Scientific Machine Learning (SciML) Surrogates for Industry, Part 1: The Guiding Questions

Surrogates are rapidly growing in importance as a technique from scientific machine learning for accelerating modeling and simulation. However, much of the current work on surrogate modeling has kept to the domain of academic literature and many techniques have not broadly been adopted in standard industrial practices. What is required for surrogates to become commonplace or standard in industrial design and control? In this position paper we discuss the various challenges associated with translating surrogate techniques of scientific machine learning into a method for industrial usage. We highlight the issues which academic oriented research overlooks that prevent these techniques from scaling to real world applications that are maintained by engineers with no machine learning background. We then motivate a series of requirements that address these issues which allow these techniques to be reliable and usable in those very same environments. This then then become the design basis of a new surrogates-enabled component-based modeling software, JuliaSim which democratizes the process of generating surrogates by engineers without requiring machine learning expertise.