On-the-fly clustering for exascale molecular dynamics simulations.

The algorithm begins with a projection step, transitioning from atomic data to a cell-based representation. Following this, we will perform connected component labeling and analysis. This process involves moving from atom-based quantities to binarized data and then to aggregation.

The projection steps are designed to enhance the distribution of the quantities. Key parameters for the projection include the atomic projection splat, which serves as a smoothing parameter for the fields, and the size of the simulation subcell, which dictates our discretization approach.

After projecting the atomic quantities—such as density, velocity, or any mechanical field computed per atom—we obtain a visual representation of these fields. At this stage, we will establish a generalization threshold. For instance, when considering density, we have various threshold options. A stringent criterion may require selected voxels to be entirely void, while a more lenient approach allows for some flexibility. The choice of threshold will influence the detection of void quantities, particularly in circular structures.

Once the projection is complete, we will have a visualized image annotated with unique ID labels for each cell, ranging from 0 to 5. These IDs correspond to the cells' positions within the global simulation, providing a convenient reference that will be utilized in subsequent analyses.