CBE Centennial Seminar Series: Arun Yethiraj (UW-Madison)

Machine Learning for Phase Diagrams of Complex Fluids

Machine learning (ML) has become an important tool in computational chemistry.  This work describes the use of supervised and unsupervised ML methods to obtain the phase diagram of complex fluids from computer simulations.  A convolutional neural network approach, based on a grid interpolation of particle positions, successfully predicts the phase behavior of off-lattice systems, e.g., the Widom-Rowlinson mixture and symmetric polymer blends.  The method is too computationally intensive, however, for more complex polymeric systems.  A deep neural network approach, based on structural and thermodynamic input from computer simulations predicts the phase diagram of polymers in ionic liquids.   The disadvantage of supervised methods is that they require some knowledge of the phase diagram for training purposes.  Unsupervised methods can predict the phase diagram of off-lattice systems, without any prior knowledge.   We propose a method that combines a judicious choice of the input features with a variational autoencoder.  The ML approach can predict the phase diagram of a number of systems, critical point and critical exponent included.

Biography

Arun Yethiraj is the V.W. Meloche-Bascom Professor of Chemistry at the University of Wisconsin-Madision. He received his Ph.D. from North Carolina State University in 1991. His research focuses on computational and theoretical studies of soft condensed matter.  While it is clear that the short-range structure of complex fluids plays an important role in determining the physical and chemical properties, the complexity of these systems makes modeling them on an atomistic level computationally challenging.  A judicious choice of coarse-grained models that hopefully capture the essential features without incorporating much of the detail is therefore a crucial step in the theoretical study of these systems. We are interested in constructing such models, and then employing liquid state theory, computer simulation, and machine learning to investigate their properties, with the final aim of predicting experimental observables.  Our research has two components: the development of force fields and methods, and the application of these to understand the structure and dynamics of condensed phases.