Abstract: Increasingly, it is recognized that new automated forms of analysis are required to understand the high-dimensional output obtained from atomistic simulations. Recently, we introduced a new dimensionality reduction algorithm, sketch-map, that was designed specifically to work with data from molecular dynamics trajectories. In what follows, we provide more details on how this algorithm works and on how to set its parameters. We also test it on two well-studied Lennard-Jones clusters and show that the coordinates we extract using this algorithm are extremely robust. In particular, we demonstrate that the coordinates constructed for one particular Lennard-Jones cluster can be used to describe the configurations adopted by a second, different cluster and even to tell apart different phases of bulk Lennard-Jonesium.
The Challenge of High-Dimensional Molecular Dynamics
Molecular Dynamics (MD) simulations generate vast amounts of data, consisting of thousands of high-dimensional snapshots of atomic positions. To extract meaningful physical insights, researchers often employ dimensionality reduction techniques to project this data into 2D or 3D "free energy surfaces." However, standard tools frequently fall short when applied to complex molecular systems.
Principal Component Analysis (PCA), for instance, assumes that data lies on a flat plane, a condition rarely met in the rugged landscapes of molecular configurations. Manifold learning techniques like Isomap assume uniform data sampling, but MD data is inherently "clumpy." Systems spend the majority of their time vibrating within stable basins (noise) and only rarely traverse the transition states that are most critical for understanding chemical processes.
As illustrated in the figure above, sketch-map is designed to ignore irrelevant thermal fluctuations within these basins while accurately capturing the global relationships between different structural states.
The Sketch-Map Algorithm
Sketch-map builds upon the principles of Multidimensional Scaling (MDS) but introduces a novel filtering approach to handle the specific challenges of MD data. The core of the algorithm relies on two primary innovations:
- Sigmoid Filters: Instead of attempting to perfectly match all high-dimensional distances ($R$) to low-dimensional distances ($r$), sketch-map utilizes sigmoid functions ($F(R)$ and $f(r)$). These functions act as a "lens" focused on a specific length scale, $\sigma$. Distances much smaller than $\sigma$ (representing thermal vibrations) are pushed toward zero, while distances much larger than $\sigma$ are pushed toward one. This forces the algorithm to prioritize the intermediate distances that define the global shape of the energy landscape.
- Landmark-Based Optimization: To manage the computational cost of large datasets, the algorithm selects a subset of "landmark" points to define the map. New data points are then projected onto this fixed map using an "out-of-sample" procedure, effectively triangulating their position based on the landmarks.
Validating with Lennard-Jones Systems
The authors tested sketch-map using "standard candles" of computational chemistry: Lennard-Jones clusters (LJ38 and LJ55) and bulk Lennard-Jonesium. LJ38 is a particularly challenging test case due to its "doubly funneled" landscape, which contains two distinct structural basins (fcc-like and icosahedral).
Sketch-map successfully achieved clear structural separation of these basins and identified several metastable, defective states that are often blurred by other methods. A key finding was the importance of the parameter $\sigma$, which defines the resolution of the map. By setting $\sigma$ larger than typical thermal fluctuations, the map automatically filters out high-frequency noise without requiring manual data averaging.
Transferability and Phase Discrimination
A standout feature of sketch-map is its transferability. The study demonstrated that a map trained on a small 38-atom cluster could successfully describe the behavior of a 55-atom cluster or even different phases of bulk material. This was made possible by using histograms of coordination numbers as high-dimensional inputs, creating a representation invariant to atom labeling.
Quantitatively, the projection quality remained high even in extreme cases. When projecting bulk phases onto a map trained on small clusters, the residual stress (measured by a $\chi^2$ function) often remained below 10%. The algorithm successfully placed bulk liquid, bcc, and $I\bar{4}3d$ phases in distinct, logical regions of the map, far removed from the original cluster data.
Limitations and Future Outlook
Despite its power, sketch-map's effectiveness depends on the choice of high-dimensional descriptors. For example, the coordination numbers used in this study could not distinguish between hcp and fcc bulk phases because their first two coordination shells are identical. Additionally, while the paper provides rules of thumb for parameter tuning, the process still requires physical intuition and some trial and error.
Looking forward, these sketch-map coordinates could serve as "collective variables" for enhanced sampling techniques like Metadynamics. This would enable researchers to actively drive simulations into unexplored regions of the structural map, accelerating the discovery of new chemical configurations.
Summary updated