Session: 37-02 Radial Turbomachinery Optimization

Paper Number: 125831

125831 - Machine Learning Application to Centrifugal Compressor Design 

It is common practice for centrifugal compressors to be custom designed for a specific application in the world of industrial compression. This applies to a number of compressor applications including reinjection, boil off gas, pipeline, process gas, etc. As with all engineered solutions, the design process starts with a concept that demonstrates key features required to meet the design objectives, and then the design matures through the process to the final manufacture of the engineered product. An important step in this process is the initial conceptual design, often referred to as preliminary machine selection. It’s called a selection since the design solution is a sum of selected, pre-engineered solutions for aerodynamic performance, mechanical integrity, etc.

Significant effort can be expended in conducting a preliminary selection. First, a selection tool (or set of tools) must exist that allow an engineer to evaluate various combinations of pre-engineered solutions. Second, an application engineer must try various combinations of selections based on their own experience, similar machines previously selected, or the experience of others. In the case where response time is important, it is desirable to complete such a preliminary design in hours as opposed to days. In practice, fast response is usually feasible only with applications very similar to previously executed projects or in the presence of an application engineer with extensive experience in the application range. The common thread here is experienced based knowledge. This is the essence of machine learning (ML). However, instead of the learning being contained in the individual, it is contained and maintained by a digital library.

In this work, the authors present results from a study to leverage all previous industrial compressor builds (already known to be successful designs) to inform a digital library. This library is then used as training data for a ML model capable of presenting the application engineer with a plausible design solution within minutes. Preliminary efforts to prototype this concept for a narrow application space using a spreadsheet approach has yielded promising results. A sample compressor solution selection obtained from this approach to meet new process requirements is shown in Figure 1.

Starting from this prototype, the authors have extended the work to a more general case (wide range of gas composition, flow rates, pressure levels, etc.) and providing more detail for the preliminary design including down select of internal components from the standard component libraries. It is also possible with such a machine learning approach to layer on an optimization routine so that various objectives can be optimized such as machine footprint, capital cost or efficiency.

Presenting Author: David Ransom Siemens Energy

Presenting Author Biography: David Ransom has over 25 years of experience in the turbomachinery industry. He joined Siemens Energy as Head of Technology & Systems in November of 2019. His current responsibilities include leadership development, leading and managing technology development, support for Operations and Services, leading Root Cause Analyses, developing innovative and disruptive technologies, and R&D strategy and planning. David has experience in several technical disciplines including rotordynamics, thermodynamics, structural dynamics, acoustics, machine design, prototype development and testing, and leading people and projects through all phases of the project, from proposal to final report.

Authors:

David Ransom Siemens Energy
Ravichandra Srinivasan Siemens Energy