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 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 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 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 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