Machine learning speeds up solid-state material simulations

Scientists are continually seeking more efficient ways to model complex solid-state materials, and new quasi-atomic methods represent a major advance in this field. Artem Chuprov from the Skolkovo Institute of Science and Technology, Egor Nuzhin and Alexei Tsukanov from the Schmidt Institute of Geophysics at the Russian Academy of Sciences, in collaboration with Nikolai Brillantov from the School of Computing and Mathematical Sciences at the University of Leicester, presented a hybrid simulation technique that combines atomic details in key regions with continuum modeling in other parts. This approach leverages optimized interaction potentials to seamlessly link the two scales and achieves computational speeds that significantly exceed fully atomistic simulations. The researchers demonstrate the accuracy and versatility of their method using both Lennard-Jones and Tersov potentials within the LAMMPS software package, paving the way for more realistic and efficient modeling of diverse phenomena in materials science.

Computationally efficient methods for modeling complex materials are developed, bridging the gap between detailed atomistic simulations and broader continuum-based approaches. This innovation promises to accelerate progress in fields that rely on understanding the behavior of materials, from engineering to nanotechnology. By streamlining simulation, engineers can now virtually design and test materials with unprecedented speed and accuracy.

This hybrid approach combines detailed atomistic simulations that model individual atoms and their interactions with continuum modeling that treats the material as a continuous entity. The innovation lies in representing large parts of the material as “quasi-atomic”, effectively grouping many atoms into a single computational unit, while preserving full atomic detail in critical regions such as interfaces and crack regions.

The parameters controlling the interactions between these quasi-atoms are automatically optimized using techniques conceptually similar to online machine learning, ensuring that the entire simulation accurately reflects the elastic properties of the material. This achievement overcomes a critical limitation in materials science, where fully atomistic simulations of large-scale systems are often computationally prohibitive.

By strategically employing both atomic and continuous descriptions, researchers can now model complex phenomena such as particle collisions, crack propagation, and material deformation with a fraction of the computational resources previously required. The speed and versatility of this method derives from its compatibility with existing molecular dynamics software, such as LAMMPS, and is enhanced by a newly developed machine learning-based optimizer.

The researchers successfully applied this hybrid method to simulate collisions between particles of different sizes, using both simple and more complex interatomic potentials, namely the Lennard-Jones and Tersov potentials, respectively. Comparing the results obtained with this new method with those of all-atom simulations demonstrated not only its accuracy but also a significant increase in computational speed.

Moreover, a direct comparison with a similar hybrid approach, the Atomic to Continuum (AtC) method, reveals significant advantages in both speed and ease of implementation. At the heart of this progress is the concept of quasi-atomicity, in which complex media are constructed from these coarse-grained units. An automated optimization procedure tunes the subatomic interactions to match the elastic modulus of the atomic system, ensuring a seamless transition between scales.

This approach enables efficient modeling of crystalline solids and opens the possibility to study phenomena such as elasticity, fracture, and indentation, where traditional methods are often inadequate. The complete model is publicly available to encourage further research and development in multiscale solid mechanics.

Hybrid atomic continuum modeling significantly speeds up solid impact simulations

Simulations reveal a new hybrid method that achieves 150 times faster computational efficiency compared to all-atom simulations while maintaining accuracy in modeling solid-state phenomena. This performance improvement stems from a unique approach that combines atomic simulation of critical regions with continuous modeling of the remaining system volume, utilizing “quasi-atomic” representations to represent the bulk of the material.

This method accurately reproduces macroscopic collision theories, such as the Hertz and Johnson-Kendall-Roberts theories, and has demonstrated its effectiveness. Specifically, particle impact simulations yielded nearly identical results in terms of impact forces and deformation patterns when compared to all-atom simulations. The core of this progress lies in an optimized interaction potential between quasi-atoms that is tailored to match the elastic properties of the composite medium.

