Scientific Machine Learning: Opportunities and Challenges - Keynote

Explore the fascinating world of Scientific Machine Learning (SciML) in this keynote address from JuliaCon 2020. Delve into the opportunities and challenges of integrating machine learning with scientific computing, examining how physics-based models and big data intersect in computational science. Learn about complex multiscale phenomena, high-dimensional parameters, and the challenges of sparse, intrusive data. Discover innovative approaches like the Lift & Learn method, illustrated through a rocket engine combustor example. Gain insights into model reduction techniques, their similarities and differences with machine learning, and how they can be combined for optimal results. Conclude with a vision of the future of computational science and programming languages, followed by a Q&A session addressing practical advice, potential pitfalls, and the impact of SciML on non-linear systems.

Welcome and information about JuliaCon 2020.
Introduction and acknowledgments.
Outline of the talk.
What is Scientific Machine Learning (SciML)?.
What are the opportunities and challenges of SciML?.
BIG DATA alone is not enough.
Using physics base models means that we are doing Computational Science.
Problem 1: Complex multiscale multiphysics phenomena.
Problem 2: High dimensional parameters.
Problem 3: Data are sparse, intrusive and expensive to acquire.
Problem 4: Rare events.
Problem 5: Uncertainty qualification.
SciML and Computational Science, summary.
Example: flow inside a rocket engine combustor.
Example: equations of flow inside a rocket engine combustor.
Physics-based model are powerful but computationally expensive.
Overview of model reduction methods.
Similarities and differences between model reduction and ML.
Can we have the best of two worlds (model reduction and ML)?.
Overview of Lift & Learn approach.
Example: Lift & Learn approach to 2D flow in a rocket engine.
Example: Training of the model.
Example: Comparing results for pressure and temperature.
Outlook of SciML.
Diverse future of computational science and programming languages.
Q&A: Advice for people bringing ML approach to scientific problems.
Q&A: Pitfalls of the interplay between domains knowledge and ML.
Q&A: Does the present approach improve the fidelity of solution of highly non-linear systems?.