Machine Learning Speeds up Solid Material Simulations

Scientists are continually seeking more efficient methods to model complex solid materials, and a new quasi-atom method offers a significant advancement in this field. Artem Chuprov from the Skolkovo Institute of Science and Technology, Egor Nuzhin and Alexey Tsukanov from the Schmidt Institute of Physics of the Earth of the Russian Academy of Sciences, working in collaboration with Nikolay Brilliantov from the School of Computing and Mathematical Sciences, University of Leicester, present a hybrid simulation technique that combines atomistic detail in critical regions with continuum modelling elsewhere. This approach utilises optimised interaction potentials to seamlessly link the two scales, achieving a computational speed substantially exceeding that of full atomistic simulations. The researchers demonstrate the accuracy and versatility of their method using both Lennard-Jones and Tersoff potentials within the LAMMPS software package, paving the way for more realistic and efficient modelling of diverse phenomena in materials science.

A computationally efficient technique to model complex materials has been developed, bridging the gap between detailed atomic simulations and broader, continuum-based approaches. This innovation promises to accelerate progress in fields reliant on understanding material behaviour, from engineering to nanotechnology. By streamlining simulations, engineers can now design and test materials virtually with unprecedented speed and accuracy.

This hybrid approach combines detailed atomistic simulations, modelling individual atoms and their interactions, with continuum modelling, which treats materials as continuous entities. The innovation lies in representing large sections of a material as “quasi-atoms”, effectively grouping many atoms into single computational units, while maintaining full atomistic detail in critical regions such as contact surfaces or crack areas.

Parameters governing the interactions between these quasi-atoms are automatically optimised using a technique conceptually similar to online machine learning, ensuring the overall simulation accurately reflects the material’s elastic properties. This work overcomes a significant limitation in materials science, where fully atomistic simulations of large-scale systems are often computationally prohibitive.

By strategically employing both atomistic and continuum descriptions, researchers can now model complex phenomena, like particle collisions, crack propagation, and material deformation, with a fraction of the computational resources previously required. The method’s speed and versatility stem from its compatibility with existing molecular dynamics software, such as LAMMPS, augmented by a newly developed machine learning-based optimiser.

The researchers successfully applied this hybrid method to simulate collisions between particles of varying sizes, using both simple and more complex interatomic potentials, the Lennard-Jones potential and the Tersoff potential, respectively. Comparing the results obtained with this new method to those from full-atomistic simulations, they demonstrated not only its accuracy but also a substantial improvement in computational speed.

Furthermore, a direct comparison with the AtC (Atomic to Continuum) method, a similar hybrid approach, revealed a significant advantage in both speed and ease of implementation. The core of this advancement is the quasi-atom concept, where a composite medium is built up from these coarse-grained units. An automated optimisation procedure calibrates the interactions between quasi-atoms to match the elastic moduli of the atomic system, ensuring a seamless transition between scales.

This approach allows for the efficient modelling of crystalline solids, opening possibilities for studying phenomena like elasticity, fracture, and indentation, where traditional methods often fall short. The complete model is publicly available, facilitating further research and development in multiscale solid mechanics.

Hybrid atomistic-continuum modelling delivers substantial speedup in solid-state impact simulations

Simulations reveal a novel hybrid method achieving a 150x speedup in computational efficiency compared to full-atomic simulations while maintaining accuracy in modelling solid-state phenomena. This performance gain stems from a unique approach combining atomistic simulation of critical regions with continuum modelling of the remaining system volume, utilising “quasi-atoms” to represent larger portions of the material.

The method accurately reproduces macroscopic collision theories, such as those of Hertz and Johnson-Kendall-Roberts, demonstrating its validity. Specifically, simulations of colliding particles yielded nearly identical results for impact forces and deformation patterns when compared to full-atomic simulations. The core of this advancement lies in an optimised interaction potential between quasi-atoms, calibrated to match the elastic properties of the composite medium.

