Machine Learning Approaches to the Simulation of Nanoparticle Crystallization

Overview & Synopsis

Nanoparticles, with sizes ranging from a few to a hundred nanometers, exhibit properties distinct from bulk matter due to their high surface-to-volume ratio and lack of long-range crystalline order. Surface atoms dominate their behavior, resulting in altered electronic structures, unusual phase stability, and unique reactivity. Combining metals into nanoalloys or forming semiconducting nanocrystals allows tailoring of optical, electronic, and chemical properties beyond what is achievable in macroscopic materials.Controlling nanoparticle formation and growth is key to designing materials with desired functionalities. Computational modeling plays a central role in understanding the underlying atomic-scale processes. First principles quantum simulations provide highly accurate predictions but are too costly to capture nucleation and early growth stages. Classical molecular dynamics enables simulations of thousands of atoms over extended timescales but relies on simplified interaction models that limit predictive accuracy.Machine learning interatomic potentials bridge this gap by reproducing quantum-accurate forces and energies at a fraction of the computational cost. These methods allow simulations of realistic nanoparticles under varied thermodynamic conditions, capturing nucleation, crystallization, diffusion, and phase transitions with unprecedented resolution.This thesis applies these machine learning potentials to study the formation, growth, and structural evolution of nanoparticles, focusing on gold-silver nanoalloys, zinc, and zinc oxide systems. For Au-Ag nanoalloys, simulations reveal details of the crystallization process and gold surface segregation, with obtained morphologies and chemical ordering validated against experimental results. For ZnO nanoparticles, growth simulations under different temperatures and deposition conditions highlight phase transitions between body-centered tetragonal and wurtzite structures. Here, an additional in-house machine learning technique based on Gaussian mixture models was key for identifying different ZnO polymorphs.The work further addresses the challenges of developing a multi-system machine learning potential for Zn ZnO, particularly due to the difficulty of modeling hexagonal close-packed zinc. Its flat energy landscape and anomalous structural ratios make accurate predictions highly sensitive to the choice of training data and model architecture. However, the latest results are promising, suggesting that continued iteration and refinement should yield an accurate model.Overall, this thesis demonstrates that machine learning-assisted simulations offer a robust framework to investigate nanoparticle synthesis and growth, offering insights that guide the rational design of nanomaterials with tailored properties.