Skip to main content

The Na Group has published a new study in Computers & Chemical Engineering journal (top-tier journal in Process Systems Engineering). This study is a study in which machine learning techniques are applied to the field of Process Systems Engineering. We have developed a technique that can effectively infer the parameters of complex multiscale simulations representing chemical systems.

This study was conducted by Prof. Jonggeol Na when he was at Carnegie Mellon University as a Postdoc in Prof. Nick Sahinidis Optimization Group. In addition, Dr. Ji Hyun Bak in Redwood Center for Theoretical Neuroscience, University of California, Berkeley helped us with the core of the algorithm, so we were able to success this study.

Highlights

  • Bayesian inference from expensive first-principles does not have to be computationally demanding•
  • Low-complexity surrogate models of first-principles models facilitate inferencing of probability distributions•
  • The integration of mathematical programming and deep learning learns uncertainty distributions much faster than sampling techniques

Abstract

Bayesian inference is a key method for estimating parametric uncertainty from data. However, most Bayesian inference methods require the explicit likelihood function or many samples, both of which are unrealistic to provide for complex first-principles-based models. Here, we propose a novel Bayesian inference methodology for estimating uncertain parameters of computationally intensive first-principles-based models. Our approach exploits both low-complexity surrogate models and variational inference with arbitrarily expressive inference models. The proposed methodology indirectly predicts output responses and casts Bayesian inference as an optimization problem. We demonstrate its performance via synthetic problems, computational fluid dynamics, and kinetic Monte Carlo simulation to verify its applicability. This fast and reliable methodology enables us to capture multimodality and the shape of complicated posterior distributions with the quality of state-of-the-art Hamiltonian Monte Carlo methods but with much lower computation cost.

Schematic diagram of neural network training pipeline and model architecture.

Leave a Reply