DISSERTATION DEFENSE
Kieran Nehil-Puleo, Interdisciplinary Materials Science
*under the supervision of Peter Cummings
“Towards virtual autonomous material design”
3.4.26 | 3:00 pm | 134 Olin Hall | Zoom
Virtual autonomous material design is a computational paradigm in which artificial intelligence, high-throughput simulations, and adaptive optimization algorithms work together to discover and refine new materials entirely in silico. In this approach, models generate candidate structures and compositions, physics-based simulations evaluate their properties, and closed-loop learning strategies use the results to iteratively guide the next set of virtual experiments. By relying solely on digital workflows, researchers can explore vast chemical and structural design spaces with exceptional speed and efficiency, screening for performance before any physical synthesis is considered. This simulation-driven autonomy not only reduces cost and accelerates innovation, but also reveals complex structure–property relationships through continuously evolving data and predictive modeling, positioning materials science as a fully computational, intelligent discovery process.
Despite rapid progress, virtual autonomous material design requires the development of new methods to represent, generate, and evaluate materials in ways that are both physically rigorous and computationally scalable. Graph theory provides a natural framework for encoding atomic structures and bonding environments as mathematically tractable objects, enabling machine-learning models to navigate complex chemical spaces while preserving connectivity, topology, and local environments. Group theory is equally essential for handling symmetry, allowing automated systems to recognize equivalent configurations, reduce redundant calculations, and construct physically valid structures that obey molecular constraints. Alongside these mathematical foundations, automated simulation construction is critical: workflows must be able to translate candidate representations into reliable computational experiments without human intervention. The development of these methods is not just a technical refinement but a prerequisite for true autonomy, ensuring that virtual design platforms can generate meaningful materials, run robust simulations at scale, and learn from the results in a closed, continuously improving loop.
In this dissertation, I will present my work towards virtual autonomous materials design in which advances in mathematical representation, machine learning, and automated simulations are developed in direct response to the requirements of real-world applications, and then leveraged to enable end-to-end computational discovery. On the methodological side, this includes symmetry-aware and graph-based learning architectures for capturing complex, interaction-dependent properties, physically grounded strategies that remove manual tuning from molecular simulation workflows, and hierarchical, group-theoretic approaches for generating chemically meaningful design spaces that are both exhaustive and computationally tractable. These developments establish the foundations for reliable, scalable, and fully automated in-silico experimentation. Built upon these capabilities, the materials design component demonstrates how multi-fidelity property prediction, active learning, and high-throughput virtual screening can be coupled directly to performance metrics at the device or process level, allowing candidate materials to be evaluated in terms of their functional impact rather than isolated properties. Together, this work illustrates a closed, co-evolutionary loop in which application demands shape the creation of new computational methods, and those methods, in turn, expand the scope, speed, and physical realism of autonomous materials discovery.