Starr Lab Develops New Model to Study Novel Phase Change Material

In Professor of Physics Francis Starr’s Lab, researchers focus on studying the complexities of soft matter and materials. One target of their investigations are phase change materials, or substances that can transition from one state to another, similar to common transitions between a solid and liquid, but in this case the material can very rapidly switch between two different solid phases. In a recently published paper, Starr and his student co-authors developed a novel model to simulate a phase change that could one day have an impact on such practical matters as how quickly our smartphones process data.

For their study, Starr’s team used a material known as GST. This particular alloy, which is made of three elements (germanium, antimony and tellurium), can transition between solid states at a nanosecond timescale, said Starr. In the different states, at the molecular level, the atoms of the material are either neatly arranged as a crystal, or disordered and amorphous. Heat triggers the change between phases, from amorphous to crystalline and back again. Each phase can be regarded as a zero or a one, and hence represent information in the way computers store binary data.

“Because of the applications to high-performance memory in devices like your smartphone, there’s a lot of interest to understand at a fundamental level how to manipulate changing between these two different states,” said Starr. “To do that, we needed to create a computational model of the interactions between these elements that would accurately and efficiently capture the process of shifting back and forth between these states.”

First author Owen Dunton ’25 took on this project of creating an efficient GST model as a freshman in 2023. Since then, his knowledge of phase change physics and GST has grown significantly. “I’ve definitely grown a very strong appreciation for [phase change materials] because they have a lot of great applications for things like computer memory,” Dunton said. He also spoke about his appreciation for efficient computational modeling, which was the chief accomplishments of the project.

The research team ran into a roadblock because the usual method for modeling the interactions GST—which involves complex quantum mechanics—would have been accurate but too slow to model the phase transition. Quantum-based simulations with just hundreds of atoms could take months to years to complete, and the Starr Lab hoped to study systems containing millions of atoms.

Starr’s student co-authors, which included recent alum Thomas Arbaugh ’23, MA ’24, had the idea to turn to machine learning, a type of artificial intelligence, to solve the problem. The team trained a machine learning model, known as ACE, to mimic the quantum mechanical calculations, but at a fraction of the computational cost.

Their model worked. “We were able to show that, number one, the model accurately reproduces what you get from the full quantum mechanical calculation,” said Starr. “And it’s not only much faster than the quantum mechanical calculation; it’s more than 30 times faster than the fastest neural network model that anybody’s produced before for this material.”

Starr points out that his group are not the first to use machine learning for this purpose but that their model raises the bar. “We can do simulations at the scale of lab experiments that people were never able to do before,” he said. “That’s going to open up a whole world of questions that we can start to address than people have previously been able to consider.”

With their model, Starr and other researchers can perform GST simulations to understand the phase change with atomic level detail. This phase change has a variety of applications. Take the example of the reversible change between the phases, which can correspond to binary zeros and ones for the storage of data on a computer or smartphone. Storing more data and accessing it more quickly requires new materials. “If we want to have bigger and faster memory in our devices you have to consider using new types of materials. That’s exactly what GST is. It’s a material that starts to allow you to exceed the limitations of the traditional silicon-based materials that we have been using,” said Starr.

Dunton agreed. “The main result of this project is that we can now use [the model] to do simulations of system sizes that would actually be used on chips for computer memory,” he said. “We’ve made this potential, this model, publicly available when we published the paper so any researchers who want to study GST have access to this very fast model and can use it for large-scale simulations of GST.”

Increasing device memory is one exciting potential application of his research. But Starr’s work lies in revealing the fundamental physics of phase change that others can learn from. “We wanted to have a much more fundamental scientific understanding and explanation of how this transition happens,” he said. “Because if you have that understanding then you can start to actually exert some control over that.”

Starr credits his student co-authors with pushing the envelope on the work. “Using machine learning to create this new model is not something that my group has worked on in the past. So I really have to give credit to the students for taking up this new approach and working with them to make all of that happen,” Starr said.

Their study was published in The Journal of Chemical Physics.