Learning Physics from Machines: Physics-Informed Machine Learning


Principal Investigator

Ashwin Dani

Co-Principal Investigator

George Bollas

Project Summary

This project aims to develop methods and toolchains for computationally efficient and precise modeling of large-scale systems. In the first year of the project, a spatially discretized heat exchanger system model with phase change was used as a testbed for exploring novel model reduction methods and constrained machine learning (ML) algorithms. These two approaches were used for the synthesis of reduced models of a heat exchanger, which were tested at nominal and T-optimal conditions (the latter being conditions within the system allowable state space that maximize the error of the reduced, machine learning generated, or hybrid models as compared to the original model). The project is performed in collaboration with Pratt & Whitney and UTAS, with the former leading the effort. P&W has identified the application of interest for the developed method in aerospace engine systems. There is an identified need for simplified, accurate system models for real-time applications. Therefore, the technology, methods and tools developed in the first year of the project are focused on reducing model complexity and improving model accuracy. The ongoing effort seeks to combine machine learning and model reduction techniques in order to create useful hybrid models, and to further develop the individual tools and techniques applied.

Program Applications

-Initial: Generic counter-flow heat exchanger

-Interim: P&W NGPF Partial Engine Model

-Final: Full engine model, additional P&W/UTAS applications

Project Outputs

-3 PhD and 1 Masters students trained

-6 peer-reviewed scientific papers

-4 conference presentations, 1 invited talk

-Software artifacts for Constrained Learning, Model Reduction, Hybrid Modeling, and Optimal Experiment Design

Business Unit Benefit

-Structured frameworks for model reduction, constrained machine learning, hybrid modeling, and model selection with broad applications to cyber-physical systems

-Integration of above tools for the creation of models with tailored complexity for specific applications

-Improvement of key estimation accuracy in deployable real-time models

Talent Creation

-4 Students have worked on this P&W/UTAS project

Excellence Artifacts

-6 scientific papers, 4 conference presentations, 1 invited lecture

-AIChE Computing and Systems Technology Division Director’s Award for poster presentation on Model Reduction by Term Elimination and Optimal Selection, 2017

Business Unit Engagement

-4+ collaborators from P&W/UTAS

Example Publications and Further Reading

-B. Baillie and G. Bollas. Model reduction by term elimination and optimal selection. Computer Aided Chemical Engineering, 2017: 40:277-283.

-B. Baillie, H. Ravichandar I. Salehi, A. P. Dani, and G. Bollas “Approaches for Creation and Evaluation of Computationally Efficient Thermofluid System Models”, IFAC International Symposium on Advanced Control of Chemical Processes, 2018; Accepted for publication.

-H. Ravichandar, I. Salehi, and A. Dani, "Learning partially contracting dynamical systems from demonstrations," in 1st Annual Conference on Robot Learning, In: Proceedings of Machine Learning Research, vol. 78, pp. 369-378

-H. Ravichandar, I. Salehi, B. Baillie, G. Bollas, and A. Dani, "Learning Stable Nonlinear Dynamical Systems with External Inputs using Gaussian Mixture Models" American Controls Conference, 2018, Accepted for publication.