(368aa) Computational Research Focusing on Particle Based Simulations and Machine Learning

Overview: I am a fifth year PhD student set to graduate in May 2025. While my major is mechanical engineering with a concentration on data science and engineering, my research falls more into chemical engineering and materials science. I do computational research, focusing on particle-based simulations (molecular dynamics, Monte Carlo) of soft materials, mostly coarse-grained polymers and colloids. I have a robust background in Machine Learning (ML) which I frequently apply and is central for my research. I have extensive coding experience in Python and C++ and I am proficient in both. I have also worked for a year as the Computational Teaching Assistant of the Materials Science department at the University of Illinois at Urbana-Champaign. In this position, I have helped instructors and undergrads of different levels; in programming, simulation and data analysis related components of their classes.

After completing a double major in Physics and Mechanical Engineering, I pursued graduate studies with a focus on computation, soft materials, and machine learning. Throughout my research and education, I have immersed myself in various scientific disciplines, which has equipped me with the confidence to tackle novel problems and adapt to evolving environments. I am looking forward to being part of a dynamic team addressing future technological challenges with innovative solutions. Below, I outline brief summaries and skills related to three of the projects I have worked on during my graduate studies.

• Project 1: Microplastics in aqueous environments
  

Project summary: Micro-Nanoplastics (MP-NP) are found everywhere, have unknown long-term health and environmental implications. MPs’ aggregation behavior in water, influenced by shape and flow conditions among many other factors, is critical to assessing their impact. In this project, we investigated the heteroaggregation of microplastics with natural colloids under various conditions, using coarse-grained molecular dynamics and Monte Carlo simulations. Our findings revealed that particle shape, which is often overlooked in large scale MP transport models, plays a crucial role in the structure and stability of heteroaggregates in environment. For instance, spherical and round microplastics form compact aggregates with higher resistance to shear flow, while faceted shapes with pronounced edges and corners like cubes and plates form more fractal aggregates that break up more easily under flow.

Publication: B. R. Argun, A. Statt, “Influence of Shape on Heteroaggregation of Model Microplastics: A Simulation Study”, Soft Matter, 2023, 19, 8081–8090.

Skills gained: We have used the in-house C++ code of our group for Monte Carlo simulations. The code was originally for spherical particles and had to be significantly modified to simulate different non-spherical microplastic shapes as composite-rigid groups of small spherical beads. As the system of interest was aggregating MP particles, configurations were getting kinetically trapped using traditional MC moves. To overcome this, I have researched and implemented the state-of-the-art cluster MC moves with intricate algorithms. Potential of the mean force calculations were also implemented for this project. Overall, I gained valuable experience in C++, and learned about object-oriented programming. After completing this project, I was more comfortable working with large coding projects, implementing new features respecting the existing infrastructure, expanding the existing infrastructure to efficiently accommodate further modifications to the code. I have used GitHub for collaboration and version control. I have learned the value of unit and system tests, documentation, and acquired systematic debugging skills.

• Project 2: Accelerating Anisotropic Particle Simulations with Neural-Nets
  

Project summary: Thanks to the advances in chemistry in the last 20-30 years, it is possible to synthesize non-spherical colloidal building blocks which can self-assemble into various targeted nano-structures with desired optical, electronic and structural properties. In addition, pollutants such as microplastics or engineered nano-particles can occur in different shapes in the environment. Thus, capturing anisotropic features of particles is essential to accurately simulate and understand the fundamentals of a wide range of phenomena in science and engineering. Rigid bodies, made of smaller composite beads, are commonly used to simulate anisotropic colloids with molecular dynamics or Monte Carlo methods. To accurately represent the shape of the colloid and to obtain smooth and realistic effective pair interactions between two rigid bodies, each body may need to contain hundreds of spherical beads. Given an interacting pair of particles, traditional molecular dynamics methods calculate all the inter-body distances between the beads of the rigid bodies within a certain distance. For a system containing many anisotropic colloids, these distance calculations are computationally costly and limit the attainable system size and

