The current study explores the possibility of using activations from AI models as a new class of order parameter. The intuition is simple: AI models are essentially order detectors. A vision model detects spatial order in images; a language model detects order among words and tokens; a scientific AI model can detect order in physical configurations. If such a model is trained to recognize different crystal structures, its output can be interpreted not only as a class label, but also as a continuous measure of how much structural order remains.
The Solution: The researchers converted atomistic configurations into three-dimensional density images and trained a convolutional neural network to classify body-centered cubic (BCC), face-centered cubic (FCC), simple cubic (SC), and random structures. They then applied controlled Lindemann-type vibrational disorder to the ideal crystals and monitored how the learned structural scores changed.
The model revealed a two-stage loss of order in BCC crystals. BCC lattice identity fell below 50% at a Lindemann ratio near 0.13, while the generic solid-like probability remained above 50% until about 0.23. In this intermediate regime, the distorted BCC structures were mainly classified as FCC-like rather than random. In contrast, FCC and SC crystals showed a more direct crossover, with lattice-specific and solid-like order disappearing at nearly the same displacement scale.
Radial distribution analysis provided a physical explanation. BCC has a relatively small separation between its first and second neighbor shells, making it more sensitive to shell mixing under vibrational disorder. The neural network appears to detect this structural ambiguity through the voxelized density field.
The Future: Future work can apply this idea to molecular-dynamics trajectories, experimental structural data, liquids, glasses, defects, and phase transformations in real materials. A key question is whether learned order parameters can complement traditional experimental variables and reveal intermediate states that are difficult to capture with a single scalar criterion.
The Impact: This work suggests that AI models can do more than classify physical structures. Their activations and output scores can provide continuous, data-driven order parameters that separate the loss of specific lattice identity from the loss of generic solid-like order. This may offer a useful route for studying melting-like disordering, structural similarity, and phase transitions in complex materials.
The research has been published in AI for Science.
Citation: Enji Li, Siyu Hu, Xiao Tian, Wentao Zhang, Tao Liu, Linwang Wang, Weile Jia. AtomFlow: Generative Prediction of Chemical Reactions in 3D Space[J]. AI for Science. DOI: 10.1088/3050-287X/ae9e53
