Research data for "Accelerated First-Principles Exploration of Structure and Reactivity in Graphene Oxide"

This dataset supports the paper: "Accelerated First-Principles Exploration of Structure and Reactivity in Graphene Oxide" (https://doi.org/10.1002/anie.202410088). The purpose is to enable readers to access the potential models and characterisation code for reproducing the work, and also to create structural models of their own of functionalised graphene sheets ("graphene oxide", GO). Contents The repository is structured in the following way: Functionalisation code: The functionalisation code uses four structural parameters (p1 to p4) to construct initial structural models of GO in a systematic way. Models: MACE model files, checkpoints for refitting and fine-tuning, training and testing databases at each iteration and the submission script for training. Structures: Structures after the 2 ns anneal from the three MD runs at 900, 1,200 and 1,500 K along with geometry optimised structures. Additional structure from 1.5 ns at 1,500 K is provided as given in Figure 3. Information regarding the iterations and DFT settings used for our MACE fit iter-0: Seeded in CASTEP-GAP (5x5 matrix) - all relevant structures were taken from the CASTEP-GAP runs up over 10 ps. A filter of any bond length < 0.5 and coordination > 6 was applied to remove high E/F structures. iter-1: iter-0 + structures held at 600K using MACE model fitted on iter-0. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-2: iter-1 + structures held at 900K using MACE model fitted on iter-1. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-3: iter-2 + structures held at 1200K using MACE model fitted on iter-2. This potential was unstable and led to failed runs at 0.40-0.00 and 0.50-0.25. As a result, all structures from every trajectory were taken and were filtered according to: min bond length < 0.5 were discarded, isolated atoms were deleted, no atoms with > 6 coordination. The structures were then downsampled using FPS to 250 structures. iter-4: iter-3 + structures held at 1500K using MACE model fitted on iter-3. This potential was unstable and led to failed runs at 0.50-0.00 and 0.50-0.25. As a result, all structures from every trajectory were taken and were filtered according to: min bond length < 0.5 were discarded, isolated atoms were deleted, no atoms with > 6 coordination. The structures were then downsampled using FPS to 250 structures. iter-5: iter-4 + structures held at 1500K using MACE model fitted on iter-4. This potential was unstable and led to failed runs at 0.40-0.00 and 0.50-0.75. As a result, all structures from every trajectory were taken and were filtered according to: min bond length < 0.5 were discarded, isolated atoms were deleted, no atoms with > 6 coordination. The structures were then downsampled using FPS to 250 structures. iter-6: iter-5 + structures held at 1500K using MACE model fitted on iter-5. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-7: iter-6 + structures held at 1500K using MACE model fitted on iter-5. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-8: iter-7 + structures held at 1500K using MACE model fitted on iter-5. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-9: iter-8 + edge structures held at 1500K using MACE model fitted on iter-8. Initial structures were sampled across p1 and p3 from 0.1-0.5 (p2 = 0.5). 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-10: iter-9 + edge structures held at 1500K using MACE model fitted on iter-9. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-11: iter-10 + edge structures held at 1500K using MACE model fitted on iter-10. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-12: iter-11 + edge structures held at 1500K using MACE model fitted on iter-11. 