WorksPhysics and Astronomy
Efficient and Accurate Machine Learning Interatomic Potential for Graphene: Capturing Stress-Strain and Vibrational Properties
Abstract
Machine learning interatomic potentials (MLIPs) offer an efficient and accurate framework for large-scale molecular dynamics (MD) simulations, effectively bridging the gap between classical force fields and ab initio methods. In this work, we present a reactive MLIP for graphene, trained on an extensive dataset generated via ab initio molecular dynamics (AIMD) simulations. The model accurately reproduces key mechanical and vibrational properties, including stress-strain behavior, elastic constants, phonon dispersion, and vibrational density of states. Notably, it captures temperature-dependent fracture mechanisms and the emergence of linear acetylenic carbon chains upon tearing. The phonon analysis also reveals the expected quadratic ZA mode and excellent agreement with experimental and DFT benchmarks. Our MLIP scales linearly with system size, enabling simulations of large graphene sheets with ab initio-level precision. This work delivers a robust and transferable MLIP, alongside an accessible training workflow that can be extended to other materials.
Introduction
The trade-off between accuracy and computational efficiency is a long-standing challenge in materials modeling, as summarized by the “No Free Lunch” theorem.[1] While ab initio Molecular Dynamics (AIMD) methods based on Density Functional Theory (DFT) can achieve high precision, their computational cost with typical scaling, where represents the number of atoms, hinders their application to systems containing more than tens of thousands of atoms even on high-performance computers.[2, 3] In contrast, classical Molecular Dynamics (MD) can simulate significantly larger systems, with hundreds of thousands of atoms; however, this comes at the cost
with typical scaling, where represents the number of atoms, hinders their application to systems containing more than tens of thousands of atoms even on high-performance computers.[2, 3] In contrast, classical Molecular Dynamics (MD) can simulate significantly larger systems, with hundreds of thousands of atoms; however, this comes at the cost of accuracy, which heavily depends on the force field parametrization.[4, 5] Force fields offer a spectrum of complexity, ranging from the computationally demanding reactive potentials,[6] capable of reproducing bond breaking and formation, to cheaper classical potentials, without reactivity considerations.[7] Furthermore, a non-negligible challenge in classical MD simulations is the task-specific calibration of force fields. This calibration can complicate the achievement of optimal agreement with experiment or DFT for various properties, including thermal,[8] vibrational, or mechanical,[9] when relying on a single parametrization optimized for a specific application.
Machine learning interatomic potentials (MLIPs) have emerged as a promising solution to bridge this gap, offering both accuracy and scalability by replacing conventional MD force fields with models trained on ab initio datasets,[10] enabling large-scale simulations that retain quantum-mechanical accuracy.[11, 12, 13] This approach applied to materials with well-established properties and current relevancy, such as graphene,[14] offers a compelling argument for the adoption of MLIPs, while also providing a more accurate and efficient pathway to improved MD simulations. In this context, previous studies have primarily focused on investigating lattice dynamics and phonon dispersion,[15] or on temperature-dependent mechanical properties.[16] However, to the best of our knowledge, no comprehensive effort has yet been made to develop a MLIP with DFT-accuracy that simultaneously captures both mechanical and vibrational properties, where collective and individual atomic effects are well represented, such as strain response, phonon dispersion, and shifts in vibrational spectra.
In this manuscript, we develop a MLIP for graphene and employ it to compute key mechanical and vibrational properties, such as: the elastic tensor, Young’s modulus, Poisson’s ratio, shear modulus, phonon dispersion curves, and vibrational density of states. The paper is organized as follows: § 2 outlines the computational methods; § 3 presents and discusses the results; and § 4 summarizes the study with the conclusions.
Methods
The workflow employed for the development of the graphene MLIP, as illustrated in Figure 1, comprised three stages.
As illustrated in Figure 1a, the relevant phase space was first explored by performing AIMD simulations on systems containing 4, 8, 16, and 32 C atoms, with the latter two considered in both and configurations, where and denote the cell’s dimensions along the and directions, respectively. Figure S1a-f depicts each of these structures. These simulations included NVT, NPT, and NVE ensembles, with varying strain rates and small random atomic displacements introduced to enhance configurational diversity. In the second stage (Figure 1b,c), we used the DeepMD framework[17] to develop a Deep Potential[18] based force field, using embedded atomic descriptors. Third, once satisfactory precision was achieved, the trained MLIP was deployed in large-scale MD simulations using LAMMPS,[19] targeting stress-driven behavior. This final stage is depicted in Figure 1d, where the deformation of a large graphene sheet is shown under applied strain.
