(3aq) Data-Driven Design of Advanced Functional Materials

The success or failure of many next generation energy storage and catalysis technologies will depend on the discovery and optimization of materials with unique, and often rare, properties. Decades of experiments and electronic structure theory calculations/modeling have led to the proliferation of large datasets which can be mined by machine learning algorithms to inform new material design strategies. I will leverage my unique domain knowledge in computational modeling and machine learning to develop a flexible set of algorithms to design novel materials. Specifically, I am interested in developing algorithms in three primary directions (Figure 1):

1. Accelerate calculation of electronic structure methods/properties using graph representations
2. Efficiently identify complex, multi-dimensional, descriptors for solid state materials and catalysts
3. Develop generative machine learning models which optimize the crystal structure for a given property

Having gained experience in studying a diverse set of problems, e.g., solid state electrolytes [1,2] and near surface alloys [3], I have acquired unique domain knowledge, and will adopt these approaches for a larger range of applications. For instance, electrolytes, photovoltaics, thermoelectrics, catalytic surfaces, and small molecule additives for electrochemical systems are suitable and impactful targets for data-driven accelerated design.

Research Background

One class of advanced materials I have studied so far is solid state electrolytes (SSEs), which improve the safety and energy density of lithium-ion batteries by replacing the combustible organic liquid electrolyte with a ceramic. However, no materials exhibit the necessary superionic conductivity and voltage stability needed for commercial viability. During my PhD I developed machine learning algorithms to accelerate the force/energy calculations in molecular dynamics, and efficiently predict solid state electrolyte properties.

I developed an artificial neural network to efficiently run molecular dynamics (MD) for systems that require a large number of simulated atoms while maintaining high accuracy.[1] Simulating these experimentally relevant low concentration requires > 150 atoms, which is a system size that exceeds the feasibility limit of traditional ab initio MD simulations. Using our approach, we calculated an optimal Cl dopant concentration of 1.3% in a Li-Ge-P-S system maximizes ionic conductivity by balancing the competing effects of effective ionic radius and dielectric screening. This approach can similarly enable simulations of other phenomena at larger time and special scales.

To directly investigate the role of crystal structure and atomic composition in voltage stability and ionic conductivity, we applied partial least squares (PLS) machine learning algorithms using the valence electronic density in both real and Fourier space.[2] The PLS model successfully identified the physically relevant BCC anion substructure and channels as effective descriptors, and was used to identify five new promising SSE candidates. We also investigated two new unsupervised learning approaches: variational autoencoders and generative adversarial networks. These generative models allow for greater flexibility and automation in the design and optimization of a crystal structure. The innovative and multi-faceted computational framework for the design of SSEs is transformative and can be easily extended to the design of advanced functional materials with diverse applications.

Teaching Interests

Even as a graduate student, I have studied pedagogy in a Future Faculty Program class and taught several guest lectures in undergraduate numerical methods and graduate catalysis classes. I would also be interested in teaching courses in thermodynamics, solid state chemistry, and quantum chemistry. Due to my specific work in machine learning, I would like to start an undergraduate/graduate course in advanced data analysis and modeling. The ultimate goal would be to give the students a set of general machine learning tools to apply to data generated in their own areas of research and inspire independent creativity and analysis. The skills taught would also be highly marketable and useful for students pursuing careers in industry. I have already developed a series of recorded lectures used by my research groups as reference material on machine learning and a teaching tool for new graduate students.

Selected Publications:

[1] K. Rao, Y. Yao, L. C. Grabow, "Accelerated Modeling of Lithium Diffusion in Solid State Electrolytes using Artificial Neural Networks", Advanced Theory and Simulation, Accepted

[2] K. Rao, Y. Yao, M. Nikolaou, L. C. Grabow, “Machine Learning the Fundamental Tradeoffs between Conductivity and Voltage Stability in Solid State Electrolytes”, In Preparation

[3] K. Rao, Q. K. Do, K. Pham, D. Maiti, L. C. Grabow, "Extendable Machine Learning Model for the Stability of Single Atom Alloys", Topics in Catalysis, 2020