Additional data for the polyanion sodium cathode materials dataset

As an addition to the Polyanion sodium cathode materials dataset (https://doi.org/10.11583/DTU.27202446), we have used three computational methods used for battery electrodes to generate 88,535 structures.We have performed atomic structural relaxation of four sodium ion polyanionic cathode materials NaTMPO₄(olivine) ,NaTMPO₄(maricite), Na₂TMSiO₄ and Na_2.56TM_1.72(SO₄)₃ with three and four different types of transition metal ions (TM) in each structure. The four possible transition metal ions in the cathode materials are Fe, Mn, Co and Ni and the concentration level of the transition metal ion and Na atoms differs from each structure. The dataset consist of 514 structure optimized.We have also performed ab-initio molecular dynamics simulation (AIMD) of the cathode materials NaTMPO₄(olivine) and NaTMPO₄(maricite) with three different pairs of transition metal ions (TM=Fe_1-yMn_y, TM=Fe_1-yNi_yand TM=Mn_1-yCo_y) with different concentration of transition metal ion and Na. The dataset consist of 87967 structural trajectory points.We have performed nudged elastic band (NEB) calculation of six cathode materials, FePO₄(olivine), MnPO₄(olivine), NiPO₄(olivine), NaFePO₄(olivine), NaMnPO₄(olivine) and NaNiPO₄(olivine). The NEB calculation estimates the energy barrier a Na atom needs to overcome when moving from one spot to a vacant position. The dataset consist of 36 structural trajectory point along the energy barriers noted by "_NEB" and 12 structure points for the initial and final configuration of the diffusion path.For each sampled structure, we record its crystal composition, total energy, atom-wise force vectors, atom-wise magnetic moments, and atomic charges obtained through Bader analysis.All computational calculation are done using density functional theory (DFT) calculation with the Vienna Ab initio simulation package (VASP) version 6.4 package. The Perdew-Burke-Ernzerhof (PBE) functional with Hubbard-U corrections were appliedwas utilized for all calculations. The U-values are similar to the ones used for materials project (Fe: 5.3eV, Mn: 3.9eV, Co: 3.32eV, Ni: 6.2eV). For all calculations, an energy cutoff of 520eV was applied, with a smearing width of 0.01eV and convergence criteria set to 1e-5eV for energy and 0.03eV/Å for forces. All calculations were performed with spin polarization. The k-points employed for the four materials were fixed, with NaMPO₄(olivine) and NaMPO₄(maricite) utilizing [3,4,6] gamma points, Na₂MSiO₄employing [3,4,4] gamma points and Na_2.56M_1.72(SO₄)₃ utilizing [2,3,4] gamma points. When constructing supercells, the gamma point in the direction of cell enlargement was halved. These settings match the ones used for the Polyanion sodium cathode materials dataset (https://doi.org/10.11583/DTU.27202446)All AIMD simulations are conducted using the Langevin thermostat with a friction constant of 0.003. The temperature is maintained at 1000K to facilitate diffusion events, and a time step of 1fs as employed throughout the simulations. All simulations are executed within the canonical (NVT) ensemble and a sample frequency was set to 1fs.All NEB calculations are performed using the ASE version 3.23.0 NEB wrapper with the FIRE optimization algorithm. Five intermediate images were used in the NEB optimization, converging to a maximum atom-wise force of 0.03ev/Å. The initial and final images were manually determined based on the cation's redox position, followed by structural optimization before initiating the NEB calculations. Throughout both structural optimizations and NEB calculations, the cell parameters were kept constant to the theoretical optimized experimental cell parameters.The dataset is presented in XYZ format. The dataset is divided into three folder, one for each computational method.