Plastic Heterogeneity Data of CuZr/AlSm/NiNb/FeP Metallic Glasses

This group of datasets contain the atomic-scale plastic heterogeneity data of a series of Cu-Zr (Cu₆₅Zr₃₅, Cu₅₀Zr₅₀, Cu₈₀Zr₂₀), Al₉₀Sm₁₀, Ni₆₂Nb₃₈ and Fe₈₀P₂₀ metallic glasses (in CSV format). In particular, a row of each CSV file corresponds to an atom, and includes the following information:<b>["id"]</b> atom id;<b>["type"]</b> atom type;<b>["x"]</b> coordinate x of the atom in the undeformed, quenched configuration;<b>["y"]</b> coordinate y of the atom in the undeformed, quenched configuration;<b>["z"]</b> coordinate z of the atom in the undeformed, quenched configuration;<b>["nonaffine_displacement"]</b> non-affine displacement of the atom at a compressive strain of 4.0% with reference to the original quenched configuration;<b>["note"]</b> denotes whether the atom is within the two compressive ends ("compr_ends") or close to the boundary ("bds"). After featurization, these atomic data could be eliminated from later machine learning.We simulate liquid melt quenching and deformation of Cu₆₅Zr₃₅, Cu₅₀Zr₅₀, Cu₈₀Zr₂₀, Ni₆₂Nb₃₈, Al₉₀Sm₁₀ and Fe₈₀P₂₀ metallic glasses using molecular dynamics simulations. We use 3 quenching rates of 5 × 10¹⁰, 5 × 10¹¹ and 5 × 10¹² K s^-1 for Cu-Zr metallic glasses, and 5 × 10¹⁰ K s^-1 for the other metallic glasses. We construct 3 large slab samples for each Cu-Zr glass, each of which contains 345600 atoms with dimensions ~120 (X) × 24 (Y) ×240 (Z) ų. Data from 2 glass samples are concatenated, equally undersampled and used in 5-fold cross-validation training the ML models, whereas the remaining sample is set-aside for rigorous generalization tests. For Ni₆₂Nb₃₈, Al₉₀Sm₁₀ and Fe₈₀P₂₀ metallic glasses, we construct samples of 131072 atoms. The timestep is 1 fs. During simulation, the initial configuration is built by randomly substituting into an fcc (Cu-Zr, N_i62Nb₃₈ and Al₉₀Sm₁₀) or bcc (Fe₈₀P₂₀) lattice. The samples are annealed at 2000 K for 1 ns, quenched to 50 K with each quenching rate, and relaxed at 50 K for 1 ns. After quenching, the Cu-Zr MGs are compressed along Z axis under a strain rate of 2.5 × 10⁷ s^-1 in a quasi-static mode (constantly apply a small strain and then relax, up to the strain of 10%) at a low temperature of 50 K (see Supplementary Figure 3 for typical stress-strain curves). Periodic boundary conditions (PBCs) are imposed in Y and Z axes and free surfaces are applied along X axis to allow shear offsets. For Ni₆₂Nb₃₈, Al₉₀Sm₁₀ and Fe₈₀P₂₀ we simulate both tensile and compressive deformation with strain rates of 2.5 × 10⁷ s^-1 and 1.0 × 10⁸ s^-1 as well as with PBCs in all directions (the data of compressive deformation under 2.5 × 10⁷ s^-1 with PBC in all directions are presented here). After feature extraction, we select atoms of ~10 - 20 Å away from the surfaces or deformation ends to construct the ML datasets. We analyze non-affine displacement D² of each atom at a strain of 4.0% with reference to the undeformed configuration as a signature of plastic heterogeneity. Please see more details of D² in ML Falk, Phys. Rev. E 57, 7192-7205 (1998).<br> <b>Note on data source</b>This data is part of our paper:Qi Wang, Anubhav Jain. A transferable machine-learning framework linking interstice distribution and plastic heterogeneity in metallic glasses. Nature Communications 10, 5537 (2019). https://doi.org/10.1038/s41467-019-13511-9. If you found this dataset useful and would like to cite it in your work, please be sure to cite its original sources below rather than or in addition to this page.

