Nanoparticle Catalytic Discovery using Machine Learning Models with Chemisorption Inputs

The mechanisms guiding nanocatalysis are poorly understood and extremely hard to model. Proposed effort attempts to provide a machine learning based platform that would guide the design of nanocatalytic materials for a variety of applications, beginning with the use of case of chloro-organic contaminant remediation. The current approaches either use conventional molecular dynamics coupled with ab initio approaches to uncover reaction kinetics or machine learning platforms that neglect or under-estimate the soft epitaxy between adsorbant and surface. Using reactive descriptors sourced from MD trajectories with validated potentials, the artificial neural networks will have molecular imprint information key to catalytic pathways. This would elucidate reactive sites using descriptors that eschew the more computationally expensive DFT approaches, leading to larger phase exploration and more translatable models. The artificial neural network models would be based on reactivity descriptors that are very specific to the molecular geometry and nanoparticle feature size. Sensitivity analyses would be done to improve the initial training of data sets for a similar set of contaminants. Robustness of the model would be tested by running the models on a dissimilar system. Finally, hypothetical nanomaterials and facets would be proposed to experimental collaborators.