simulation time. For example, it is not feasible to investigate the effect of shape on the amount of defects or long-range order of self-assembled structures when the driving mechanism is attractive interactions. Similarly, we are limited to tens of particles when simulating aggregation of non-spherical nanopollutants into clusters. In this study, we bypass the costly distance calculations with neural-nets. In principle, there is a function capable of directly mapping the center of mass distance and orientation to the interaction energy between the two rigid bodies, which would completely bypass inter-bead distance calculations. It is challenging to derive this function analytically, so in this study, we approximate it using neural-networks. We achieve up to 30 times speed up over traditional molecular dynamics, depending on hardware and system size while accurately reproducing both structural quantities and dynamic quantities observed in the traditional simulations.

Publication: B. R. Argun, A. Statt, “Molecular dynamics simulations of anisotropic particles accelerated by neural-net predicted interactions”, J. Chem. Phys., 2024, 160, 244901

Skills gained: This work has started as a project for my “Scientific Machine Learning” class which was designed to help grad students to put the theoretical knowledge of ML into practice and evolved into an actual research project. Through this work, I have experimented with various supervised learning techniques such as decision trees, k-nearest neighbors, gaussian process regression etc. I gained extensive experience working with feed-forward neural networks for regression and classification tasks. I have appreciated how critical feature engineering is, since the most challenging part of the project was preprocessing the raw data into a format that is easy to learn for the ML model. The main goal of this project was to develop a faster alternative to traditional MD simulations, so I spent significant time on code performance optimization and various related tools. For example, I have learned how to quickly identify performance bottlenecks, how to benefit from GPUs to perform costly computations such as geometric operations and inference through the neural nets, specifically I used TENSORRT, libtorch packages for C++ and cupy, pytorch packages for python. To efficiently simulate non-spherical particles, I gained expertise on handling rotations and orientations using quaternions and Euler angles. I also had the chance to mentor an undergraduate student for this project.

• Project 3: Structural and Mechanical Properties of Polymer Microgel Particles:
  

Project summary: Polymer networks are found in a wide range of material systems with significant technological significance, such as membranes, gels, and coatings. During synthesis, internal and nanoscopic physicochemical heterogeneities develop in the networks, which in turn dictate their mechanical properties and swelling-deswelling behaviour relevant for applications including drug delivery and water harvesting. It is not possible to resolve the internal heterogeneities using conventional imaging methods like scattering or electron microscopes, which limits our understanding of their effect on the network’s properties. In this collaborative project, our aim was to explore the internal heterogeneities using a model network of pNIPAM microgel particles.

Our experimental partners imaged the microgels using liquid-phase TEM (LP-TEM). In parallel, we have modeled the microgels using a coarse-grained molecular dynamics simulation with implicit solvent. The model consisted of bead-spring chains and crosslinker particles that react during the simulation to form the microgel. To obtain atomistic representations of the model microgels, “backmapping” is employed where coarse-grained beads were replaced with corresponding atoms, respecting the chemistry and topology of pNIPAM. Atomistic microgels were used to generate micrographs (simulated TEM images) that were compared with the real microgel TEM images. The parameters of the coarse-grained microgel model were iteratively tuned to match the internal heterogeneity and radial density profiles of the real microgels. By combining simulation and experiments, we have confirmed the hypothesis that the high-density regions observed in the LP-TEM images correspond to a clustered crosslink distribution and shown that most of the nano structural response occurs in the portions of the network that connect “locked” clusters.

Publications: B. R. Argun, A. Statt, “Interplay of Spatial and Topological Defects in Polymer Networks” ACS Engineering Au, 2024

Heterogeneous Internal Nanostructure and Dynamics of Responsive Polymer Networks from Direct Nanoscopic Imaging” In preparation

Skills gained: In this project, I had the chance to work closely with experimental researchers, and learned to effectively communicate with people with different skill sets and backgrounds, such as chemists who synthesize the particles or material scientists looking at them with LP-TEM. I have learned about both capabilities and limitations of models and simulations in connection with real-world systems. I have performed atomistic simulations, experimented with force-fields and used tools from graph theory to analyze microgel networks.