Machine-Learning Assisted Atomic Simulations of Defect Dynamics in Multicomponent Concentrated Alloys

This dissertation investigates the complex defect diffusion behaviors in concentrated solid solution alloys (CSAs) including high-entropy alloys (HEAs), which are critical for understanding their exceptional mechanical and radiation-resistant properties. Through a combination of multiple atomistic-level simulation techniques and novel machine learning methods, this work reveals how the intricacies of local atomic arrangements and chemical heterogeneities influence diffusion processes, thereby offering new insights into alloy design and optimization. The research initially focuses on the vacancy-mediated diffusion employing binary Ni-Fe concentrated alloys as model systems. To evaluate the impact of local chemical short-range orders (SROs) on vacancy diffusion, both random solid solution configurations and alloys with SROs are prepared using hybrid molecular dynamics (MD) and metropolis Monte Carlo (MMC) methods. The results demonstrates that the development of SROs can significantly impede vacancy-mediated diffusion and enhance the chemically biased diffusion between Fe and Ni sites. Such findings suggest that the diffusion behavior in CSAs can be intricately controlled by adjusting the chemical ordering, a principle that could revolutionize alloy design strategies. Moreover, the study establishes a linear correlation between changes in the enthalpy of mixing and the formation of SROs, indicating that the reduction of enthalpy of mixing towards the more negative direction within an alloy system acts as a driving force for the observed diffusional slowdown. Advancing the methodological frontier, this dissertation introduces a state-of-the-art approach that integrates machine learning (ML) with kinetic Monte Carlo (KMC) simulations to efficiently investigate the controversial phenomenon of "sluggish diffusion" in concentrated alloys. As the first step, the Ni-Fe concentrated alloys are used as model systems. The complexity of defect diffusion in varying local atomic environment in CSAs makes it impractical to apply the standard nudged elastic band (NEB) method for on-the-fly determination of defect migration barriers at each step. By developing an artificial neural network (ANN) model trained on a dataset of NEB-computed migration barriers, it enables precise, efficient, and on-the-fly predictions of vacancy migration barriers for arbitrary local atomic environments during KMC simulations, including both random solution configuration and alloys with SROs. The diffusivities derived from this ANN-KMC modeling closely align with those from independent MD and temperature-accelerated dynamics (TAD) simulations at their accessible temperatures. The research delves into the sluggish diffusion mechanisms over the entire composition range of the Ni-Fe alloy system, elucidating them through the lens of ANN-KMC-derived insights at both high and low temperatures. The exploration then extends to quinary FeNiCrCoCu HEAs, utilizing a similar but improved ANN model to predict vacancy migration barriers across a wide compositional range. Due to the challenges of exploring the vast HEA compositional space, to date most experimental and computational studies have been limited to equiatomic compositions. This model, remarkably effective despite being trained solely on equiatomic HEA data, accurately predicts vacancy migration barriers in non-equiatomic compositions and their binary to quinary subsystems. Implementing this ANN model as an on-the-fly barrier calculator for KMC simulations, such ANN-KMC framework derives diffusivities nearly identical to the those from independent MD simulations but with far higher efficiency. This capability facilitates an extensive study of over 1,500 HEA compositions, uncovering the presence of sluggish diffusion in many non-equiatomic compositions. The analysis provides critical insights into the interplay between compositions, complex potential energy landscape, and percolation effect of the faster diffuser (i.e., Cu) on sluggish diffusion behaviors, offering invaluable perspectives for experimental alloy design and development. Lastly, the dissertation delves into interstitial-mediated diffusion in FeNiCrCoCu HEAs, confirming the presence of sluggish interstitial diffusion by comparing the equiatomic HEA with a range of reference systems. To study the non-monotonic concentration dependences in interstitial diffusion, a machine learning KMC (ML-KMC) method has been developed to simulate 〈100〉 dumbbell interstitial diffusion across various HEA compositions. Diverging from conventional KMC (C-KMC) and random sample KMC (RS-KMC) approaches, which approximate transition energies through a mean-field and random sampling methods, respectively, the ML-KMC predicts dumbbell formation energy on-the-fly based on local atomic configurations. This enables it to effectively replicate diffusion patterns from independent MD simulations. This novel ML-KMC approach offers a promising high-throughput method for studying HEAs, avoiding the expensive computational overhead associated with calculating dumbbell migration barriers. The impact of the percolation effect of faster diffusing elements (Cr, Cu) is also analyzed. Insights from this study can advance the understanding of compositional-dependent diffusion and provide valuable insights for the HEA design. Beyond the achievement of these completed works, two promising future projects have been evaluated that could significantly advance the field of diffusion research. The first initiative seeks to broaden the scope of the ANN-KMC framework, aiming to significantly enhance simulation efficiency across a broad range of HEA compositions. An accurate ANN model for predicting interstitial migration barriers has already been developed, and its full integration into the KMC framework could enable more accurate diffusion simulations. The second project aims to develop a comprehensive ML interatomic potential tailored specifically for HEAs, intended to improve the predictive accuracy of MD simulations. Although progress has been made in modeling an equiatomic CoCrFeMnNi HEA, constructing a robust ML potential for HEAs faces substantial challenges, primarily due to the extensive data requirements and computational demands.