Ab initio Molecular Dynamics. To generate a comprehensive dataset that captures the relevant potential energy surface (PES), we performed BornOppenheimer AIMD simulations using the projector-augmented wave method,[20] as implemented in the GPAW DFT calculator[21, 22] within the Python ASE framework.[23] The simulations were conducted using a -grid with a resolution of at least Å, a plane-wave cutoff of 500 eV, and the Local Density Approximation (LDA) exchange-correlation functional, chosen due to its relatively lower computational cost[24] and its reasonable agreement with more advanced functionals in graphene-based systems.[25]

Figure 1. (a) Illustration of the 32-C graphene structure used in the AIMD simulations. (b) Generation of the coordinate matrix, , and subsequent calculation of the descriptors, . The first element of , , is illustrated. (c) These descriptors served as input elements for the NN, trained to predict the total energy . (d) The trained MLIP was deployed as a force field in large-scale MD simulations. The panel depicts the application of strain on a large graphene sheet at K.
AIMD simulations were performed within three thermodynamic ensembles:[26] (i) the microcanonical (NVE) ensemble, employing the Velocity Verlet algorithm for time propagation; (ii) the canonical (NVT) ensemble at K, using a Langevin thermostat with a friction coefficient of fs; and (iii) the isothermal-isobaric (NPT) ensemble, which allows control over temperature and pressure, enabling the application of stress along various directions and at different strain rates.[27, 28]
All simulations used a timestep of fs, and a total of frames (positions, stress, forces, energy, and cell lengths) were collected for all system sizes. Additional frames were obtained via the application of random displacements of amplitude Å on the atomic positions of each geometry, except for the 32-atom system, where this was deemed unnecessary due to its inherent structural richness in comparison with the smaller systems. Simulation times ranged from 0.5 ps to 5 ps, and NPT runs contained strain rates between fs to fs. During these latter runs, tearing of the graphene structures was observed. Further details about each dataset are provided in Table S1.
Machine Learning Interatomic Potential
We employed the local frame two-body embedding descriptor[29] within the Deep Potential framework.[18] The local environment of each -th — comprising its position, the positions of its neighboring atoms, and their relative angles — is defined based on a cutoff radius, . This environment is then used to construct the coordinate matrix, ,
where is a switching function ensuring a smooth decay starting from the switching radius, ,
with . The coordinate matrix is then used to generate a set of descriptors, , for each atom via auxiliary embedding residual neural networks (ResNet).[30, 31] Each neighboring atom is processed through an embedding function , which receives the associate and is trained to enrich the descriptor by vectorizing the discrete input data. The resulting descriptors can be written as
These descriptors serve as inputs to the Deep Potential Neural Network (NN), , which is trained to predict individual atomic energies, . The NN is optimized using AIMD data, including atomic positions, forces, the virial tensor, the simulation cell, and total energy of the system. The overall objective of the NN is to predict the total energy of the entire system, ,
Once is defined as a function of , the atomic forces, , and virial tensor, , can be obtained via differentiation as and , where represents the atomic index, , and refer to Cartesian coordinates, and denotes the corresponding virial direction.
Given that part of our dataset includes strained graphene structures, we set the search and smooth decay radii to Å and Å. The size of the embedding matrix was chosen as 16, with three hidden layers containing 50, 100, and 200 neurons. The Deep Potential network consisted of three hidden layers of 240 neurons each. The starting and final loss pre-factors,[17] for energy, force, and virial were set as , respectively. Training was conducted for a total of steps, using an initial learning rate of with a decay of every steps. 20% of the data was reserved for validation.
The model’s performance was assessed using the root mean square error (RMSE) computed across all validation datasets, defined as
where denotes one of the physical properties (energy, force, or virial) predicted by the NN or obtained from AIMD data, evaluated over the same set of descriptors.
Molecular Dynamics. Classical MD simulations were performed using the LAMMPS software,[19] to study the stress-strain behavior and the vibrational properties. Unless stated otherwise, simulations employed a rectangular graphene sheet containing 9072 carbon atoms, with Å, Å, depicted in Figure S1g, with a timestep of fs. For the stress-strain analysis, we employed two approaches. First, to obtain deformation curves up to the breaking point, uniaxial strain was applied along both the and directions at a strain rate of , maintaining the system in an NPT ensemble at K and K. Second, to extract the elastic tensor, an initial energy minimization at K ensured a mechanically stable configuration, and then a series of small finite deformations () was applied, keeping the system within the linear elastic regime. The elastic constants were calculated using the finite deformation method, with the symmetric pressure tensor computed as in Refs.,[32, 33]
where stands for the cell volume, , the number of atoms, , the -th component of the position vector of atom , and , the -th component of the force acting on the same atom. Given the definition of the stress tensor as , the elastic constants were extracted using the generalized Hooke’s law, , where represents the stiffness tensor components and the applied strain. Then, the independent elastic constants and were obtained in Voigt notation. Based on these, the Young’s modulus and Poisson’s ratio were determined as and respectively.[9]
In addition to mechanical properties, vibrational properties were analyzed through phonon dispersion and the vibrational density of states (VDOS). The phonon dispersion curve was obtained using Phonopy’s[34] interface with LAMMPS, using a supercell based on graphene’s hexagonal unit cell. The VDOS was extracted from the Fourier transform of the velocity-velocity autocorrelation function,
which provides insight into the first-order Raman spectrum even without explicit projection of DFT-derived vibrational modes.[35, 36, 37] The VDOS calculations were performed in the NVT ensemble at K, with a 25 ps thermalization step followed by 1 ps time averages. Simulations were carried out for both the undeformed and strained graphene sheets at , enabling comparison of the MLIP’s predictions with experimental observations of the red-shift in graphene’s optical spectrum upon uniaxial strain.[38, 39, 40] Fourier transforms were computed using NumPy’s[41] discrete Fourier transform package.