本数据集组包含一系列铜锆（Cu₆₅Zr₃5、Cu₅₀Zr₅0、Cu₈₀Zr₂0）、Al₉₀Sm₁0、Ni₆₂Nb₃8以及Fe₈₀P₂0金属玻璃的原子尺度塑性异质性数据，格式为CSV文件。具体而言，每个CSV文件的一行对应一个原子，包含以下信息：
**["id"]**：原子编号；
**["type"]**：原子种类；
**["x"]**：原子在未变形淬火态下的x坐标；
**["y"]**：原子在未变形淬火态下的y坐标；
**["z"]**：原子在未变形淬火态下的z坐标；
**["nonaffine_displacement"]**：以原始淬火态为参考，在4.0%压缩应变下原子的非仿射位移（non-affine displacement）；
**["note"]**：标记原子是否位于两个压缩端（"compr_ends"）或靠近边界（"bds"）。完成特征化处理后，这些原子数据可在后续机器学习流程中剔除。

本研究通过分子动力学模拟（molecular dynamics simulations），制备了Cu₆₅Zr₃5、Cu₅₀Zr₅0、Cu₈₀Zr₂0、Ni₆₂Nb₃8、Al₉₀Sm₁0及Fe₈₀P₂0金属玻璃的熔体淬火过程与变形行为。其中铜锆基金属玻璃采用3种淬火速率：5×10¹⁰、5×10¹¹和5×10¹² K·s⁻¹，其余金属玻璃仅采用5×10¹⁰ K·s⁻¹的淬火速率。针对每种铜锆基金属玻璃，我们构建了3个大尺寸薄板试样，每个试样包含345600个原子，尺寸约为120(X)×24(Y)×240(Z) ų。将2个玻璃试样的数据进行拼接、均等欠采样后，用于5折交叉验证以训练机器学习（ML）模型，剩余1个试样留作严格的泛化性能测试。对于Ni₆₂Nb₃8、Al₉₀Sm₁0和Fe₈₀P₂0金属玻璃，我们构建的试样包含131072个原子。模拟的时间步长为1 fs。

模拟初始时，我们通过在面心立方（face-centered cubic, fcc，适用于Cu-Zr、Ni₆₂Nb₃8及Al₉₀Sm₁0）或体心立方（body-centered cubic, bcc，适用于Fe₈₀P₂0）晶格中随机替换原子来构建初始构型。将试样在2000 K下退火1 ns，随后以各指定淬火速率冷却至50 K，并在50 K下弛豫1 ns。淬火完成后，铜锆基金属玻璃在50 K低温下，以2.5×10⁷ s⁻¹的应变速率沿Z轴进行准静态压缩（依次施加小应变并弛豫，直至总应变达到10%，典型应力-应变曲线参见补充图3）。模拟中，Y、Z轴施加周期性边界条件（Periodic Boundary Conditions, PBCs），X轴采用自由表面以允许剪切偏移。

针对Ni₆₂Nb₃8、Al₉₀Sm₁0和Fe₈₀P₂0金属玻璃，我们分别以2.5×10⁷ s⁻¹和1.0×10⁸ s⁻¹的应变速率模拟了拉伸与压缩变形，且所有方向均施加周期性边界条件（本文仅提供2.5×10⁷ s⁻¹应变速率下全周期边界条件的压缩变形数据）。完成特征提取后，我们选取距离表面或变形端约10~20 Å的原子以构建机器学习数据集。我们以未变形构型为参考，计算每个原子在4.0%应变下的非仿射位移D²，以此作为塑性异质性的特征指标。关于D²的更多细节可参见ML Falk发表于《物理评论E》第57卷，第7192-7205页（1998年）的研究。

**数据来源说明**
本数据集属于我们发表的论文：Qi Wang, Anubhav Jain. A transferable machine-learning framework linking interstice distribution and plastic heterogeneity in metallic glasses. *Nature Communications* 10, 5537 (2019). https://doi.org/10.1038/s41467-019-13511-9。若您认为本数据集对您的研究有帮助并希望引用，请务必引用其原始文献，而非仅引用本数据集页面。