Combining Simulation and Machine Learning for Materials Optimization: Polymer Compatibilization, Disinfection, and Heat Transfer

In the last half-century, considerable advances have been achieved in molecular simulation techniques aiming at offering a comprehensive understanding of the structure-property relationship of soft materials on several time and length scales. So far, however, the optimal design of candidates for the next-generation soft materials is still a challenging task due to the enormous chemical and configurational space. The machine learning (ML) techniques, which are utilized to extract actionable insights from big data generated from simulations, can overcome the bottlenecks in the tasks of soft materials optimization. Hence, this thesis has developed a framework based on the mutual communication between multiscale simulations (atomistic and coarse-grained) and ML toward rational investigations of soft matter. One objective of this thesis is to evaluate the detailed structure-composition-property-performance rela- tionships of soft materials in a forward way. We firstly investigate the compatibilizing performance of block copolymers (i.e., linear and graft) on the interface between two incompatible polymer phases by dissipative- particle-dynamics (DPD) simulations. A phenomenological analytical power-law fit is developed to quantify the variation of compatibilization efficiency of linear block copolymers with the polymer chemistries, the molecular architecture, and the number of copolymer molecules. However, graft copolymers have larger diversities in the space of architectural parameters as compared to linear block copolymers, which limits the traditional empirical fitting process. Accordingly, we feed DPD results to ML models and find that the combination of DPD/ML is able to accurately predict the compatibilization efficiency of graft copolymers at the molecular level. For a given graft copolymer with several descriptors (e.g., molecular architectures and chemistries), its compatibilization efficiency can be well predicted from the trained ML models. Moreover, ML techniques provide a descriptor importance measure for the correlation between descriptors and DPD predictions. We find that as the blend changes from weakly incompatible to strongly incompatible, the number of side chains of graft copolymers gradually dominates their compatibilization efficiency while the side chain length becomes unimportant. This finding can narrow the search space in further simulations and experiments. Furthermore, we attempt to understand the compatibilization mechanism of the linear and graft copolymers by characterizing the beads distributions, the number of unlike contacts between different species, and the molecular conformations. Specifically, the relative shape anisometry of copolymers, defined as the ratio of their gyration tensor elements in directions normal and parallel to the surface, is strongly correlated with their compatibilization efficiency for both linear and graft copolymers. We also evaluate the alcohol-induced changes on coronavirus membranes of different compositions with DPD models, i.e., pure dipalmitoylphosphatidylcholine, dioleoylphosphatidylcholine, and dimyristoylphos- phatidylcholine as well as their binary and ternary mixed membranes. The principal finding of this study is that a maximum ethanol concentration of 32 mol % (55 wt. % ) in alcoholic-based disinfectants is sufficient to decompose any coronavirus model membranes composed of these three lipids. However, given the wide variations in compositions and structures of mixed membranes, identifying their transitions from the intact to the disrupted state is challenging. For example, we find that the transition point cannot be quantitatively predicted based on physical descriptors such as the area per lipid molecule, the membrane thickness, and the orientational order parameter. Additionally, the visual inspection of simulation profiles is cumbersome to characterize the state of these membranes. Developing a simple and robust tool to characterize the stability of membranes against ethanolic disinfectants, can therefore accelerate the optimization process of disinfection investigations. This target is achieved by the developed DPD/deep-neural-network framework in this study, which accesses the integrity of lipid membranes in place of visual inspections. The other objective of this thesis is to design materials with optimal performance on desired properties, com- positions, and structures in the reverse direction, i.e., inverse design (performance-property-composition- structure). We employ a hybrid framework by combining the genetic algorithm and the atomistic molecular dynamics simulation, to design polyethylene-polypropylene copolymers with high thermal conductivity. We find that polyethylene-polypropylene copolymers with various sequences at the same monomer ratio have a broad distribution of thermal conductivities. This indicates that the monomer sequence has a crucial effect on the thermal energy transport of the copolymers. A non-periodic and non-intuitive optimal sequence is indeed identified by this hybrid framework, which gives the highest thermal conductivity compared with both homopolymers and any regular block copolymers, e.g., diblock, triblock, and hexablock. In comparison to bulk density, chain conformations, and vibrational density of states, the monomer sequence has the strongest impact on the efficiency of the thermal energy transport via inter- and intra-molecular interactions. The success of ML, providing property predictions of materials in both large compositional and confor- mational spaces, relies on the availability of training data from simulations. In turn, ML methods allow a robust posteriori data analysis (e.g., descriptor importance measure) for exploring correlations between descriptors and target properties in simulations, which can narrow the search space of descriptors for further investigations. In short, the computational framework of integrating multiscale simulations with ML algorithms has a significant potential for accelerating the design of soft matter. We believe our work provides efficient and practical approaches to develop the advanced hybrid framework for the material optimization