Results
MLIP training. A direct comparison between the NN predictions trained on the 32-C validation dataset and the reference AIMD results obtained via GPAW is shown in Figure 2. The predicted energy per atom, force components, and virial stress exhibit only minor deviations from the AIMD reference. These deviations remain consistently low across all properties, demonstrating the NN’s ability to reproduce the interatomic interactions of graphene with high accuracy. The quantitative agreement is further supported by the RMSE values reported in Table 1, which summarize the model’s performance across different system sizes.
Having established the accuracy of the trained NN, the next step was to deploy it in MD simulations. To further assess the viability of the MLIP, we evaluated its computational efficiency by measuring the time required to complete 100 NPT steps for various system sizes. As shown in Figure 3, the results exhibit a linear scaling behavior with the system size, , , confirming that our MLIP maintains computational efficiency even as the number of atoms increases. These findings underscore the model’s suitability for large-scale graphene simulations while preserving ab initio accuracy.

Figure 3. Average time required for 100 NPT steps as a function of system size. The trained MLIP exhibits linear scaling behavior, , as indicated by the dashed line.
Strain Engineering. Figure 4 depicts the stress-strain curves at two different temperatures (1 and 300 K), with stress applied independently along both directions. The MLIP successfully reproduces two expected results related to temperature and strain direction.[42] The temperature effect is evident in the difference in ultimate tensile strength for the same direction and different temperatures, with the fracture occurring earlier at higher temperatures. The directional effect appears in the curves at a given temperature: graphene strained along the direction (i.e., armchair orientation) exhibits greater ductility than when strained along the direction (zigzag orientation).

Figure 4. Stress-strain curves obtained using the MLIP. Solid lines correspond to simulations at K, while dashed lines represent results at K.
| # | [eV/atom] | [eV/Å] | [eV/Å atom] |
| 4 | |||
| 8 | |||
| 16 | |||
| 32 |
Table 1. Root mean square error (RMSE) of the MLIP model predictions for different system sizes used in the AIMD training set. The errors are reported for energy per atom (), force components (), and the virial stress tensor ().

Figure 2. Comparison between AIMD reference data and NN predictions (black scattered points), derived using AIMD simulation frames from the 32-C system as input: (a) energy per atom, (b) -component of atomic forces, and (c) -component of the virial stress tensor. Red dashed lines indicate perfect agreement.
Snapshots of the torn graphene sheets for K are shown in Figure 5, where we note, particularly in Figure 5b, the formation of linear acetylenic carbon (LAC) chains[43, 44] after the graphene sheets fractured.

Figure 5. Snapshots from strain simulations for K, with strain (a) and (b) . Atoms are colored according to their Von Mises stress.
Even though LACs exhibit mechanical properties that differ significantly from those of graphene,[45] their formation has been closely associated with applying strain to graphene. Previous studies have reported LAC formation, either by individually pulling atoms[46] or by assembling carbon atoms atop one another,[47] suggesting a close connection between mechanical deformation and the emergence of carbyne. We also note that these structures appeared in the torn regions of the graphene sheets at all temperatures, with a more pronounced presence when stress was applied in the direction, as shown in the videos available in the Supporting Information.
The elastic tensor components and derived mechanical properties obtained using our MLIP, i.e., the Young modulus, Poisson’s ratio, shear modulus and the constants and , are presented in Table 2 alongside reference values from the literature based on DFT, MD, and experimental studies. Regarding our results, we note the excellent agreement between the Poisson’s ratio, , and Young modulus, , obtained experimentally and via DFT – Refs. [50] and [51, 52], respectively. We also note that the discrepancies observed in the elastic constants and shear modulus, when compared to previous MD results, can be attributed to the known tendency of conventional MD force fields to model graphene as a significantly “squishier” material.[9]
Table 2. Comparison of elastic properties obtained in this work with values from previous studies. The table includes the elastic constants and , shear modulus , Poisson’s ratio, , and Young’s modulus . All dimensional quantities are reported in TPa. MM stands for Molecular Mechanics.
Vibrational Analysis. Finally, we extended our investigation of the MLIP beyond its original training scope, focused on stress/strain properties, to evaluate its predictive capabilities for vibrational and thermal properties. Specifically, we examined the phonon band structure to characterize thermal behavior and analyzed the VDOS—which is inherently linked to Raman spectroscopy—to gain preliminary insights into graphene’s optical properties. This analysis provided an initial assessment of the model’s transferability to vibrational phenomena.
