Adversarial Learning Methods for Modeling and Inverse Design of Soft Materials

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

The properties of the material world emerges from countless interactions between molecules and other microscopic structures. Soft materials are those which exhibit behaviors having a significant dependence on temperature. This includes liquid crystals used in display technology, gels and colloids used by industry in foods and consumer products, and constituents of biological systems. Insights into behaviors and the design of soft materials with specified target properties poses significant challenges given subtleties in how interactions and rearrangements at the microscopic level can vary with temperature, density, and other physical conditions. This calls for the further development of advanced computational methods for modeling, simulation, and optimization for soft materials. This project contributes new data-driven techniques and software tools for soft materials by leveraging and further developing emerging machine learning methods and simulation approaches. This includes adversarial training methods for learning representations of materials leveraging computational properties of competitive games coupled with further development of deep neural network architectures. The approaches will be used to develop tools for modeling and designing soft materials with target properties and for improving the fidelity and efficiency of computational simulations. Open source software also will be developed and released for use by the community. Outreach activities are planned for promoting diversity and engaging under-represented students both at the University of California Santa Barbara and in the local community. This includes working with local area K-12 schools and colleges on programs to engage students on topics in computation, data science, machine learning, and engineering. Educational activities are also planned providing unique opportunities to train the next generation of researchers and students on recent emerging machine learning approaches at the interface of engineering, mathematics, statistics, and data science. The project addresses challenges in developing data-driven approaches for modeling, simulation, and design of soft materials. The properties of soft materials arise from collective microstructure interactions having energies comparable to thermal fluctuations and from effects spanning a wide range of spatial-temporal scales. Given the role of fluctuations, collective entropic effects play a significant role. This presents computational challenges resulting in expensive large-scale forward simulations to characterize and design materials. The project develops new machine learning approaches and software tools for data-driven modeling and simulation of soft materials. This includes approaches for model reduction by identifying coarse degrees of freedom from high-fidelity simulations, methods for learning model parameters and force interactions, and optimization approaches for design of materials with target properties. The project leverages and further develops recent adversarial learning approaches to learn implicit generative models and other representations for improving the efficiency and fidelity of simulations. Methods are also developed for specific applications for data-driven modeling of colloidal systems and polymeric materials with target properties. Software tools also will be developed and released for the approaches to provide general methods for performing simulations, optimization, and analysis of materials. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.