This optimization is conceptually coupled with online machine learning techniques and enables computationally efficient determination of subatomic parameters. A parallel sampling strategy implemented within the optimization process further increases speed and enables rapid calibration of even complex systems. Representing the bulk material quasi-atomically reduces the number of particles that require fully atomic processing, leading to significant computational savings.

The validation involves modeling collisions of particles consisting of both real atoms in the contact region and pseudoatoms of various sizes that are two, four, and eight times larger than the real atoms. The simulations successfully captured the energy transfer and deformation behavior observed in all-atom simulations, confirming that the method can accurately represent material responses at multiple scales. Additionally, the implementation of the Python, optimization, and LAMMPS bridge provides a convenient and flexible framework for integrating the hybrid approach with existing molecular dynamics software, simplifying the setup and execution of multiscale simulations.

Multiscale modeling with adaptive quasi-atomic optimization and Python-LAMMPS integration

Python, optimization, and the LAMMPS bridge underpin the methodology adopted in this study, facilitating a new hybrid approach to simultaneously simulate solids at both atomic and continuous scales. Critical regions such as contact surfaces and crack regions undergo detailed atomic processing using the LAMMPS software package, a widely used molecular dynamics simulator.

The rest of the system is modeled as a complex medium consisting of “quasi-atoms,” which are coarse-grained units that effectively represent collections of atoms, significantly reducing computational complexity. The size of these quasi-atoms is not fixed, allowing for adaptive resolution across the simulated region. Central to this method is the optimization of the interaction potential between quasi-atoms, ensuring that the collective behavior of the quasi-atoms accurately reproduces the elastic properties of fully atomic materials.

This optimization process is conceptually consistent with online machine learning techniques and allows for a computationally efficient calibration procedure. A parallel sampling strategy was implemented to further accelerate this optimization, significantly increasing the speed of potential decisions. The resulting potential is directly integrated into the LAMMPS simulation, creating a seamless connection between the atomic and continuum regimes.

The system was modeled using both a simple pairwise Lennard-Jones potential to account for van der Waals forces and a more complex many-body Tersov potential to account for covalent bonds. To rigorously test the performance of this method, collisions of particles of different sizes were simulated. This hybrid approach distinguishes itself from other techniques, especially the Atomic to Continuum (AtC) technique, by offering both increased computational speed and ease of implementation. This work avoids the limitations of purely atomic or coarse-grained simulations and shows significant advantages in modeling complex phenomena that require simultaneous resolution at multiple scales.

Learn how to bridge atomic and continuum simulations with machine learning optimization

The persistent challenge of accurately modeling material behavior at multiple scales has puzzled physicists and engineers for years. Atomic modeling, which directly simulates atomic interactions, provides detailed insights but is computationally prohibitive for all but the smallest systems. Conversely, continuum mechanics provides speed but sacrifices details important to understanding phenomena at interfaces, cracks, or contact points.

This study presents a sensible solution, a hybrid approach that intelligently combines the best of both worlds. The feature of this method is not simply the combination of an atomic model and a continuous model, but the method of learning the interaction between them. By employing a machine learning-inspired optimization process, the system effectively “trains” the continuum model to mimic the behavior of atomic models, significantly reducing computational costs without sacrificing accuracy.

This is an important step beyond previous hybrid methods, which often relied on less adaptive or more cumbersome parameterization schemes. The demonstrated speed advantages are compelling and open the door to simulation of larger and more complex systems than previously possible. Extending this to more realistic many-body potentials, which are essential for modeling complex materials, will be an important test.

Moreover, although the validation against fully atomic simulations is encouraging, the performance of this method in predicting entirely new phenomena remains to be seen. In the future, this approach could be integrated with advanced machine learning techniques, allowing models to adapt in real time during simulation. The fundamental coupling principles learned have potential applications beyond materials science in a variety of fields, from fluid mechanics to biological modeling, where scale bridging is always an obstacle. The potential lies not only in faster simulations, but also in fundamentally more efficient ways to explore the complex interactions between the microscopic and macroscopic worlds.