Figure 6 depicts the phonon band structure obtained using the MLIP, as well as experimental points from a highenergy electron energy-loss spectroscopy (HEELS) experiment, reproduced from Yanagisawa et al.[54] Notably, the MLIP captures the quadratic behavior of the in-plane ZA mode near the point, and shows good agreement with the experimental data, specially along the K- path.

Figure 6. Phonon dispersion of graphene obtained using the MLIP in LAMMPS. The in-plane (TO) and out-of-plane (ZO, ZA) phonon branches are labeled accordingly. Notably, the ZA vibrational mode exhibits quadratic behavior around point. The scattered open circles represent experimental data from Ref.[54]
Figure 7 presents the component-separated velocity-velocity autocorrelation function (VACF) and the corresponding VDOS for the carbon atom graphene system, with various strain values along . The corresponding results for strain applied in can be found in Figure S2. For the relaxed structures, the two most prominent peaks can be found in Figure 7b 1570.08 cm (dotted), 1476.77 cm (dashed), Figure 7d 1572.12 cm (dotted), 1473.75 cm (dashed) and Figure 7f 651.81 cm (dotted), 459.01 cm (dashed). As indicated by Figure 7a,Figure 7c, the in-plane vibrations (along and ) correspond to higher frequencies, typically associated with the upper region of the phonon dispersion, in agreement with Figure 6.

Figure 7. (left) and corresponding VDOS (right) for (a)–(b), (c)–(d), and (e)–(f). The values in all panels are shifted along the -axis for clarity. The color scheme indicates strain applied along , as shown in panel (a).
We also observed red (blue) frequency shift in Figure 7 for the in-plane (out-of-plane) VDOS upon the introduction of strain. The amplitudes of the shifts, from to on the greatest peaks are 100.06 cm and 69.75 cm for Figure 7b,Figure 7d, corresponding to the red-shift and 17.08 cm for Figure 7f, corresponding to the blue shift.
Remarkably, the in-plane vibrations, shown in Figure 7(a)-(Figure 7d), align well with pre vious investigations on graphene’s thermal conduction using Nonequilibrium MD,[55] as well as ab initio calculations on the vibrational spectrum.[51] Additionally, our results reproduce the red-shift observed on the (Raman equivalent) G peaks,[56] around , under uniaxial strain, consistent with experimental observations.[39, 57]
Overall, the strong agreement between the shifts observed here and other works in the literature validates the MLIP’s ability to capture vibrational anharmonicity and strain-induced effects, key features for reliable predictions of thermal conductivity and phonon-limited transport phenomena.
Conclusions
In this work, we developed a MLIP for graphene based on the Deep Potential framework. The model was trained using a diverse and comprehensive dataset generated via AIMD simulations under various thermodynamic conditions and strain regimes. This enabled the construction of a reactive, accurate, and computationally efficient potential that captures key mechanical and vibrational properties.
The MLIP demonstrated excellent agreement with AIMD and experimental data for stress–strain responses, elastic constants, phonon dispersion, and VDOS. Notably, it accurately captured temperature-dependent fracture behavior and the formation of linear acetylenic carbon (LAC) chains during tearing—an emergent phenomenon not trivially reproduced by traditional force fields. Additionally, the model correctly predicted key phonon features, such as the quadratic ZA mode and strain-induced red and blue shifts in the VDOS, highlighting its robustness and transferability.
Thanks to its linear computational scaling and high fidelity, the proposed MLIP enables large-scale molecular dynamics simulations of graphene with ab initio-level accuracy, making it a valuable tool for investigating thermal, mechanical, and phonon-mediated transport phenomena.
All training data and model parameters are publicly available through the NOMAD repository,[58] and the trained MLIP is hosted on GitHub to facilitate reproducibility and further developments.[59]
Future extensions may include expanding the training dataset to cover a wider range of temperature and defect configurations, as well as applying the workflow to other two-dimensional materials, paving the way for accurate, large-scale ML-driven simulations across diverse nanomaterials.
Acknowledgement The authors would like to thank CNPq and CAPES for financial support, the Center for Computing in Engineering and Sciences at Unicamp and the HPC Cluster Coaraci, made available under FAPESP grants 2013/08293-7 and 2019/17874-0, respectively, the Brazilian LNCC (Laboratório Nacional de Computação Científica), for access to the SDumont Cluster, under the SINA-PAD/2014 project 01.14.192.00. The authors also gratefully acknowledge the computing time made available to them on the high-performance computer “Lise” at the NHR Center NHR@ZIB. This center is jointly supported by the Federal Ministry of Education and Research and the state governments participating in the NHR (www.nhr-verein.de/unsere-partner).
Supporting Information for Efficient and Accurate Machine Learning Interatomic Potential for Graphene: Capturing Stress-Strain and Vibrational Properties
Felipe Hawthorne; Paulo R. E. Raulino; Ronaldo Rodrigues Pelá; Cristiano F. Woellner
Visualization of the structures

Figure 1 (PDF p. 17). Illustrations of the graphene structures used in the AIMD simulations, containing (a) 4, (b) 8, (c) 16, (d) 16, (e) 32, (f) 32 and (g) 9072 carbon atoms. The axis illustrated in panel (a) is shared across all others.