250 structures were added (1 ps intervals (10 ps total) on 5x5 matrix). Structures were cleaned where single atoms were removed. iter-12 was cleaned up by removing any structures with forces > 50 eV/A. 7 Structures were removed from the training data and 0 from the test data. The production model can be found in iter-12-final-model. The final database can be found in iter-12-clean/structures/iter-12-train-filtered.xyz. MACE fitting settings: --name="MACE_model" \ --train_file="" \ --valid_fraction=0.10 \ --test_file="" \ --config_type_weights='{"Default":1.0}' \ --E0s='{1:-13.59395639138, 6:-148.314002, 8:-432.8647463978}' \ --model="MACE" \ --hidden_irreps='128x0e' \ --loss='huber' \ --r_max=3.7 \ --batch_size=25 \ --max_num_epochs=1200 \ --swa \ --default_dtype='float32' \ --energy_key='QM_energy' \ --forces_key='QM_forces' \ --stress_key=None \ --start_swa=500 \ --ema \ --ema_decay=0.99 \ --amsgrad \ --restart_latest \ --device=cuda \ --seed=123 \ DFT settings (CASTEP 23.1): calculate_stress = false popn_calculate = false xc_functional PBE spin_polarized : false mixing_scheme : Pulay cut_off_energy = 550 eV elec_energy_tol = 1e-5 eV max_scf_cycles 200 fix_occupancy false opt_strategy speed smearing_scheme Gaussian smearing_width 0.1 eV WRITE_CHECKPOINT : MINIMAL本数据集配套论文《氧化石墨烯结构与反应性的第一性原理加速探索》（https://doi.org/10.1002/anie.202410088）。 本数据集旨在支持读者复现该研究工作，使其能够获取相关势函数模型与表征代码，同时支持用户自行构建官能化石墨烯片层（氧化石墨烯，GO）的结构模型。 ## 数据集内容 本数据集仓库的组织架构如下： ### 官能化代码 该代码通过4个结构参数（p1至p4），以系统化方式构建氧化石墨烯的初始结构模型。 ### 模型文件 包含MACE模型文件、用于重拟合与微调的检查点文件、各迭代轮次的训练与测试数据库，以及训练提交脚本。 ### 结构文件 包含在900 K、1200 K与1500 K下开展的三次分子动力学（MD）模拟经2 ns退火后的结构，以及几何优化后的结构。此外还提供了1500 K下1.5 ns模拟得到的结构，对应论文图3中的内容。 --- ## MACE拟合相关的迭代轮次与密度泛函理论（DFT）设置说明 ### 迭代轮次说明 iter-0：基于CASTEP-GAP构建初始数据集（5×5超胞）——所有相关结构取自CASTEP-GAP模拟时长超过10 ps的结果。我们设置过滤规则：移除所有键长小于0.5且配位数大于6的结构，以去除高能、高受力的异常结构。 iter-1：在iter-0数据集基础上，添加使用iter-0拟合得到的MACE模型在600 K下模拟得到的结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-2：在iter-1数据集基础上，添加使用iter-1拟合得到的MACE模型在900 K下模拟得到的结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-3：在iter-2数据集基础上，添加使用iter-2拟合得到的MACE模型在1200 K下模拟得到的结构。该势函数存在不稳定性，导致0.40-0.00与0.50-0.25时段的模拟运行失败。因此我们采集了所有轨迹中的结构，并按照以下规则过滤：移除键长小于0.5的结构、删除孤立原子、排除配位数大于6的原子。随后通过最远点采样（FPS）将结构下采样至250个。 iter-4：在iter-3数据集基础上，添加使用iter-3拟合得到的MACE模型在1500 K下模拟得到的结构。该势函数存在不稳定性，导致0.50-0.00与0.50-0.25时段的模拟运行失败。因此我们采集了所有轨迹中的结构，并按照以下规则过滤：移除键长小于0.5的结构、删除孤立原子、排除配位数大于6的原子。随后通过最远点采样（FPS）将结构下采样至250个。 iter-5：在iter-4数据集基础上，添加使用iter-4拟合得到的MACE模型在1500 K下模拟得到的结构。该势函数存在不稳定性，导致0.40-0.00与0.50-0.75时段的模拟运行失败。因此我们采集了所有轨迹中的结构，并按照以下规则过滤：移除键长小于0.5的结构、删除孤立原子、排除配位数大于6的原子。随后通过最远点采样（FPS）将结构下采样至250个。 iter-6：在iter-5数据集基础上，添加使用iter-5拟合得到的MACE模型在1500 K下模拟得到的结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-7：在iter-6数据集基础上，添加使用iter-5拟合得到的MACE模型在1500 K下模拟得到的结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-8：在iter-7数据集基础上，添加使用iter-5拟合得到的MACE模型在1500 K下模拟得到的结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-9：在iter-8数据集基础上，添加使用iter-8拟合得到的MACE模型在1500 K下模拟得到的边缘结构。初始结构通过在p1与p3参数0.1~0.5范围内采样得到（p2=0.5）。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-10：在iter-9数据集基础上，添加使用iter-9拟合得到的MACE模型在1500 K下模拟得到的边缘结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-11：在iter-10数据集基础上，添加使用iter-10拟合得到的MACE模型在1500 K下模拟得到的边缘结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。 iter-12：在iter-11数据集基础上，添加使用iter-11拟合得到的MACE模型在1500 K下模拟得到的边缘结构。本次在5×5超胞上以1 ps为间隔（总时长10 ps）采集了250个结构，并通过移除孤立原子完成结构清洗。此外，iter-12还额外进行了清洗：移除所有受力大于50 eV/Å的结构，最终从训练集中移除7个结构，测试集无结构被移除。 最终生产模型位于iter-12-final-model目录下，最终数据库位于iter-12-clean/structures/iter-12-train-filtered.xyz文件中。 --- ## MACE模型拟合参数设置 bash --name="MACE_model" --train_file="" --valid_fraction=0.10 --test_file="" --config_type_weights='{"Default":1.0}' --E0s='{1:-13.59395639138, 6:-148.314002, 8:-432.8647463978}' --model="MACE" --hidden_irreps='128x0e' --loss='huber' --r_max=3.7 --batch_size=25 --max_num_epochs=1200 --swa --default_dtype='float32' --energy_key='QM_energy' --forces_key='QM_forces' --stress_key=None --start_swa=500 --ema --ema_decay=0.99 --amsgrad --restart_latest --device=cuda --seed=123 --- ## 密度泛函理论（DFT）设置（CASTEP 23.1版本） calculate_stress = false popn_calculate = false xc_functional PBE spin_polarized : false mixing_scheme : Pulay cut_off_energy = 550 eV elec_energy_tol = 1e-5 eV max_scf_cycles 200 fix_occupancy false opt_strategy speed smearing_scheme Gaussian smearing_width 0.1 eV WRITE_CHECKPOINT : MINIMAL