Home / SmartTech / DeepMind's AI models the transition from glass from a liquid to a solid

DeepMind's AI models the transition from glass from a liquid to a solid

In an article published in the journal Nature Physics DeepMind researchers describe an AI system that can predict the movement of glass molecules during the transition between liquid and solid states. The techniques and trained models that were made available in open source could be used to predict other interesting properties of glass, says DeepMind.

In addition to glass, the researchers claim that the work provides insights into general substance and biological transitions. and that it could lead to advances in industries like manufacturing and medicine. "Machine learning is well positioned to examine the nature of fundamental problems in a number of areas," a DeepMind spokesman told VentureBeat. "We will apply some of the insights and techniques that have been proven and developed through the modeling of glass dynamics to other key questions in science to reveal new things about the world around us."

Glass Dynamics

Glass becomes made by cooling a mixture of melted high temperature sand and minerals. It acts like a solid once it has cooled above its crystallization point and resists tension by pulling or stretching. However, the molecules are structurally similar to those of an amorphous liquid at the microscopic level.

The solution to the physical secrets of glass motivated an annual conference of the Simons Foundation, in which a group of 92 researchers from the USA, Europe, Japan and Brazil and India in New York participated last year. In the three years since the first meeting, they have made breakthroughs like undercooled liquid simulation algorithms, but they still need to develop a full description of the glass transition and the predictive theory of glass dynamics.

 DeepMind glass modeling AI

  VB TRansform 2020: The AI ​​event for managers. San Francisco July 15-16

This is because there are countless unknowns about the nature of the glass formation process, e.g. B. whether it corresponds to a structural phase transition (similar to the freezing of water) and why the viscosity increases by a factor of a trillion during cooling. It goes without saying that modeling the glass transition is a worthwhile task – the underlying physics is based on behavior modeling, drug delivery methods, materials science and food processing. However, the complexity makes it difficult to crack a nut.

AI and machine learning

Fortunately, there are structural markers that help identify and classify phase transitions of matter, and glasses are relatively easy to simulate and enter into particles. based models. Glasses can be modeled randomly as particles that interact via a repulsive potential with a short range. This potential is relational (because only particle pairs interact) and local (because only particles nearby interact).

The DeepMind The team used this to train a neural diagram network – a kind of AI model that works directly on a diagram, a nonlinear data structure made up of nodes (vertices) and edges (lines or arcs, any two Connect knot) exists – to predict glass dynamics. They first created an input diagram in which the nodes and edges represented particles or interactions between particles, so that a particle was connected to its neighboring particles within a certain radius. Two encoder models then embedded the labels (i.e., they were translated into mathematical objects that the AI ​​system could understand). Next, the edge embeds were updated iteratively, initially based on their previous embeds and the embeddings of the two nodes to which they were connected.

 DeepMind Glass Modeling AI

After all edges of the diagram were updated in parallel with the same model, another model updated the nodes based on the sum of their neighboring edge embeddings and their previous embeddings. This process was repeated several times so that local information can spread through the graph. A decoder model then extracted mobilities – measurements of how much a particle typically moves – for each particle from the final embeddings of the corresponding node.

Testing the Model

The team validated their model by constructing multiple data sets that corresponded to mobility predictions on different time horizons for different temperatures. After applying graph networks to the simulated 3D glasses, they found that the system “outperformed” existing physically inspired baselines as well as cutting-edge AI models.

In short, the network was "extremely good" and remained "well-coordinated" until the relaxation time of the glass (which would take up to thousands of years for actual glass), with a correlation of 96% with the basic truth for short periods and for the relaxation time of 64% a correlation of 64% was achieved. In the latter case, this is a 40% improvement over the prior art.

 DeepMind Glass Modeling AI

In a separate experiment, the team examined which factors were important for the success of the diagram model . They measured the sensitivity of the prediction to the central particle when another particle was modified, and were able to judge how large an area the network used to extract its prediction. This provided an estimate of the distance over which particles in the system affected each other.

They report "convincing evidence" that as the glass transition approaches, there are growing spatial correlations and the network has learned to extract them. "These results are consistent with a physical picture in which the correlation length increases as the glass transition approaches," DeepMind wrote in a blog post. "The definition and study of correlation lengths is a cornerstone of the study of the phase transition in physics."


DeepMind claims the knowledge gained could be useful to predict the other properties of glass. As already mentioned, the phenomenon of glass transition manifests itself in more than just window glasses (silica). The associated malfunction transition can be found in ice cream (Kolloll suspension), piles of sand (granular materials) and cell migration during embryonic development as well as in social behavior such as traffic jams.

Glasses are archetypal for this type of complex systems, which operate under restrictions in which the position of elements inhibits the movement of others. It is believed that a better understanding of them will affect many areas of research. For example, imagine a new type of stable but dissolvable glass structure that could be used for drug delivery and building renewable polymers.

“Graph networks can not only help us make better predictions for a number of systems,” wrote DeepMind. However, state what physical correlates are important to modeling so that machine learning systems may help researchers derive basic physical theories and ultimately broaden human understanding rather than replace it. "

Source link