Detailed Dataset Frames
Table 1: Detailed information about the Born-Oppenheimer Molecular Dynamics performed to generate the dataset used to train the MLIP. The notation denotes the strain rate applied in the direction. The #disp./time values indicate how many different geometries were used in the simulations, and for how much time each of these simulations were performed on average.
| #-C | NPT (K) | NVT (K) | NVE | ||
| /time | /time | # disp./time | # disp./time | # disp./time | |
| 4 | 10/2 ps | 10/2 ps | 10/2.2 ps | ||
| 4 | 10/2 ps | 10/2 ps | 10/2.2 ps | ||
| 4 | 10/2 ps | 10/2 ps | 10/2.2 ps | ||
| 4 | 10/2 ps | 10/2 ps | 10/2.2 ps | ||
| 8 | 10/3 ps | 10/4.1 ps | 10/2.2 ps | ||
| 8 | 10/3 ps | 10/4.1 ps | 10/2.2 ps | ||
| 8 | 10/3 ps | 10/4.1 ps | 10/2.2 ps | ||
| 8 | 10/3 ps | 10/4.1 ps | 10/2.2 ps | ||
| 16 () | 10/2 ps | 10/1.6 ps | 10/2.2 ps | ||
| 16 () | 10/2 ps | 10/1.6 ps | 10/2.2 ps | ||
| 16 () | 10/2 ps | 10/1.6 ps | 10/2.2 ps | ||
| 16 () | 10/2 ps | 10/1.6 ps | 10/2.2 ps | ||
| 16 () | 10/1 ps | 10/0.5 ps | 10/1.1 ps | ||
| 16 () | 10/1 ps | 10/0.5 ps | 10/1.1 ps | ||
| 16 () | 10/1 ps | 10/0.5 ps | 10/1.1 ps | ||
| 16 () | 10/1 ps | 10/0.5 ps | 10/1.1 ps | ||
| 32 () | 1/1 ps | 1/1 ps | |||
| 32 () | /time; | 1/1 ps | 1/1 ps | ||
| 32 () | /time; | 1/1 ps | 1/1 ps | ||
| 32 () | /time; | 1/1 ps | 1/1 ps | ||
| 32 () | 1/1 ps | 1/1 ps | |||
| 32 () | /time; | 1/1 ps | 1/1 ps | ||
| 32 () | /time; | 1/1 ps | 1/1 ps | ||
| 32 () | /time; | 1/1 ps | 1/1 ps |
Table 1 (PDF p. 18).

Figure 2 (PDF p. 19). Corresponding velocity autocorrelation functions and VDOS (left and right panels, respectively) for graphene strain with .
Link to the videos
The videos of the strain applications at K and K can be found here: https://www.youtube.com/watch?v=dauxu3UeUh4.
References
- [1]Wolpert, D.; Macready, W. No free lunch theorems for optimization. IEEE Transactions on Evolutionary Computation 1997, 1, 67–82.DOI
- [2]Butler, K. T.; Frost, J. M.; Skelton, J. M.; Svane, K. L.; Walsh, A. Computational materials design of crystalline solids. Chemical Society Reviews 2016, 45, 6138–6146.pubmed.ncbi.nlm.nih.gov/26992173
- [3]Müller, T.; Sharma, S.; Gross, E.; Dewhurst, J. Extending Solid-State Calculations to Ultra-Long-Range Length Scales. Physical Review Letters 2020, 125, 256402.DOI
- [4]terials. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. 2020.arxiv.org/abs/2005.00223
- [5]Iftimie, R.; Minary, P.; Tuckerman, M. E. Ab initio molecular dynamics: Concepts, recent developments, and future trends. Proceedings of the National Academy of Sciences 2005, 102, 6654–6659.pubmed.ncbi.nlm.nih.gov/15870204
- [6]Senftle, T. P.; Hong, S.; Islam, M. M.; Kylasa, S. B.; Zheng, Y.; Shin, Y. K.; Junkermeier, C.; Engel-Herbert, R.; Janik, M. J.; Aktulga, H. M.; Verstraelen, T.; Grama, A.; van Duin, A. C. T. The ReaxFF reactive force-field: development, applications and future directions. npj Computational Materials 2016, 2.DOI
- [7]Rappe, A. K.; Casewit, C. J.; Colwell, K. S.; Goddard, W. A.; Skiff, W. M. UFF, a full periodic table force field for molecular mechanics and molecular dynamics simulations. Journal of the American Chemical Society 1992, 114, 10024–10035.DOI
- [8]de Lima, J.; Bastos, L. V.; Woellner, C. F. A comparative study of the electronic and thermal properties of α-graphyne and α-graphdiyne nanoscrolls: Implications for thermoelectric applications. Computational Materials Science 2025, 250, 113728.DOI
- [9]Tsai, J.-L.; Tu, J.-F. Characterizing mechanical properties of graphite using molecular dynamics simulation. Materials and Design 2010, 31, 194–199.DOI
- [10]Friederich, P.; Häse, F.; Proppe, J.; Aspuru-Guzik, A. Machine-learned potentials for next-generation matter simulations. Nature Materials 2021, 20, 750–761.
