Active Learning of Anharmonic Phonon Transport Across Carbon Allotropes

Machine-learning potentials (MLPs) face significant challenges in sampling anharmonic lattice dynamics due to computational bottlenecks. This work presents an active learning (AL) framework that integrates pretrained potentials with committee models to overcome these limitations. This approach achieves a 59.9% reduction in force errors while maintaining exceptional transferability across diverse carbon allotropes. The optimized potential model reveals distinct anharmonic regimes: in diamond, the longitudinal acoustic phonons exhibit a temperature-induced sign reversal in their lifetime scaling exponent (from n = -1.54 at 200 K to n = 0.17 at 800 K), marking a fundamental transition from Umklapp scattering to fluctuation-induced localization where thermal disorder preferentially suppresses low-frequency phonon propagation; in bilayer graphene, the strongly negative Grüneisen parameters of flexural acoustic (ZA) phonons [γ(q) = -33.7 to -8.4] reveal anomalous vibrational stiffening under volume expansion, originating from stress release in the interlayer coupling that enhances in-plane restoring forces. These findings establish a general strategy for MLPs development and uncover emergent phonon phenomena, providing new pathways for controlling thermal transport in quantum materials and nanoscale devices.