(334at) Combining Molecular Simulations and Machine Learning for Nanomaterial and Reaction Design

A significant challenge in chemical design is the large number of parameters that can be manipulated. Identifying an optimal set of parameters empirically by trial-and-error approaches is cost-prohibitive, inspiring the development of computational approaches, such as molecular simulations, to help garner insight into how these parameters relate to system properties. My research is focused on combining classical molecular dynamics (MD) simulations and machine learning techniques to systematically tune chemical properties for two relevant application areas. In both applications, I develop accurate predictive models that correlate simulation-derived descriptors to experimental data measured by collaborators, enabling the high-throughput screening of parameters using computationally efficient methods.

Application 1 is motivated by the ability of gold nanoparticles (GNP) to target biomolecules, such as receptors on cancer cells, by attaching target-specific molecules (i.e. ligands) on the surface of the particle. Due to the vast number of tuning parameters of the GNP (e.g. core size, shape, and ligand selection), it is unfeasible to experimentally screen GNPs. Therefore, I developed a generalized workflow for developing GNP systems using MD simulations for any arbitrary gold core shape and size, and ligand selection. I used this workflow in conjunction with unsupervised machine learning clustering algorithms to characterize GNP surface properties. These properties dictate GNP adsorption onto lipid bilayers, which I further interrogate using enhanced sampling techniques (e.g. umbrella sampling). These tools could be used to systematically narrow down optimal GNP parameters for drug delivery and cell therapy applications.

Application 2 focuses on converting biomass (renewable, organic material from plants) into transportation fuels or commodity chemicals. Biomass conversion is performed through acid-catalyzed reactions in aqueous solutions that are hindered by low reactivity. One way to improve reactivity is to modify the solvent composition by mixing water with organic cosolvents. Thus, I developed MD simulations to understand and predict the effects of adding an organic solvent on biomass conversion reactivity. I coupled MD simulations with convolutional neural networks, a machine learning model used to classify images, and generalized the reaction rate predictions across 84 reactant-solvent combinations; enabling fast solvent screening for enhanced reaction rates.

My research interests include developing new materials or chemical processes using computational tools, such as molecular simulations (e.g. molecular docking, multiscale modeling) in conjunction with machine learning techniques (e.g. clustering, neural networks). By joining industry, I aim to broaden my ability to generate efficient and accurate computational models that could accelerate the discovery of new, beneficial products.