Session A51: Big Data, Polymers, and Soft Matter: New Developments in Machine Learning, Data Mining and High-Throughput Studies

Machined-learned softness as a structural order parameter for understanding glassy systems

All solids flow at high enough applied stress and melt at high enough temperature. Crystalline solids flow and premelt via localized particle rearrangements that occur preferentially at structural defects known as dislocations. The population of dislocations therefore controls both how crystalline solids flow and how they melt. In disordered solids, there is considerable evidence that localized particle rearrangements induced by stress or temperature occur at localized flow defects but all attempts to identify them directly from the structure have failed. Here we describe a an application of machine learning data mining methods to diagnose flow defects, or “soft” particles from their local structural environments. We follow the softness of each particle as it evolves under deformation or temperature. Our results show that machine learning methods can be used to gain a conceptual understanding of glassy dynamics and of plasticity that has not been achieved with conventional approaches.   

Accelerated Discovery in Polymer Materials Domain: Knowledge Extraction and Representation

Acceleration of scientific discovery requires resolution of multiple bottlenecks along the flow connecting the perception of the available data, data analysis and hypothesis generation, and data acquisition via experiments. The challenges of a) ingestion of fast-growing volume of unstructured scientific data, and b) efficient generation of actionable hypotheses come from the limitations of human cognition. We investigate various approaches to the augmentation of the respective capabilities of human subject matter experts in the domain of polymer materials. Transition from curated datasets/databases to scientific knowledge graphs (sKGs) plays central role in this effort. In this talk, I discuss the technical aspects of the construction of sKGs from unstructured data in polymer materials domain, and utilization of sKGs for hypothesis generation, including human-in-the-loop approaches.
  

Exploring Free Energy Landscapes with Neural Networks

The use of adaptive sampling algorithms is an indispensable part of modern molecular simulations. A wide array of techniques have been developed to accelerate the sampling of phase behavior and molecular conformations, with those exhibiting the proper mix of simplicity and power gaining wide acceptance within the simulation community. Here, we discuss recent efforts to incorporate machine-learning techniques, in particular, artificial neural networks, to drive sampling and efficiently obtain free energy landscapes from incomplete representations of the mean force and partition functions. We will discuss conceptual and numerical aspects of these approaches, presenting new improvements alongside applications of the algorithms to soft and biological materials.
  

Text and Data Mining for Material Synthesis

Data has become a fundamental ingredient for accelerating and optimizing materials design and synthesis. Molecular synthesis planning, driven by advances in machine learning, has recently achieved human-level performance for the retrosynthetic design of organic molecules. The acceleration of data-driven synthesis planning and related analyses has, in part, been enabled by access to massive datasets which tabulate known chemical reactions. While macromolecule, polymer and inorganic materials databases also exist, the focus of these databases is primarily on materials structures and properties, rather than reactions and synthesis. Indeed, there is currently no comprehensive dataset which organizes the methods by which these materials are synthesized or even extensive property information. Comprehensively extracting the knowledge contained within written inorganic materials syntheses, without the use of significant human effort, is a key step towards reducing the overall discovery and development time for novel materials. This presentation will describe work to extract information from peer reviewed academic literature across a range of inorganic solid state materials synthesis approaches. We have demonstrated not only the potential of the natural language processing (NLP) approach to assemble materials data from the literature, but we have also shown that one can develop hypotheses for what synthesis conditions drive a particular target material outcome using learning approaches.
  

Data-driven learning of collective variables to understand and accelerate biomolecular folding

Data-driven modeling and machine learning have opened new paradigms and opportunities in the understanding and design of soft and biological materials. Nonlinear dimensionality reduction and deep learning present a powerful means to identify the underlying dynamical modes governing the assembly and folding of soft materials such as colloids, peptides, and polymers by direct analysis of molecular simulation data. Recovery of these modes, together with nonlinear collective variables with which to parameterize them, provides fundamental understanding of the microscopic forces and emergent dynamical motions governing the long-time evolution of the molecular system. We will discuss our use of diffusion maps, deep neural networks with novel topologies and loss functions, and enhanced sampling techniques to recover the high variance and/or slow collective variables from molecular dynamics simulations of protein folding, and our subsequent use of these coordinates to understand folding mechanisms and guide and accelerate sampling.