The Health Benefits of Cranberries: Essential Insights for Your Thanksgiving Feast

Are cranberries good for you? What to know before Thanksgiving. Are you team canned or team fresh cranberry sauce? This Thanksgiving, we're answering plenty of your burning, commonly-searched food questions. Here, we're tackling the nutritional facts behind cranberries. Here's how certain cranberry dishes may or may not boost your nutrition this holiday season. And remember
HomeTechnologyHarnessing Machine Learning to Accelerate Simulations of Non-Uniform Particle Shapes

Harnessing Machine Learning to Accelerate Simulations of Non-Uniform Particle Shapes

Simulating spherical particles is relatively straightforward. However, in reality, most particles are not perfect spheres; they exhibit irregular shapes and sizes. This makes simulating such particles significantly more complex and time-consuming.

The capacity to simulate particles is vital for understanding their behavior. For instance, microplastics represent a type of pollution that has surged due to the drastic increase in plastic waste, which degrades uncontrollably in the environment through mechanical or UV processes. These minuscule particles are now virtually omnipresent globally. To address this environmental issue effectively, it’s crucial to gain more insight into these particles and their behaviors.

As part of the effort to tackle this problem, researchers from the University of Illinois Urbana-Champaign have utilized neural networks to predict how irregularly shaped particles interact, significantly speeding up molecular dynamics simulations. This innovative approach allows simulations to be conducted up to 23 times quicker than conventional methods and can accommodate any irregular shape given adequate training data.

“Microplastics are ubiquitous in the environment, and they’re mostly not spherical; they are highly varied, featuring corners and edges. To understand their environmental behavior, we need to create new methods that enable us to simulate them more rapidly, affordably, and efficiently,” states Antonia Statt, a professor of materials science and engineering.

Spheres are simple to simulate because the primary factor in determining interactions between two particles is the distance from the center of each sphere. However, shifting from spheres to more complex shapes—like cubes or cylinders—requires knowledge of not only the distance between particles but also the angles and relative orientations. Traditionally, simulating a cube involves constructing it from numerous tiny spheres.

“Describing a cube through multiple small spheres is a convoluted method,” Statt notes. “It’s also resource-intensive since you need to compute the interactions among all those small spheres. To streamline this process, we employed machine learning—specifically, a feed-forward neural network—which essentially means we are finding a complex function that remains unknown. Neural networks excel in this area; with sufficient data, they can approximate almost anything.”

This technique eliminates the need to calculate every individual distance between the small spheres. Instead, only the distance from center to center and their relative orientation are necessary, streamlining the process considerably. Moreover, this method maintains the accuracy of traditional simulations, as it is trained on data derived from those approaches, but it operates with greater efficiency.

Looking ahead, Statt aims to simulate more intricate irregular shapes and combinations of different shapes, such as a cube paired with a cylinder instead of just two cubes. “We need to understand all the unique interactions, but our method is versatile enough to achieve that,” she explains.

This study, titled “Molecular dynamics simulations of anisotropic particles accelerated by neural-net predicted interactions,” was recently published in The Journal of Chemical Physics and has been included in the 2024 JCP Emerging Investigators Special Collection. It is also highlighted on the cover of this issue of JCP.

Antonia Statt is additionally affiliated with the Materials Research Laboratory, the Department of Chemical and Biomolecular Engineering, and the Beckman Institute for Advanced Science and Technology at Illinois.

Other contributors to this research include B. RuÅŸen Argun from the Department of Mechanical Engineering and Yu Fu from the Department of Physics, both at Illinois.

This research received funding from the Molecule Maker Lab Institute (MMLI), an AI Research Institutes initiative supported by the National Science Foundation.