- [11]Fedik, N.; Zubatyuk, R.; Kulichenko, M.; Lubbers, N.; Smith, J. S.; Nebgen, B.; Messerly, R.; Li, Y. W.; Boldyrev, A. I.; Barros, K.; Isayev, O.; Tretiak, S. Extending machine learning beyond interatomic potentials for predicting molecular properties. Nature Reviews Chemistry 2022, 6, 653–672.DOI
- [12]Wang, G.; Wang, C.; Zhang, X.; Li, Z.; Zhou, J.; Sun, Z. Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations. iScience 2024, 27, 109673.DOI
- [13]Unke, O. T.; Chmiela, S.; Sauceda, H. E.; Gastegger, M.; Poltavsky, I.; Schütt, K. T.; Tkat͡chenko, A.; Müller, K.-R. Machine Learning Force Fields. Chemical Reviews 2021, 121, 10142–10186.DOI
- [14]Urade, A. R.; Lahiri, I.; Suresh, K. S. Graphene Properties, Synthesis and Applications: A Review. JOM 2022, 75, 614–630.DOI
- [15]Rowe, P.; Csányi, G.; Alfè, D.; Michaelides, A. Development of a machine learning potential for graphene. Physical Review B 2018, 97, 054303.DOI
- [16]Singh, A.; Li, Y. Reliable machine learning potentials based on artificial neural network for graphene. Computational Materials Science 2023, 227, 112272.DOI
- [17]Zeng, J.; Zhang, D.; Lu, D.; Mo, P.; Li, Z.; Chen, Y.; Rynik, M.; Huang, L.; Li, Z.; Shi, S.; Wang, Y.; Ye, H.; Tuo, P.; Yang, J.; Ding, Y.; Li, Y.; Tisi, D.; Zeng, Q.; Bao, H.; Xia, Y.; Huang, J.; Muraoka, K.; Wang, Y.; Chang, J.; Yuan, F.; Bore, S. L.; Cai, C.; Lin, Y.; Wang, B.; Xu, J.; Zhu, J.-X.; Luo, C.; Zhang, Y.; Goodall, R. E. A.; Liang, W.; Singh, A. K.; Yao, S.; Zhang, J.; Wentzcovitch, R.; Han, J.; Liu, J.; Jia, W.; York, D. M.; E, W.; Car, R.; Zhang, L.; Wang, H. DeepMD-kit v2: A software package for deep potential models. The Journal of Chemical Physics 2023, 159.DOI
- [18]Zhang, L.; Han, J.; Wang, H.; Car, R.; E, W. Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics. Physical Review Letters 2018, 120, 143001.DOI
- [19]Thompson, A. P.; Aktulga, H. M.; Berger, R.; Bolintineanu, D. S.; Brown, W. M.; Crozier, P. S.; in ’t Veld, P. J.; Kohlmeyer, A.; Moore, S. G.; Nguyen, T. D.; Shan, R.; Stevens, M. J.; Tranchida, J.; Trott, C.; Plimpton, S. J. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comp. Phys. Comm. 2022, 271, 108171.DOI
- [20]Blöchl, P. E. Projector augmented-wave method. Physical Review B 1994, 50, 17953–17979.pubmed.ncbi.nlm.nih.gov/9976227
- [21]Enkovaara, J.; Rostgaard, C.; Mortensen, J. J.; Chen, J.; Dulak, M.; Ferrighi, L.; Gavnholt, J.; Glinsvad, C.; Haikola, V.; Hansen, H. A.; Kristoffersen, H. H.; Kuisma, M.; Larsen, A. H.; Lehtovaara, L.; Ljungberg, M.; Lopez-Acevedo, O.; Moses, P. G.; Ojanen, J.; Olsen, T.; Petzold, V.; Romero, N. A.; Staussholm-Møller, J.; Strange, M.; T ritsaris, G. A.; Vanin, M.; Walter, M.; Hammer, B.; Häkkinen, H.; Madsen, G. K. H.; Nieminen, R. M.; Nørskov, J. K.; Puska, M.; Rantala, T. T.; Schiøtz, J.; Thygesen, K. S.; Jacobsen, K. W. Electronic structure calculations with GPAW: a real-space implementation of the projector augmented-wave method. Journal of Physics: Condensed Matter 2010, 22, 253202.
