Title: FlowerMD: A Flexible Library of Organic Workflows and Extensible Recipes for Molecular Dynamics and Machine-Learned Coarse-Grained Simulations of Isotropic and Anisotropic Systems

Program: Materials Science and Engineering MS

Advisor: Eric Jankowski

Committee: Jeunghoon Lee, Mahmood Mamivand

Abstract:

Molecular dynamics (MD) simulations are essential tools for understanding and predicting material behavior at the atomic and molecular scale. While numerous open-source software packages exist for different stages of MD simulations, assembling a seamless, end-to-end workflow for complex multi-step simulations remains a significant challenge. In this work, we develop FlowerMD, a flexible and extensible Python-based software package that automates molecular simulation workflows, improving both reproducibility and ease of use.

FlowerMD provides modular components that simplify the end-to-end execution of MD workflows, from system initialization and force field application to simulation execution. It also automates multi-step simulation workflows through Recipes—predefined, modular workflows that integrate complex processes into a cohesive pipeline. We demonstrate this functionality by developing Recipes for fusion welding, tensile testing, and surface wetting, showcasing FlowerMD's ability to automate intricate molecular simulations. FlowerMD Recipes are transferable across different chemistries, simulation resolutions (atomistic vs. coarse-grained), physical models (isotropic vs. anisotropic), and force fields (traditional vs. ML-driven). A key advantage of FlowerMD is its user-friendly design, allowing researchers without extensive programming expertise to efficiently set up simulations. With a high-level API and predefined workflows, it eliminates the need to manually integrate multiple software packages, enabling researchers to focus on scientific analysis rather than technical implementation.

Additionally, we integrate machine learning-based potentials into coarse-grained molecular dynamics (CGMD), focusing on both isotropic and anisotropic particles. For isotropic CGMD, we develop a force-matching neural network that learns Lennard-Jones (LJ) potentials to accurately reproduce interparticle forces. For anisotropic CGMD, we implement a force and torque matching model that captures orientation-dependent interactions in non-spherical particles. These capabilities enable FlowerMD to bridge the gap between traditional force fields and ML-driven molecular simulations, offering a scalable and reproducible framework for complex simulation workflows.