Beyond Ideal Crystals: The Case For Scale In Atomistic Modeling

Almost all computer simulations face the same trade-off: larger models can be more realistic and therefore more useful, but they also take longer to run. Engineers and scientists are therefore faced with an almost daily challenge of choosing a model that is detailed enough to capture the important details without making the calculation impractically expensive.

“All models are wrong, but some are useful.”
–George E. P. Box

Atomistic simulations are no exception. A simple crystal can often be modeled with only a few atoms repeated periodically, which is enough to predict many electronic, optical, magnetic, and thermal properties. But real materials are rarely perfect – defects, interfaces, grain boundaries, composition variations, and temperature effects often matter a great deal in practice. If these effects are left out, the simulation may miss the behavior that determines performance, but they cannot be captured with the most precise and predictive techniques. And if reducing the model size is not an option, the only alternative is to simplify the model complexity, for instance by simulating atoms as classical particles rather than quantum-mechanically, which limits accuracy and transferability.

In the past, such simplifications were necessary because of limited computing power. Today, faster hardware and better methods make it possible to perform larger, more realistic simulations that include imperfections and finite-temperature effects, while retaining important physics.

Basics of first-principles atomistic modeling

A central method in first-principles atomistic modeling is density functional theory (DFT). DFT can predict the properties of molecules, crystals, and more complex materials without relying on experimental fitting. A widely used version is based on plane waves, which works especially well for perfectly periodic crystals.

Plane-wave DFT has been highly successful in simulating a wide variety of material properties, but it becomes less efficient when the material is not perfectly periodic. Defects, surfaces, amorphous materials, random alloys, and low-concentration impurities all require large simulation cells, which quickly increases computational cost. In practice, this makes very large or highly disordered systems difficult to treat.

Temperature adds another challenge. Standard DFT is essentially a zero-temperature approach, so thermal effects must be represented indirectly through atomic motion or vibrational snapshots. Because long-wavelength vibrations require large models, cramming the model into a small simulation cell can distort finite-temperature behavior.

For more complex systems, DFT based on a linear combination of atomic orbitals (LCAO) is therefore often a better fit. Because the basis is tied directly to atoms, it handles defects, interfaces, surfaces, and vacuum regions more naturally than plane waves. It also uses fewer basis functions per atom, making larger calculations practical.

Even so, standard DFT still becomes expensive as system size grows, with runtime and memory increasing steeply. The basis set efficiency of LCAO helps, but it does not remove the underlying scaling problem. This is why recent progress has focused on both faster hardware and new algorithms.

Cutting-edge techniques to the rescue for large-scale modeling

The advent of GPUs for high-performance computing has been a major advance for atomistic modeling. Because much of the work in DFT involves large matrix operations, GPUs can often accelerate calculations dramatically compared with CPUs. Systems that once took days or weeks may now run in a few hours, making many simulations far more practical.

But the main scaling law is still there, so faster hardware alone does not solve the problem of very large models. To go further, researchers need methods that avoid the most expensive part of standard DFT. Some linear-scaling approaches exist, but they are best suited to predict global quantities such as total energy, and typically do not provide the local, chemically detailed information engineers need.

A more promising alternative is to use machine learning (ML) to predict parts of the DFT solution directly from the atomic structure, using a physics-informed graph neural network inference model. These models can reduce compute time and change the scaling from cubic in standard DFT to linear. This could extend simulations to much larger systems while preserving access to useful local properties, especially when studying many similar systems, such as those with added thermal vibrations or other small distortions. For some properties, it may even extend DFT calculations to millions of atoms! Such ML-DFT approaches are still in early stages, but industrial platforms such as QuantumATK are beginning to make it more practical, especially on GPUs.

Even larger models, and as a function of time

So far, the discussion has focused mainly on static properties. In many problems, however, the important behavior is dynamic: atoms diffuse, structures respond to temperature and pressure, and properties evolve over time. Although DFT can, in principle, be used for molecular dynamics, it is usually too expensive except for very small structures.

To study larger systems over longer times, interatomic potentials for force fields have been a staple in the atomistic modeling engineer’s toolbox. In this picture, atoms interact through an effective potential, while DFT is used “behind the scenes” to help build and validate the force field. Traditional force fields were often limited by fixed functional forms and could be hard to generalize, but modern machine-learned interatomic potentials (MLIPs) are now changing that by offering much greater flexibility and accuracy across a wide range of materials. With active-learning workflows, they can approach first-principles quality at far lower cost, and have expanded the use of force fields into the field of materials discovery. MLIPs also take great advantage of GPUs, both for the ML algorithms that constitute the training phase, and in the inference model employed when the potential is used to predict the energies, forces and stress needed to drive molecular dynamics and other simulations.

With force fields, simulations can reach systems with millions of atoms, making it possible to study disordered materials, polymers, and biological systems. Unlike DFT, their main limit is often not size but time: molecular dynamics must still move sequentially, time step by time step, so truly long timescales remain hard to access. Specialized acceleration methods can help extend those timescales by a few orders of magnitude, and improved software workflows are making these techniques easier to use in practice, but we are still typically limited to simulation times far shorter than those we experience in the real world, and at which scale materials processing may occur.

Hierarchical and multiscale modeling

So, even with all the advances outlined above, atomistic modeling cannot cover every relevant length and time scale directly. Many real materials problems involve microstructures, crack growth, diffusion, or processing histories that unfold on much larger scales. That is where multiscale modeling comes in. In these workflows, lower-level methods such as DFT or force-field simulations provide parameters and physical insight for mesoscale or continuum models. In other cases, different resolutions are coupled within the same simulation so that only the most important region is treated in full detail. This allows atomistic modeling to remain predictive while still connecting to engineering-scale questions.

What appears to be a simple piece of material often hides a detailed microstructure—grain boundaries, disordered alloys, and underlying crystal structures. Capturing all the physics that determine performance requires a hierarchy of modeling techniques for each level of detail.

The recent acquisition of Ansys by Synopsys provides a practical context for linking atomistic methods in QuantumATK with continuum tools such as Ansys Mechanical, Fluent, Rocky, and ChemKin Pro. In this setting, workflow continuity is important: when tools interoperate within a consistent software environment, it becomes easier to build, validate, support, and maintain cross-scale simulation pipelines.

Generating and analyzing large-scale models

Large-scale simulation is not only about solving equations efficiently. It also depends on being able to build realistic complex models, such as disordered structures, interfaces, polymers, or thermally displaced configurations. All the extra computational scale does not help us unless we can also prepare input structures for the simulations that accurately represent the materials that materials scientists are tasked with investigating,

Visualization of results is equally important. As models become larger and more complex, researchers need tools that help interpret the output data, visualize motion and spatial variation, and combine results of different calculations into physical observables. The goal is to turn modeling output into actionable insight and support collaboration across teams. Environments such as Synopsys QuantumATK’s NanoLab aim to support this by combining model building, job submission, visualization, and analysis in one interface.

Conclusion

We may now return to the original question: When is a larger model actually worth the additional cost? The deciding factor is not simply whether a larger model can be run, but whether it reveals physics that smaller models miss. For realistic materials, the answer is often yes: defects, disorder, temperature, interfaces, and time-dependent behavior all require more scale than idealized textbook systems. That is why large-scale atomistic modeling matters, and why advances in methods, hardware, and workflows remain important for modern materials science. More information on how these approaches are applied in practice is available through Synopsys QuantumATK.