- [22]Mortensen, J. J. r.; Larsen, A. H.; Kuisma, M.; Ivanov, A. V.; Taghizadeh, A.; Peterson, A.; Haldar, A.; Dohn, A. O.; Schäfer, C.; Jónsson, E. O.; Hermes, E. D.; Nilsson, F. A.; Kastlunger, G.; Levi, G.; Jónsson, H.; Häkkinen, H.; Fojt, J.; Kangsabanik, J.; Sø dequist, J.; Lehtomäki, J.; Heske, J.; Enkovaara, J.; Winther, K. T. s.; Dulak, M.; Melander, M. M.; Ovesen, M.; Louhivuori, M.; Walter, M.; Gjerding, M.; Lopez-Acevedo, O.; Erhart, P.; Warmbier, R.; Würdemann, R.; Kaappa, S.; Latini, S.; Boland, T. M.; Bligaard, T.; Skovhus, T. r.; Susi, T.; Maxson, T.; Rossi, T.; Chen, X.; Schmerwitz, Y. L. A.; Schiø tz, J.; Olsen, T.; Jacobsen, K. W.; Thygesen, K. S. GPAW: An open Python package for electronic structure calculations. The Journal of Chemical Physics 2024, 160, 092503.
- [23]Hjorth Larsen, A.; Jørgen Mortensen, J.; Blomqvist, J.; Castelli, I. E.; Christensen, R.; Dulak, M.; Friis, J.; Groves, M. N.; Hammer, B.; Hargus, C.; Hermes, E. D.; Jennings, P. C.; Bjerre Jensen, P.; Kermode, J.; Kitchin, J. R.; Leonhard Kolsbjerg, E.; Kubal, J.; Kaasbjerg, K.; Lysgaard, S.; Bergmann Maronsson, J.; Maxson, T.; Olsen, T.; Pastewka, L.; Peterson, A.; Rostgaard, C.; Schiøtz, J.; Schütt, O.; Strange, M.; Thygesen, K. S.; Vegge, T.; Vilhelmsen, L.; Walter, M.; Zeng, Z.; Jacobsen, K. W. The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter 2017, 29, 273002.DOI
- [24]Ziesche, P.; Kurth, S.; Perdew, J. P. Density functionals from LDA to GGA. Computational Materials Science 1998, 11, 122–127.DOI
- [25]Wang, Y.; Xu, Z.; Moe, Y. N. On the performance of local density approximation in describing the adsorption of electron donating/accepting molecules on graphene. Chemical Physics 2012, 406, 78–85.pubmed.ncbi.nlm.nih.gov/23180902
- [26]Frenkel, D.; Smit, B. Understanding molecular simulation, 2nd ed.; Computational science series 1; Acad. Press: San Diego [u.a.], 2011; Hier auch später erschienene, unveränderte Nachdrucke.
- [27]Melchionna, S.; Ciccotti, G.; Lee Holian, B. Hoover NPT dynamics for systems varying in shape and size. Molecular Physics 1993, 78, 533–544.DOI
- [28]Melchionna, S. Constrained systems and statistical distribution. Physical Review E 2000, 61, 6165–6170.pubmed.ncbi.nlm.nih.gov/11088290
- [29]Zhang, L.; Han, J.; Wang, H.; Saidi, W.; Car, R.; E, W. End-to-end Symmetry Preserving Inter- atomic Potential Energy Model for Finite and Extended Systems. Advances in Neural Information Processing Systems. 2018.
- [30]Leibe, B. In Computer Vision - ECCV 2016; Matas, J., Sebe, N., Welling, M., Eds.; Lecture Notes in Computer Science Ser. v.9908; Springer International Publishing AG: Cham, 2016; Description based on publisher supplied metadata and other sources.
- [31]Chollet, F. Deep Learning with Python, Second Edition; Manning Publications Co. LLC: New York, 2021; Description based upon print version of record.
- [32]Surblys, D.; Matsubara, H.; Kikugawa, G.; Ohara, T. Application of atomic stress to compute heat flux via molecular dynamics for systems with many-body interactions. Physical Review E 2019, 99, 051301.
- [33]Surblys, D.; Matsubara, H.; Kikugawa, G.; Ohara, T. Methodology and meaning of computing heat flux via atomic stress in systems with constraint dynamics. Journal of Applied Physics 2021, 130.DOI
- [34]Togo, A.; Chaput, L.; Tadano, T.; Tanaka, I. Implementation strategies in phonopy and phono3py. J. Phys. Condens. Matter 2023, 35, 353001.DOI
- [35]DICKEY, J. M.; PASKIN, A. Computer Simulation of the Lattice Dynamics of Solids. Physical Review 1969, 188, 1407–1418.DOI
- [36]Thomas, M.; Brehm, M.; Fligg, R.; Vöhringer, P.; Kirchner, B. Computing vibrational spectra from ab initio molecular dynamics. Physical Chemistry Chemical Physics 2013, 15, 6608.DOI
- [37]Bosse, J.; Götze, W.; Zippelius, A. Velocity-autocorrelation spectrum of simple classical liquids. Physical Review A 1978, 18, 1214–1221.DOI
- [38]Ni, Z. H.; Yu, T.; Lu, Y. H.; Wang, Y. Y.; Feng, Y. P.; Shen, Z. X. Uniaxial Strain on Graphene: Raman Spectroscopy Study and Band-Gap Opening. ACS Nano 2009, 3, 483–483.
