Carlos León, Roderick Melnik
8 min
This study explores the application of machine learning interatomic potentials (MLIPs) to investigate shape memory effects in graphene oxide nanoribbons (GONRs). Shape memory materials, which can return to an original state after deformation, are increasingly vital in biomedical engineering for applications such as stents, actuators, and drug delivery systems. While density functional theory (DFT) is highly accurate, it is computationally prohibitive for larger nanoribbon systems. The authors develop Moment Tensor Potentials (MTP) to bridge this gap, providing a data-driven approach to simulate phase transitions in highly ordered graphene oxide nanoribbons.
To develop the MTP, the researchers trained models using static DFT calculations of energy, forces, and stress. The training process utilized an iterative active learning loop where the potential was refined by identifying configurations where the model's predictions deviated significantly from DFT results. This allowed the authors to study the lattice contraction and expansion phases—critical to the shape memory effect—across various nanoribbon widths. By incorporating physics-based descriptors that respect rotational invariance and permutation symmetry, the model maintains a robust balance between computational efficiency and predictive accuracy, allowing for the simulation of systems containing thousands of atoms.
The researchers identified that the shape memory effect in GONRs depends strongly on width. The metastable phase, which characterizes the shape memory behavior, only emerges at a critical width of approximately 51 armchair lines. For narrower nanoribbons, the confinement effect suppresses this phase, and the two-phase behavior is lost. The team also examined the impact of structural imperfections; while the shape memory effect remains robust against minor line defects and low concentrations of oxygen vacancies, it is significantly degraded by high concentrations of impurities like nitrogen or boron, which disrupt the electronic and geometric configuration required for the shape memory phase transition.
This work provides a framework for integrating data-driven techniques into materials science, specifically for designing next-generation smart nanomaterials. By demonstrating that MLIPs can accurately predict complex phase transformations in graphene-based structures, the authors offer a toolset that could significantly accelerate the development of biocompatible, low-energy diagnostic and therapeutic devices.
Shape memory materials have been playing an important role in a wide range of bioengineering applications. At the same time, recent developments of graphene-based nanostructures, such as nanoribbons, have demonstrated that, due to the unique properties of graphene, they can manifest superior electronic, thermal, mechanical, and optical characteristics ideally suited for their potential usage for the next generation of diagnostic devices, drug delivery systems, and other biomedical applications. One of the most intriguing parts of these new developments lies in the fact that certain types of such graphene nanoribbons can exhibit shape memory effects. In this paper, we apply machine learning tools to build an interatomic potential from DFT calculations for highly ordered graphene oxide nanoribbons, a material that had demonstrated shape memory effects with a recovery strain up to 14.5% for 2D layers. The graphene oxide layer can shrink to a metastable phase with lower constant lattice through the application of an electric field, and returns to the initial phase through an external mechanical force. The deformation leads to an electronic rearrangement and induces magnetization around the oxygen atoms. DFT calculations show no magnetization for sufficiently narrow nanoribbons, while the machine learning model can predict the suppression of the metastable phase for the same narrower nanoribbons. We can improve the prediction accuracy by analyzing only the evolution of the metastable phase, where no magnetization is found according to DFT calculations. The model developed here allows also us to study the evolution of the phases for wider nanoribbons, that would be computationally inaccessible through a pure DFT approach. Moreover, we extend our analysis to realistic systems that include vacancies and boron or nitrogen impurities at the oxygen atomic positions. Finally, we provide a brief overview of the current and potential applications of the materials exhibiting shape memory effects in bioengineering and biomedical fields, focusing on data-driven approaches with machine learning interatomic potentials.
Alex: And what is the single idea that lets the learned model replace some of that work?
Sam: The model estimates total energy by adding contributions from local atomic neighborhoods. Each neighborhood records nearby atoms and their species. The approach is called a moment tensor potential, and its construction respects physical symmetries, such as rotating a structure without changing its energy.
Alex: That gives the model a useful constraint before training. How do the authors choose which quantum calculations to spend effort on?
Sam: For the sheet, they use active learning: the model flags unfamiliar configurations encountered during structural relaxation. Those configurations receive new quantum calculations and enter training. Relaxation means moving atoms toward a lower-energy arrangement.
