between nodes indicate the distance between molecules. Bapst said that because the neural network "learns" by changing its own structure to reflect its input structure, "graphical neural network is very suitable for representing the interaction of particles".
Bapst and his colleagues first used the simulation results to train their AI system: they created a virtual glass cube containing 4096 molecules, simulated the evolution of molecules based on 400 unique starting positions at different temperatures, and calculated the inertia of particles. After training the neural network to accurately predict these trends, the researchers then sent 400 previously invisible particle configurations ("snapshots of glass molecular configurations") to the trained network.
Using only these structural snapshots, neural networks can predict molecular characteristics at different temperatures with unprecedented accuracy. Compared with the latest machine learning prediction method, the prediction distance for the future will reach 463 times.
Biroli said that DeepMind neural network can predict the future motion of molecules only by its snapshot of the current structure, which provides a powerful new method for exploring the dynamics of glass and other materials.
However, what pattern does the network detect in these snapshots to make predictions? The system can't easily reverse engineer to determine the preventive measures it learns in training-this is a common problem for researchers who try to use artificial intelligence for scientific research. But in this case, they found some clues.
According to Agnieszka Grabska-Barwinska, a team member, the graphic neural network learned what coding physicists call the correlation length pattern. That is to say, as DeepMind's graphical neural network reorganizes itself to reflect the training data, it shows the following trend: when the prediction trend is at a higher temperature (the molecular movement looks more like a liquid than a solid, but not a solid), for the prediction of each node, the network relies on information from neighboring nodes (there are two or three connections in the figure). But at a lower temperature close to the glass transition, the number (correlation length) increases to 5.
Thomas Keck, a physicist in DeepMind's team, said: "As the temperature drops, we find that the network extracts information from larger and larger blocks." "At these different temperatures, the glass looks exactly the same with the naked eye. But with the application of our AI technology, Tu neural network saw something different. "
The increase of correlation length is a sign of phase transition, in which particles change from disorder to ordered arrangement, and vice versa. For example, this happens when the atoms in a piece of iron line up to magnetize the iron. As the block approaches this transition, each atom affects the atoms farther and farther away in the block.
For a physicist like Biroli, the ability of neural network to understand the correlation length and incorporate it into the prediction shows that some hidden order will inevitably be formed in the glass structure during the glass transition. Peter Wolynes, a glass expert at Rice University, said that the relevant length of machine learning provides evidence that materials will "approach thermodynamic phase transition" when they become glassy.
However, the knowledge obtained by neural network cannot be easily converted into new equations. Pushmeet Kohli, head of DeepMind's scientific team, said, "We can't say,' Oh, actually our network is studying this correlation, and I can provide you with a formula'. For some glass physicists, this warning limits the practicality of graphic neural networks. " Can this be explained in human terms? "Wolins said," they didn't do it. This doesn't mean that they can't do this in the future. "