環状高分子のすり抜け絡み合いを機械学習で検出

Dense melts of nonconcatenated ring polymers exhibit anomalous dynamics and nonlinear rheology, in some cases linked to long-lived inter-ring threading entanglement. Detecting these geometric constraints remains computationally demanding, with different methods being based on different geometric features tuned to specific models. Here, we introduce a machine-learning framework that detects threading directly from the writhe of ring conformations and we demonstrate its broad generalisability. Trained on simulations of isolated, distance-constrained pairs of short rings, a one-dimensional convolutional neural network (CNN1D) operating on the writhe profile classifies threading states with over 98% accuracy on held-out ring pairs. Remarkably, this model generalises without retraining to equilibrium dense melts spanning a wide range of densities and chain lengths (N = 100 to 1600), maintaining true-positive rates above 91% with false-positive rates near zero. Misclassified conformations are limited to physically ambiguous, shallow threading events which are arguably not impacting the dynamics of the rings. Effectively, writhe captures the interaction between the two rings, allowing writhe-based classifiers to improve the topological analysis of large ring-polymer systems with a 3-fold speed-up over the state-of-the-art minimal-surface detection. These results establish the writhe-trained neural networks as an accurate, scalable tool for characterising entanglement in ring polymer melts.