(2bf) Developing Workflows to Understand and Design Complex Alloy Catalysts Using Density Functional Theory, Machine Learning, and Catalysis First-Principles

Electrochemical devices, such as fuel cells, represent a promising technological solution to the transition from a fossil fuel-based economy to a more sustainable, circular economy with clean hydrogen as energy currency. However, a cost-effective, practical implementation of fuel cells as power sources in vehicles and other applications is hampered by acute challenges at the fundamental level, such as the sluggish kinetics of the oxygen reduction reaction (ORR) at the cathode. These fundamental problems traverse scales to ultimately manifest as economic problems—for example, Pt, which is commercially used to catalyze ORR, has a low earth-abundance and consequently higher cost, that leads to lower cost-effectiveness of hydrogen-powered cars. Therefore, my research objective entails identifying key fundamental challenges in sustainable chemical processes and technologies, such as fuel cells, and providing solutions that ultimately have an impact, not only at the fundamental level, but also at the process or economic level. Furthermore, my goal is to employ a computational and data-driven approach in collaboration with experimental groups towards solving these problems. The broad focus of my research during my PhD has been elucidating key atomic-scale insights into the stability and activity of Pt-based alloy catalysts for the ORR and further, proposing design rules for the development of more active high-entropy alloy (HEA) catalysts.

On the applications side, I have developed a rigorous thermodynamic analysis of the stability of Pt-based bimetallic alloy surfaces, in particular, the effect of various mechanisms such as segregation, leaching, and surface oxidation on formation of Pt-skin layers that ultimately lead to the improvement of the ORR activity of these alloys over Pt. Further, I extended the analysis to incorporate multiple elements to model HEAs and demonstrated how the activity of such alloys can be maximized under the constraints of the ORR volcano in collaboration with the Wang group at Johns Hopkins University [1]. However, a particular challenge with modeling disordered alloys with multiple elements was the intractability of the vast combinatorial space of catalyst structures using DFT calculations exclusively.

Therefore, on the methodological side, I worked on developing workflows that incorporated density functional theory (DFT) calculations, catalysis first principles, and deep learning for the rapid screening of catalyst structures. I demonstrated the potential to reduce the number of required DFT calculations from ~18,000 to ~2,500 to identify stable structures for four different Pt-bimetallic alloys with this workflow. A unique aspect of my work in this domain has been incorporating physical and chemical principles into iterative active learning schemes as well as trying to interpret deep learning models using dimensionality reduction techniques in collaboration with the Kevrekidis group at Johns Hopkins University [2].

Based on my previous work, I have identified three areas of interest for future research:

Reaction networks and informatics

With the advent of HEAs, a large, complex design space has been unlocked for catalysis. However, modeling the surfaces of these alloys such that the atomic-scale insights correspond well with experimental measurements is a huge challenge. While my previous work has addressed the challenge of catalyst structure determination from a large combinatorial space for HEAs, the challenge is raised by an order of magnitude when modeling reactions, especially those involving a large number of reaction intermediates, on these surfaces. It is possible for complex reaction networks with hundreds of unique binding sites and corresponding energetics to emerge due to the heterogeneity of an HEA surface. Handling such complexity through purely intuition-driven reaction network generation can then become intractable. Therefore, approaches involving automated reaction network generation and reaction informatics can be helpful in solving complex reaction networks on HEAs.

Physics-driven machine learning

In the past few years, many graph neural network models have emerged that have an excellent ability to map material structure to property. However, such models have typically offered little more than excellent prediction ability, which, while useful for screening, is not commensurate with the traditional idea of a model. The ability to interpret and extract physical insights from a model remains an important goal in the sciences. On this front, dimensionality reduction methods like principal components analysis (PCA) have shown promise in revealing the correspondence between the graph neural network models and physical properties relevant to the problem identified by intuition. Further, incorporation of scientific knowledge in deep learning schemes—such as in physics-informed loss functions—has helped in the reduction of required data. Physics-driven machine learning workflows therefore present a unique opportunity to leverage scientific intuition to make data-driven schemes more efficient and transparent.

Bridging the gap between theory and experiment

Ideally, catalyst structure-property relationships developed using DFT should lead to the design and development of more optimal catalysts in an iterative feedback loop with experimentalists. However, in reality, many of the simplifying assumptions made in theoretical models can lead to significant deviations from experimental measurements, thus making it difficult for seamless catalyst development. In the case of Pt-based HEAs, factors such as presence of Pt-skins with variable thickness and strain-induced relaxation in the surface can significantly alter the prediction for most active HEA composition. To help bridge this gap, machine learning models built within the delta learning paradigm can be deployed. This requires organizing the theoretical and experimental data into databases and careful benchmarking between the two. Machine learning models can then be trained to predict synthesis-structure-property relationships that are informed both by experimental measurements and theoretical calculations.

Teaching Interests

I am very passionate about science communication and have written and shared articles on the internet on solving chemical engineering problems—such as computational fluid dynamics (CFD)—using computers and programming [3]. I have received positive feedback from students as well as faculty on these articles, and this drives to push myself further in terms of outreach. I consider it my mission, not just a potential profession, to share what I have learned in an accessible medium with those who are interested in and passionate about a career in chemical engineering.

My teaching philosophy is to distill complex concepts into simple models that students can play around with and visualize to improve their intuition. For example, a course on thermodynamics can quickly become needlessly abstract if not tethered to a real-world application. My approach to teaching such a course would involve, for example, a discussion on the application of activity coefficient models on modeling flash vaporization of mixtures and the development of computer code that would help students visualize how the product compositions would change with the change in the models. I believe that the usage of computers and programming in teaching chemical engineering is not only beneficial, but essential.

I have been a teaching assistant for two courses in my PhD—Graduate Engineering Mathematics and Heat and Mass Transfer. I have delivered recitations as well as lectures in the latter to undergraduates. I would be interested in teaching any undergraduate course in chemical engineering, but if given a preference, I would like to teach either thermodynamics or chemical reaction engineering. I am also interested in teaching graduate courses on the application of DFT simulations to study reaction kinetics and thermodynamics as well as applications of deep learning to molecules and materials.

References:

[1] Xu, F.*, Deshmukh G.*, et al., Advanced Oxygen Reduction Electrocatalysts Based on Multi-Principal Element Alloys, in preparation. (*equal contribution)

[2] Deshmukh G., Wichrowski, N., Evangelou, N., et al., Active Learning of Ternary Alloy Structures and Energies, submitted (preprint link: https://chemrxiv.org/engage/chemrxiv/article-details/647fe2af4f8b1884b7f...)

[3] Deshmukh G., Computational Fluid Dynamics using Python: Modeling Laminar Flow (Link: https://medium.com/towards-data-science/computational-fluid-dynamics-usi...)