<br>To extract structural compositions and physical properties, the ase.io.read function from ASE version 3.23.0 is used. An example of how to extract data and plot the physical properties is provided in https://github.com/dtu-energy/cPaiNN/blob/main/read_data.py and https://github.com/dtu-energy/cPaiNN/blob/main/attach_bader_charge.ipynb shows how to attached Bader charges to an ASE atom object.<br>To cite the data please use the doi https://doi.org/110.11583/DTU.27411681<br>Along with the data is all the pretrained cPaiNN models and PaiNN models used to test these datasets. To load the pretrained models, please use the descriptions presented here: https://github.com/dtu-energy/cPaiNN<br>Finetuned model and model trained from scratch for CHGNet and Mace-MP-0 large are also included.<br>Versions:<br>Use the data from version 42: All data is uploaded and pretrained model is trained on structures with atomic charge<br>3: The NEB data is updated to consist of theoretical optimized cell parameters instead of experimental parameters. Moreover, are all pretrained model updated to be trained on structures with and without atomic charge, but still predicts atomic charge4: Included more pretrained models as well as the initial and final configuration of the structural optimization. The atomic charge has also changed sign.<br>

作为聚阴离子钠正极材料数据集（https://doi.org/10.11583/DTU.27202446）的补充，我们采用了三种适用于电池电极的计算方法，生成了88,535个结构。
我们对四种钠离子聚阴离子正极材料（NaTMPO₄（橄榄石型）、NaTMPO₄（马氏体）、Na₂TMSiO₄及Na_2.56TM_1.72(SO₄)₃）进行了原子结构弛豫计算，每种结构中包含3至4种不同的过渡金属离子（TM）。正极材料中可能存在的四种过渡金属离子为Fe、Mn、Co和Ni，且过渡金属离子与Na原子的浓度水平在各结构中存在差异。该数据集包含514个优化后的结构。
我们还对正极材料NaTMPO₄（橄榄石型）和NaTMPO₄（马氏体）进行了从头算分子动力学模拟（AIMD），涉及三种不同的过渡金属离子对（TM=Fe_1-yMn_y、TM=Fe_1-yNi_y及TM=Mn_1-yCo_y），且过渡金属离子与Na的浓度各不相同。该数据集包含87,967个结构轨迹点。
我们对六种正极材料（FePO₄（橄榄石型）、MnPO₄（橄榄石型）、NiPO₄（橄榄石型）、NaFePO₄（橄榄石型）、NaMnPO₄（橄榄石型）及NaNiPO₄（橄榄石型））进行了nudged elastic band（NEB）计算。NEB计算用于估算Na原子从一个位置迁移至空位时需克服的能垒。该数据集包含沿能垒分布的36个结构轨迹点（标记为“_NEB”），以及扩散路径初始与最终构型对应的12个结构点。
对于每个采样结构，我们记录其晶体组成、总能量、原子级力向量、原子级磁矩，以及通过Bader分析获得的原子电荷。
所有计算均基于密度泛函理论（DFT），使用Vienna Ab initio模拟包（VASP）6.4版本完成。所有计算均采用带Hubbard-U修正的Perdew-Burke-Ernzerhof（PBE）泛函。U值与Materials Project所用值一致（Fe：5.3eV，Mn：3.9eV，Co：3.32eV，Ni：6.2eV）。所有计算的能量截断值设为520eV，展宽为0.01eV，收敛判据为能量1e-5eV、力0.03eV/Å。所有计算均考虑自旋极化。四种材料的k点设置固定：NaMPO₄（橄榄石型）和NaMPO₄（马氏体）采用[3,4,6] Γ点，Na₂MSiO₄采用[3,4,4] Γ点，Na_2.56M_1.72(SO₄)₃采用[2,3,4] Γ点。构建超胞时，胞扩张方向的Γ点数量减半。这些设置与聚阴离子钠正极材料数据集（https://doi.org/10.11583/DTU.27202446）所用设置一致。
所有AIMD模拟均使用朗之万热浴（Langevin thermostat），摩擦常数为0.003。温度维持在1000K以促进扩散事件发生，模拟全程采用1fs的时间步长。所有模拟在正则系综（NVT）下进行，采样频率设为1fs。
所有NEB计算均使用ASE 3.23.0版本的NEB封装器，结合FIRE优化算法。NEB优化中使用5个中间图像，收敛至最大原子力0.03eV/Å。初始与最终图像基于阳离子的氧化还原位置手动确定，在启动NEB计算前需先进行结构优化。在结构优化与NEB计算过程中，胞参数均固定为理论优化的实验胞参数。
数据集以XYZ格式呈现，分为三个文件夹，每种计算方法对应一个文件夹。<br>
若需提取结构组成与物理性质，可使用ASE 3.23.0版本的ase.io.read函数。数据提取与物理性质绘图的示例见https://github.com/dtu-energy/cathode-generation-workflow/tree/main/extract_data/read_data.py；https://github.com/dtu-energy/cathode-generation-workflow/tree/main/extract_data/utils.py包含两个函数，一个用于将Bader电荷附加至ASE原子对象，另一个用于合并多个XYZ数据文件。<br>
引用本数据请使用DOI：https://doi.org/110.11583/DTU.27411681<br>
数据中还包含用于测试这些数据集的所有预训练cPaiNN模型。加载预训练模型请参考此处说明：https://github.com/dtu-energy/cPaiNN<br>
版本说明：<br>使用版本22的数据：所有数据已上传，预训练模型基于含原子电荷的结构训练<br>版本3：NEB数据已更新为理论优化的胞参数（替代实验参数）；此外，所有预训练模型已更新，可基于含或不含原子电荷的结构训练，但仍能预测原子电荷