DescriptionMany of the macroscopic properties of materials are rooted in the details of their structure at the atomic scale. For example, the properties of real materials are often strikingly different from those predicted by assuming a perfectly crystalline state. Indeed, perhaps contrary to intuition, nano or micro-scale features such as point defects, dislocations, or grain boundaries, often dictate the performance of materials. In order to optimize desirable properties or avoid catastrophic failure, it is hence crucial to be able to perform simulation of materials with full atomistic resolution in both space and time. One of the most powerful methods to do so is Molecular Dynamics (MD), i.e., the direct integration of atomic equations of motion. MD is extremely powerful but also computationally intensive, due to the need to resolve the motion of each individual atom. Leveraging HPC resources is therefore critical and a large fraction of the computing budget of national supercomputing centers is currently spent on such calculations.
Through different examples, I will show how massively-parallel HPC platforms provide unique opportunities to access the time and length scales required to make accurate predictions of the behavior of materials. Doing so, I will pay special attention to recently-developed techniques that leverage parallelism to extend the simulation timescales that are amenable to direct MD simulations into the milliseconds, thereby approaching experimentally relevant timescales.