Ian Steegmayer

Investigation of Nanoplastics and Mineral Interactions using Machine-Learning

The invention of plastics has shaped our society, but despite all its usefulness, it also brings severe problems. Only a small fraction of plastics are recycled, and large amounts of plastic waste end up in the environment. Mechanical wear and weathering finally lead to micro and nanoplastic particles of millimeters down to nanometer size. Those small pollutants are found virtually everywhere, including oceans, rivers, lakes, soils, and sediments, as well as in many organisms, including humans. The abundance of nanoplastics and its predicted increase demands an understanding of natural processes with nanoplastic interaction. One field of particular interest is the interaction of nanoplastics with geological materials such as clays, which are abundant in aqueous environments (e.g., ocean sediments) and on land. However, nanoplastic-mineral interactions have yet to be understood as they have not been investigated on the molecular scale (by molecular dynamics). This was found out by an extensive literature review during which available publications were searched for combinations of keywords like nano plastic, mineral, interaction, simulation, molecular dynamics, etc. While we encountered many experimental articles concerning nano-/microplastics interactions with various materials, publications on interactions between nano-/microplastics and living organisms, and various other topics, a molecular study of nanoplastic-mineral interactions like this work is targeting could not be found. Hence, this study aims to shed light on the fate of nanoplastics when in contact with minerals. To this end, molecular dynamics simulation of different nanoplastics and geologically relevant materials will be used. In a preliminary study, classical force fields (e.g., clayff) will be used to obtain the first insights into the systems to be investigated in detail. High-precision simulations of the most promising systems will be performed using state-of-the-art machine learning potentials such as DeePMD. These modern force fields/potentials do not rely on selected functions and parameters but are trained for the specific problem at hand using first-principles calculations. To prepare and train the problem-specific potential, we will develop a DFT dataset including all the relevant structures. The newly developed potential can be verified using well-tested force fields, for example, clayff or ReaxFF, albeit the accuracy of DeepMD is supposed to be on par with ab initio methods, which the classical force fields are not.
Once the framework for high-accuracy molecular simulations is set up, the objective will shift toward how nanoplastics can be captured and bound to geological materials. Here one or more machine learning tasks will be defined to optimize a system according to previously specified goals, similar to, e.g., in. The exact formulation of the problem will strongly depend on the findings of the first part of the project and cannot (should not) be fixed now. Possible investigations could encompass the surface topology of a given material, porosity-related properties, and much more. As machine learning is evolving extremely fast, the type of algorithm or the architecture to be used will also be decided just before that part of the project starts. This pathway ensures the possibility of choosing from the newest methods at that date. In the current CompSci course(s), we will also be introduced to more machine-learning techniques and gain computational background knowledge that will be incorporated into the project.