(334bv) Multi-Scale Modeling of Heterogeneous Catalysts Using First-Principles and Deep Learning

Although computational models have been instrumental in advancing the areas of renewable energy production and storage, traditional catalyst design primarily relies noticeably on human intuition. As such, the usual timeline for discovery, development and deployment of novel catalytic materials and chemical process is typically long-term and capital intensive. A promising solution to address this shortcoming is integration of machine learning (ML) and artificial intelligence (AI) methods, which are known to improve iteratively through experience and more data, with human intuition and creativity. I am fascinated by this data-driven design paradigm possible through the digitization of traditional R&D pipeline using AI/ML methods; enabling teams to undertake diagnostic, predictive, and prescriptive analytics which form an integral part of Industry 4.0.

My Ph.D. research at Purdue University is focused on developing multi-scale catalyst models that capture the essential structure-functional properties of the real-world catalyst. These models include a combination of atomic-scale quantum mechanical simulations with subsequent reactor design and high throughput machine-learning based strategies to predict material properties. I am using these models to investigate the active site and reaction mechanism for water-gas shift and NO_x decomposition in vehicular exhaust. The proposed theory models, as validated through our experimental collaborators, offer an improved molecular-level mechanistic understanding of the reaction. Besides this, I am involved in developing a grand-canonical genetic algorithm to generate catalyst models which allow an improved active-site description thereby reducing the theory-experiment disparity.

Alongside developing theoretical models, I am investigating accelerated screening strategies for ranking catalyst candidates. These strategies are being developed for high-entropy alloys (HEA), which typically comprise of five or more elements. HEAs have attracted considerable interest in recent years due to their superior oxidation resistance relative to that of conventional alloys catalysts. I am using crystal-graph convolutional neural network (CGCNN) to model and rank catalyst stability with reduced computational expense compared to traditional electronic-structure optimization methods. Appropriate featurization of the chemical moieties to encode physical-chemical properties is a crucial step in developing a ML-Al model. Crystal graphs are a promising avenue to encode these structural and chemical underpinnings. In addition to generating a surrogate model, I am using the latent-space encodings generated through the CGCNN to discover lower dimensional manifolds that can be used for accelerated optimization of the target property.

Besides my graduate research, I am currently interning at Dow Chemical Company (Lake Jackson, TX) in the Chemometric and AI department. I am part of a cross-disciplinary team comprised of process system engineers, applied statisticians, and organic chemists. My primary projects are 1) developing process models for anomaly detection to improve unplanned event identification and 2) building a GPU-accelerated cloud-based deep-learning framework to screen large database (on the scale 10⁹) of organic molecules to optimize for cost, reactivity and stability.