- [39]Dong, J.; Liu, S.; Fu, Y.; Wang, Q. Investigation of strain-induced modulation on electronic properties of graphene field effect transistor. Physics Letters A 2017, 381, 292–297.DOI
- [40]Das, A.; Chakraborty, B.; Sood, A. K. Raman spectroscopy of graphene on different substrates and influence of defects. Bulletin of Materials Science 2008, 31, 579–584.arxiv.org/abs/0710.4160
- [41]Harris, C. R.; Millman, K. J.; van der Walt, S. J.; Gommers, R.; Virtanen, P.; Cournapeau, D.; Wieser, E.; Taylor, J.; Berg, S.; Smith, N. J.; Kern, R.; Picus, M.; Hoyer, S.; van Kerkwijk, M. H.; Brett, M.; Haldane, A.; del Río, J. F.; Wiebe, M.; Peterson, P.; Gérard-Marchant, P.; Sheppard, K.; Reddy, T.; Weckesser, W.; Abbasi, H.; Gohlke, C.; Oliphant, T. E. Array programming with NumPy. Nature 2020, 585, 357–362.
- [42]Zhao, H.; Aluru, N. R. Temperature and strain-rate dependent fracture strength of graphene. Journal of Applied Physics 2010, 108.DOI
- [43]Bryce, M. R. A review of functional linear carbon chains (oligoynes, polyynes, cumulenes) and their applications as molecular wires in molecular electronics and optoelectronics. Journal of Materials Chemistry C 2021, 9, 10524–10546.DOI
- [44]Zhang, K.; Zhang, Y.; Shi, L. A review of linear carbon chains. Chinese Chemical Letters 2020, 31, 1746–1756.DOI
- [45]Kotrechko, S.; Mikhailovskij, I.; Mazilova, T.; Sadanov, E.; Timoshevskii, A.; Stetsenko, N.; Matviychuk, Y. Mechanical properties of carbyne: experiment and simulations. Nanoscale Research Letters 2015, 10.DOI
- [46]Wang, Y.; Ning, X.-J.; Lin, Z.-Z.; Li, P.; Zhuang, J. Preparation of long monatomic carbon chains: Molecular dynamics studies. Physical Review B 2007, 76, 165423.DOI
- [47]Ataca, C.; Ciraci, S. Perpendicular growth of carbon chains on graphene from first-principles. Physical Review B 2011, 83, 235417.DOI
- [48]Reddy, C. D.; Rajendran, S.; Liew, K. M. Equilibrium configuration and continuum elastic properties of finite sized graphene. Nanotechnology 2006, 17, 864–870.DOI
- [49]Liang, Y.; Huang, J.; Qu, J.; Huang, J.; Hui, D. Research on the auxetic behavior and mechanical properties of periodically rotating graphene nanostructures. Nanotechnology Reviews 2022, 11, 1733–1743.DOI
- [50]Fan, N.; Ren, Z.; Jing, G.; Guo, J.; Peng, B.; Jiang, H. Numerical Investigation of the Fracture Mechanism of Defective Graphene Sheets. Materials 2017, 10, 164.DOI
- [51]Liu, F.; Ming, P.; Li, J. Ab initio calculation of ideal strength and phonon instability of graphene under tension. Physical Review B 2007, 76, 064120.DOI
- [52]Gouadec, G.; Colomban, P. Raman Spectroscopy of nanomaterials: How spectra relate to disorder, particle size and mechanical properties. Progress in Crystal Growth and Characterization of Materials 2007, 53, 1–56.DOI
- [53]LU, Q.; HUANG, R. NONLINEAR MECHANICS OF SINGLE-ATOMIC-LAYER GRAPHENE SHEETS. International Journal of Applied Mechanics 2009, 01, 443–467.DOI
- [54]Yanagisawa, H.; Tanaka, T.; Ishida, Y.; Matsue, M.; Rokuta, E.; Otani, S.; Oshima, C. Analysis of phonons in graphene sheets by means of HREELS measurement and ab initio calculation. Surface and Interface Analysis 2005, 37, 133–136.DOI
- [55]Liu, X.; Zhang, G.; Zhang, Y.-W. Thermal Conduction Across Graphene Cross-Linkers. The Journal of Physical Chemistry C 2014, 118, 12541–12547.DOI
- [56]Wu, J.-B.; Lin, M.-L.; Cong, X.; Liu, H.-N.; Tan, P.-H. Raman spectroscopy of graphene-based materials and its applications in related devices. Chemical Society Reviews 2018, 47, 1822–1873.DOI
- [57]Zheng, X.; Chen, W.; Wang, G.; Yu, Y.; Qin, S.; Fang, J.; Wang, F.; Zhang, X.-A. The Raman redshift of graphene impacted by gold nanoparticles. AIP Advances 2015, 5.
- [58]Hawthorne, F.; Raulino, P. R. E.; R., P. R.; Woellner, C. F. NOMAD Repository Dataset: Graphene MLIP dataset. Dataset, NOMAD Laboratory, 2025.DOI
- [59]Hawthorne, F.; Raulino, P. R. E.; R., P. R.; Woellner, C. F. GrapheneMLIP. https://github.com/felipehgc/GrapheneMLIP, 2025.