Nanocluster Structure and Reactivity: Experiment & Theory (including Machine Learning)

Cluster Experiments

My group is interested in the static and dynamic properties of isolated, gas phase nanoclusters. With regard to static properties, we tend to focus on the geometric and electronic structures of nanoclusters, as well as on cluster binding energies and dissociation thresholds. This typically involves studying stable clusters that we can isolate and hold in an ion trap. For example, the mass spectrum plotted below shows the distribution of clusters associated with varying numbers of all-cis hexafluorocyclohexane, C6H6F6, bound to the B12F122- dianion.

B12F12 clustered with hexafluorocyclohexane

Following mass-selection and trapping, we can conduct many different experiments to examined the properties of the isolated clusters. The image below, shows the IRMPD spectra for the bare B12F122- dianion (panel A) and clusters containing 1, 2, and 3 C6H6F6 molecules. Notice the increasing intensity of the C-F vibrational bands as we add more C6H6F6 to the cluster. For additional details, see this link.

IRMPD Spectra

When interested in the dynamic properties of clusters, we employ a technique known as differential mobility spectrometry (DMS). In a nutshell, we use the quickly oscillating asymmetric electric fields of DMS to drive ca. 20,000 cyles ion-solvent cluster formation and dissociation over the course of a few milliseconds, while monitoring motion of the ions in a high-pressure environment. Species that bind more strongly with the solvent vapour exhibit large displacements from their solvent- and field-free trajectories. We observe this as a DC compensation voltage (CV) that is required to correct ion trajectories as a function of the strength of the separation field (i.e., separation voltage; SV). This behaviour is summarized in the figure that's shown below. For additional details, see our "Hitchhiker's Guide to Dynamic Ion-Solvent Clustering".

Dispersion Curves

Chemical Theory

It is usually the case that the results of our experiments cannot be easily interpreted without guidance from chemical theory. Consequently, students in the Hopkins research group receive extensive training in computational chemistry. Usually, we conduct our computational work using commercially available software packages like Gaussian16 or ORCA. Using these programs, we predict optimal geometries and molecular properties, which we then compare with experimental outcomes. For example, below is the computed interconversion pathway between the cubic and bi-capped octahedron isomers of Li8.

Li8 Isomerization

We have also developed our own theoretical approaches to solving some chemical problems. For example, as molecules and clusters grow in size, the complexity of their potential energy landscape increases substantially. To model the species that we study experimentally, it is necessary for us to first identify the likely geometries of the samples that we probe. To do this, we have written a custom molecular simulation code that is based on the Basin-Hopping algorithm, but augmented with some techniques from unsupervised machine learning (ML). Once we have identified the low energy isomers/conformers for our system of interest, we can then treat these species with higher levels of electronic structure theory to refine molecular energies and identify the transition states that connect the minima. We can then create a coarse-grained map of the potential energy surface known as a disconnectivity graph (shown below) to learn about, e.g., isomerization pathways and likely regions of kinetic trapping.

Disconnectivity Graph

Over the past few years, we have also put some work into introducing ML to our data processing workflows. We have developed methods for using ML to classify molecules and assign molecular spectra and for predicting molecular properties. In fact, our work in the ML space has resulting in the creation of a spin-off company, WaterMine Innovation, which uses machine learning to screen new drug candidates and predict their pharmacokinetic properties.

Neural Network