Paper Number
SC4
Session
Suspensions, Colloids, and Granular Materials
Title
Constitutive model selection using neural networks
Presentation Date and Time
October 21, 2019 (Monday) 11:05
Track / Room
Track 2 / Room 304
Authors
- Blackwell, Brendan C. (University of Pennsylvania)
- Arratia, Paulo E. (University of Pennsylvania, Mechanical Engineering and Applied Mechanics)
Author and Affiliation Lines
Brendan C. Blackwell and Paulo E. Arratia
Mechanical Engineering and Applied Mechanics, University of Pennsylvania, Philadelphia, PA 19104
Speaker / Presenter
Blackwell, Brendan C.
Text of Abstract
In this work we apply the tools of machine learning to improve the process of constitutive model selection. Choosing a model to characterize empirical data is a nuanced process. A complex model with many parameters will always provide the closest fit to the data, while a simple model with fewer parameters is often more valuable to understanding the underlying physics that are at play. A good approach is to start with the simplest model possible and gradually add complexity until the fit is sufficient for the given purpose, but for many models this is a high-dimensional problem with several choices for where and how to add complexity. Beginning with experimental data of a colloidal suspension (shear rheology of aqueous kaolinite clay) that shows an unusual signature, we set out to find a model that captures the unique signature without having so many parameters as to become unwieldy. We begin with a simple, structure parameter thixotropic model, and show that the simplest version does not capture key aspects of the data. Faced with a choice of where in the model to add additional parameters, we turn to machine learning to make a complicated task computational feasible. After writing a ‘master’ model with fifteen parameters, each of which can be interpreted to have a physical meaning, we use a neural network to determine what subset of the parameters is optimal. A Bayesian inference criterion is used to penalize the number of parameters, preventing an overly complex result. The technique of the neural network allows us to evaluate all possible combinations and subsets of parameters, a task that would otherwise be computationally infeasible.