This presentation was made at CAASE18, The Conference on Advancing Analysis & Simulation in Engineering. CAASE18 brought together the leading visionaries, developers, and practitioners of CAE-related technologies in an open forum, to share experiences, discuss relevant trends, discover common themes, and explore future issues.

Resource Abstract

Historically, engineers have relied heavily on experimentation in order to make useful predictions about the performance of their designs. This has produced many significant advancements of technology. A well-known example is the Wright Brothers’ use of a home-made wind-tunnel to measure the lift and drag characteristics of various airfoils, directly leading to the first successful heavier-than-air flight in 1903.

Analytical methods have also been useful, but until the advent of modern computers they were usually limited to relatively simple cases, such as laminar flow down an axisymmetric pipe, or deflection of a beam with a regular cross-section. However, now that adequate computing power is almost a commodity, problems with more complex boundary conditions and geometries can now be solved, and a multitude of commercial packages are available for this purpose.

While modern engineering analysis tools have dramatically improved our ability to model many problems, which in turn has enabled us to reduce or eliminate experimentation, there remain some problems which do not lend themselves to conventional analytical methods. Examples include prediction of fatigue failures, lamination failures, quality “drift” in a production line, and (returning to the original example of airfoils) the accurate prediction of critical angle of attack.

Machine learning is a relatively new approach which has been applied to a wide variety of computing tasks where designing and programming explicit algorithms with acceptable accuracy and performance is either difficult or impossible. Machine learning models allow researchers, scientists, engineers, and analysts to generate reliable and repeatable predictions through learning from historical trends in (usually a large quantity of) experimental data.

This paper discusses the application of the various open-source machine learning tools to problems of engineering significance. Tools investigated include TensorFlow™, an open-source software library developed originally for internal use at Google, and also Scikit-learn (formerly scikits.learn), a free software machine learning library for the Python programming language.

As a complement to this research, the authors have also applied a commercially available low-code development platform (EASA). This type of technology, also known as “hpaPaaS” – High Productivity Application Platform as a Service – enables “citizen developers” or “authors” to “appify” and democratize “expert-only” software tools and models. It enables non-programmers to codelessly create custom, fit-for-purpose apps, complete with error trapping and design or business rules embedded. Previous work has highlighted the value of this approach in leveraging models created with Excel, MATLAB, R, or commercial codes. However, the current work demonstrates the feasibility and convenience of extending this approach to deploy models based on machine learning, which would otherwise remain usable only by computer scientists.