Alex: But a ribbon has edges, unlike an extended sheet. Does the same training procedure carry straight over?
Sam: No, and that’s an important change in the method. For ribbons, they train on already relaxed narrow ribbons together with the sheet. They use the learned potential only to estimate energies near those structures, not to relax the ribbons themselves.
Alex: So the ribbon model is interpolating between narrow structures and the sheet limit. What concrete test shows that this interpolation captures something beyond its training examples?
Sam: Their preferred model combines the sheet with a ribbon seven armchair lines wide. Armchair lines are atomic rows used to specify ribbon width. They then compare its predictions with quantum results for a wider ribbon excluded from training.
Alex: An excluded example is useful, but what did it actually get right? Predicting an energy offset is different from predicting another phase.
Sam: At twenty-five armchair lines, the reference energy curve starts changing shape, although it has no additional local minimum yet. The learned model reproduces that emerging feature. It also predicts the shift in the lowest-energy state as ribbons become wider.
Alex: So that check supports the beginning of the transition, not necessarily the whole prediction for much wider ribbons.
Sam: That’s the right scope. The model predicts gradual development of the additional phase as width increases. At four hundred and one armchair lines, it produces a clear additional local minimum, but that large-ribbon result is a model prediction.
Alex: The practical message seems to be that shape memory is not automatically inherited from the sheet. The ribbon has to be wide enough to support the relevant states.
Sam: The authors connect the emerging metastable phase to interactions between oxygen atoms along the same row. Those interactions strengthen as width increases. Confinement can therefore suppress the two-phase feature associated with shape memory.
Alex: Now let’s make the structure less ideal. Does that feature survive when oxygen atoms are missing?
Sam: In their large-ribbon simulations, sparse vacancies leave energies close to those of the ordered structure. At around ten percent missing oxygen atoms, the additional phase is destroyed. They generate multiple random vacancy arrangements, rather than testing just one pattern.
Alex: Missing atoms are one defect. What about losing the perfectly ordered oxygen rows while keeping the overall composition?
Sam: They test that too, using regions with different oxygen arrangements. The additional phase can survive disrupted ordering if the resulting ordered strips are wide enough to support both phases independently. So disorder does not have one universal effect.
Alex: And replacing oxygen changes the chemistry, not just the pattern. Did the substitutions behave similarly?
Sam: Nitrogen substitutions are more disruptive than boron substitutions in their calculations. The two-phase feature is already affected when nitrogen replaces a quarter of one percent of oxygen atoms. It is destroyed at two and a half percent substitution.
Alex: Those are useful sensitivities, but they rely on the potential. Where does the paper show that the model itself stops being reliable?
Sam: The clearest problem appears in the sheet calculations. Deformation rearranges electrons and produces magnetism in some configurations. The model can locate the two minima, yet mispredict energies around one phase.
Alex: That sounds like a failure of the local-neighborhood assumption. Does the document identify it that way?
Sam: It attributes the difficulty partly to missing training configurations and partly to long-range interactions from spin-polarized carbon atoms. Increasing the neighborhood size did not adequately solve the problem. Restricting training to nonmagnetic configurations improved accuracy, but required substantial additional training data.
Alex: Then I’d trust the checked width trend more than a general claim that this potential handles every phase transformation. What else remains outside the demonstrated scope?
Sam: Stability of larger structures at nonzero temperatures remains future work, along with accounting for magnetic effects. And none of these calculations demonstrates a biomedical device. The paper’s discussion of stents, actuators, and other shape-memory applications provides context, not validation of these ribbons in those settings.
Alex: Given those boundaries, who should read the full document, and where should they start?
Sam: Researchers developing learned atomic potentials or studying responsive graphene structures should read it. Start with the ribbon results: compare the training choices and the quantum check on the excluded width. Then read the sheet results for the magnetic failure, and the defect sections for the limits of robustness.
Alex: And for a researcher outside that area, what is the one conclusion worth carrying away?
Sam: A sheet’s shape memory is not a ribbon’s guarantee. Width and oxygen arrangement can decide whether it survives—and a useful learning model must still be checked where the physics changes.
Alex: That’s a useful boundary to keep in mind.