Multiscale Simulation Algorithms for Understanding and Steering Complex Chemical Systems

Abstract

Atomistic simulations and macroscopic continuum models often fall short in capturing the full complexity of chemical systems across disparate time and length scales. This dissertation bridges these domains through the development and application of physics-based and data-driven multiscale methods, progressing from the practical evaluation of physical materials to the pioneering of novel computational algorithms and automated laboratory architectures. Initially, a molecular dynamics and kinetic Monte Carlo workflow was deployed to investigate the fundamental trade-off between electrical conductivity and thermomechanical stability in cross-linked organic radical polymers. Addressing the temporal limitations of traditional molecular dynamics identified in applied materials research, the second project introduces a fundamentally novel computational framework to address the rare-event timescale gap. The Hybrid kinetic Monte Carlo/Molecular Dynamics (HkMCMD) algorithm was created to decouple reactive event selection from femtosecond vibrational dynamics, thereby extending accessible simulation times to the seconds scale. By incorporating dynamic reaction rate scaling to efficiently escape pseudo-steady-state kinetic traps and utilizing spatially resolved diffusion calculations, the HkMCMD framework achieves computational savings of up to four orders of magnitude compared to brute-force methods. The final project extends multiscale simulation from understanding to control. Contemporary self-driving laboratories (SDLs) are typically constrained by rigid, closed-loop action sequences that fail to replicate the dynamic, composable decision-making of human experimentalists. To overcome this, a multi-agent machine learning architecture (RL-SDL) was developed, coupling a multi-output active learning algorithm with a reinforcement learning lab manager. Evaluated within simulated analytical environments, the framework autonomously routed samples through an open-ended action space. The RL-SDL successfully learned to dynamically compose per-sample workflows, strategically prioritizing inexpensive intermediate characterizations and scrapping unpromising materials to drastically reduce the cost of optimization. Ultimately, this dissertation demonstrates that integrating physics-based simulations with data-driven machine learning models is essential for characterizing known materials, accelerating molecular simulations, and automating the future of chemical discovery.