Metallic glasses (MGs) have an amorphous atomic arrangement, but their structure and dynamics in the nanoscale are not homogeneous. Numerous studies have confirmed that the static and dynamic heterogeneities of MGs are vital for their deformation mechanism. The “defects” in MGs are envisaged to be structurally loosely packed and dynamically active to external stimuli. To date, no definite structure-property relationship has been established to identify liquid-like “defects” in MGs. In this paper, we proposed a machine-learned “defects” from atomic trajectories rather than static structural signatures. We analyzed the atomic motion behavior at different temperatures via a k-nearest neighbors machine learning model, and quantified the dynamics of individual atoms as the machine-learned temperature. Applying this new temperature-like parameter to MGs under stress-induced flow, we can recognize which atoms respond like “liquids” to the applied loads. The evolution of liquid-like regions reveals the dynamic origin of plasticity (thermo- and acousto-plasticity) of MGs and the correlation between stress-induced heterogeneity and local environment around atoms, providing new insights into thermo- and acousto-plastic forming.

金屬玻璃具有無序的原子排列, 但其結構與動力學並非各處均勻. 許多研究證實金屬玻璃的結構與動態不均勻性對於其塑性機制至關重要. 金屬玻璃的“缺陷”被視為結構上疏鬆排布、動力學上積極回應外界刺激的區域. 但迄今仍未建立明確的結構-性能關係來甄別金屬玻璃中的類液缺陷. 本文中, 我們基於模擬原子運動軌跡並結合機器學習提出了一種不依賴於靜態結構特徵的缺陷. 利用k近鄰機器學習模型分析並預測了不同溫度下的原子運動行為, 建立了溫度類標籤-原子運動特徵映射關係. 應用這個“機器學習溫度”參數理解金屬玻璃在應力下的塑性流, 識別類液區原子. 類液區的演化揭示了金屬玻璃塑性的動態起源(包括熱塑性和超聲塑性), 展示了應力誘發的非均勻性和原子局域環境的關聯, 為熱塑性成型和超聲加工提供了新見解.