This optimisation, conceptually aligned with online machine learning techniques, allows for a computationally efficient determination of the quasi-atom parameters. A parallel sampling strategy implemented within the optimisation process further enhances speed, enabling rapid calibration even for complex systems. By representing the bulk material with quasi-atoms, the number of particles requiring full atomistic treatment is drastically reduced, leading to significant computational savings.

Validation involved modelling collisions of particles comprised of both real atoms in contact regions and varying sizes of quasi-atoms, two, four, and eight times larger than the real atoms. The simulations successfully captured the energy transfer and deformation behaviour observed in full-atomic simulations, confirming the method’s ability to accurately represent material response at multiple scales. Furthermore, the implementation, a Python, Optimisation, LAMMPS bridge, offers a convenient and flexible framework for integrating the hybrid approach with existing molecular dynamics software, simplifying the setup and execution of multiscale simulations.

Multiscale modelling via adaptive quasi-atomic optimisation and Python-LAMMPS integration

A Python, Optimisation, LAMMPS bridge underpins the methodology employed in this work, facilitating a novel hybrid approach to simultaneously simulate solids at both atomistic and continuum scales. Critical regions, such as contact surfaces and crack areas, receive detailed atomistic treatment using the LAMMPS software package, a widely used molecular dynamics simulator.

The remainder of the system is modelled as a composite medium comprised of “quasi-atoms”, effectively coarse-grained units representing collections of atoms, to substantially reduce computational demands. These quasi-atoms are not fixed in size, allowing for adaptable resolution across the simulated domain. Central to this method is the optimisation of interaction potentials between quasi-atoms, ensuring their collective behaviour accurately replicates the elastic properties of the fully atomistic material.

This optimisation process conceptually aligns with online Machine Learning techniques, enabling a computationally efficient calibration procedure. A parallel sampling strategy was implemented to further accelerate this optimisation, significantly improving the speed of potential determination. The resulting potentials are then integrated directly into the LAMMPS simulations, creating a seamless connection between the atomistic and continuum regions.

Systems were modelled using both the simple, pairwise Lennard-Jones potential, describing van der Waals forces, and the more complex, multi-body Tersoff potential, which accounts for covalent bonding. Particle collisions of differing sizes were simulated to rigorously test the method’s performance. This hybrid approach distinguishes itself from other techniques, notably the Atomic to Continuum (AtC) method, by offering both increased computational speed and greater ease of implementation. The work demonstrates a significant advantage in modelling complex phenomena requiring simultaneous resolution at multiple scales, bypassing the limitations of purely atomistic or coarse-grained simulations.

Learning to bridge atomistic and continuum simulations with machine learning optimisation

The persistent challenge of accurately modelling material behaviour at multiple scales has long frustrated physicists and engineers. Simulating the interactions of atoms directly, atomistic modelling, provides detailed insight but is computationally prohibitive for all but the smallest systems. Conversely, continuum mechanics offers speed but sacrifices the fine-grained detail crucial for understanding phenomena at interfaces, cracks, or points of contact.

This work presents a clever solution, a hybrid approach that intelligently combines the strengths of both worlds. What distinguishes this method is not simply the coupling of atomistic and continuum models, but the way it learns the interaction between them. By employing a machine learning-inspired optimisation process, the system effectively ‘trains’ the continuum model to mimic the behaviour of the atomistic one, dramatically reducing computational cost without sacrificing accuracy.

This is a significant step beyond previous hybrid methods, which often rely on less adaptable or more cumbersome parameterisation schemes. The demonstrated speed advantage is compelling, opening doors to simulating larger and more complex systems than previously feasible. Extending this to more realistic, multi-body potentials, essential for modelling complex materials, will be a crucial test.

Furthermore, while the validation against full atomistic simulations is reassuring, the method’s performance in predicting entirely new phenomena remains to be seen. Looking ahead, this approach could be integrated with advanced machine learning techniques, potentially allowing the model to adapt in real-time during simulations. Beyond materials science, the underlying principle of learned coupling could find applications in diverse fields, from fluid dynamics to biological modelling, where bridging scales is a persistent hurdle. The promise lies not just in faster simulations, but in a fundamentally more efficient way to explore the complex interplay between the microscopic and macroscopic worlds.