In a paper revealed within the journal Nature Physics, DeepMind researchers describe an AI system that may predict the motion of glass molecules as they transition between liquid and strong states. The methods and skilled fashions, which have been made obtainable in open supply, could possibly be used to foretell different qualities of curiosity in glass, DeepMind says.

Beyond glass, the researchers assert the work yields insights into basic substance and organic transitions, and that it may result in advances in industries like manufacturing and drugs. “Machine learning is well placed to investigate the nature of fundamental problems in a range of fields,” a DeepMind spokesperson instructed VentureBeat. “We will apply some of the learnings and techniques proven and developed through modeling glassy dynamics to other central questions in science, with the aim of revealing new things about the world around us.”

Glassy dynamics

Glass is produced by cooling a combination of high-temperature melted sand and minerals. It acts like a strong as soon as cooled previous its crystallization level, resisting pressure from pulling or stretching. But the molecules structurally resemble that of an amorphous liquid on the microscopic stage.

Solving glass’ bodily mysteries motivated an annual convention by the Simons Foundation, which final yr hosted a gaggle of 92 researchers from the U.S., Europe, Japan, Brazil, and India in New York. In the three years because the inaugural assembly, they’ve managed breakthroughs like supercooled liquid simulation algorithms, however they’ve but to develop a whole description of the glass transition and predictive idea of glass dynamics.

DeepMind’s AI models transition of glass from a liquid to a solid

DeepMind’s AI models transition of glass from a liquid to a solid

That’s as a result of there are numerous unknowns concerning the nature of the glass formation course of, like whether or not it corresponds to a structural part transition (akin to water freezing) and why viscosity throughout cooling will increase by an element of a trillion. It’s well-understood that modeling the glass transition is a worthwhile pursuit — the physics behind it underlie habits modeling, drug supply strategies, supplies science, and meals processing. But the complexities concerned make it a tough nut to crack.

AI and machine studying

Fortunately, there exist structural markers that assist establish and classify part transitions of matter, and glasses are comparatively simple to simulate and enter into particle-based fashions. As it occurs, glasses could be modeled as particles interacting by way of a short-range repulsive potential, and this potential is relational (as a result of solely pairs of particles work together) and native (as a result of solely close by particles work together with one another).

The DeepMind workforce leveraged this to coach a graph neural community — a kind of AI mannequin that instantly operates on a graph, a non-linear knowledge construction consisting of nodes (vertices) and edges (traces or arcs that join any two nodes) — to foretell glassy dynamics. They first created an enter graph the place the nodes and edges represented particles and interactions between particles, respectively, such {that a} particle was linked to its neighboring particles inside a sure radius. Two encoder fashions then embedded the labels (i.e., translated them to mathematical objects the AI system may perceive). Next, the sting embeddings have been iteratively up to date, at first primarily based on their earlier embeddings and the embeddings of the 2 nodes to which they have been linked.

DeepMind’s AI models transition of glass from a liquid to a solid

After all the graph’s edges have been up to date in parallel utilizing the identical mannequin, one other mannequin refreshed the nodes primarily based on the sum of their neighboring edge embeddings and their earlier embeddings. This course of repeated a number of occasions to permit native data to propagate by way of the graph, after which a decoder mannequin extracted mobilities — measures of how a lot a particle sometimes strikes — for every particle from the ultimate embeddings of the corresponding node.

Testing the mannequin

The workforce validated their mannequin by establishing a number of knowledge units similar to mobilities predictions on totally different time horizons for various temperatures. After making use of graph networks to the simulated 3D glasses, they discovered that the system “strongly” outperformed each current physics-inspired baselines and state-of-the-art AI fashions.

They say that community was “extremely good” on brief occasions and remained “well matched” as much as the comfort time of the glass (which might be as much as 1000’s of years for precise glass), reaching a 96% correlation with the bottom reality for brief occasions and a 64% correlation for leisure time of the glass. In the latter case, that’s an enchancment of 40% in contrast with the earlier cutting-edge.

DeepMind’s AI models transition of glass from a liquid to a solid

In a separate experiment, to higher perceive the graph mannequin, the workforce explored which elements have been vital to its success. They measured the sensitivity of the prediction for the central particle when one other particle was modified, enabling them to evaluate how massive of an space the community used to extract its prediction. This offered an estimate of the space over which particles influenced one another within the system.

They report there’s “compelling evidence” that rising spatial correlations are current upon approaching the glass transition, and that the community discovered to extract them. “These findings are consistent with a physical picture where a correlation length grows upon approaching the glass transition,” wrote DeepMind in a weblog submit. “The definition and study of correlation lengths is a cornerstone of the study of phase transition in physics.”


DeepMind claims the insights gleaned could possibly be helpful in predicting the opposite qualities of glass; as alluded to earlier, the glass transition phenomenon manifests in additional than window (silica) glasses.  The associated jamming transition could be present in ice cream (acolloidal suspension), piles of sand (granular supplies), and cell migration throughout embryonic improvement, in addition to social behaviors equivalent to site visitors jams.

Glasses are archetypal of those sorts of advanced techniques, which function beneath constraints the place the place of components inhibits the movement of others. It’s believed that a greater understanding of them could have implications throughout many analysis areas. For occasion, think about a brand new kind of secure but dissolvable glass construction that could possibly be used for drug supply and constructing renewable polymers.

“Graph networks may not only help us make better predictions for a range of systems,” wrote DeepMind, “but indicate what physical correlates are important for modeling them that machine learning systems might be able to eventually assist researchers in deriving fundamental physical theories, ultimately helping to augment, rather than replace, human understanding.”