In this blog, we will discuss the issues we face when constructing a Machine Learning (ML) system for turbomachinery design and optimization.
We will learn how ADT’s 3D Inverse Design method and Reactive Response Surface (RRS) algorithm combine to create a practical and efficient ML tool, and how that tool, contained wholly with the TURBOdesign1 environment, delivers blade designs optimized for multiple operating points and multiple competing objectives (for example noise generation versus pressure rise).
● Axial fans - a simple idea with hidden complexity
● The challenge of Machine Learning for turbomachinery design and optimization
● ADT’s pillar components to enable Machine Learning
● Towards expert systems for practical design cases
Axial fans - A simple idea with hidden complexity
Axial fans are probably the most ubiquitous air movement machines. From consumer goods, though HVAC, automotive, to civil and military aerospace applications, axial fans are found wherever high volume, low pressure delivery is required in the most straightforward installation configuration.
Although simple in conception, the flow structures around an axial fan blade are complex, with plenty of opportunity for the flow to migrate cross-passage, spanwise and over-tip. Controlling these flows is the key to improved performance, reduced noise and lower operating costs.
Figure 1: Cross passage flow and separation in the fan blade passage
A machine learning system should, in theory, be able to find a close to optimum design for this type of blading, especially when tailored to a unique operating requirement. However a number of obstacles must be overcome to turn this into a reality.
The challenge of Machine Learning for turbomachinery design and optimization
Firstly we can rule out applying a Large Language Model (LLM) type approach to this problem. The training data for such a system relies on millions or billions of items of cheap, tagged data and the data for turbomachinery design and performance is not organised like this. It is limited and in many respects strictly proprietary and guarded.
Then there is the problem of parameterization. How to efficiently parameterize a complex 3D shape that (a) accurately describes the blade geometry and (b) provides enough flexibility in the geometry envelope to effect meaningful performance change. If one requires hundreds of parameters to define the design space that is going to be extremely costly to explore and manage.
Turbomachinery performance evaluation also requires high-fidelity simulation data, which is very costly to generate. The longstanding block against running full-fidelity optimization has been the need to evaluate thousands of candidates in 3D CFD, many of which come nowhere near improvements to the baseline, but are nevertheless required to inform the optimizer about the topology of the design space. That is a lot of core-hours and/or a lot of licenced solver time.
ADT’s pillar components to enable Machine Learning
3D Inverse Design needs only a handful of parameters to describe the blade shape, The parameters are effectively performance demands and the blade shape is the output of the given work distribution. The blade shape for a given parameter set is arrived at in a matter of seconds, along with surface pressure and velocity distributions, plus loss estimations.
Figure 2: Changing one parameter drives large changes in overall blade shape
Because performance is the input to the inverse design method, with blade shape being the output, each created candidate is guaranteed to meet the basic duty requirements (e.g. pressure rise for a given mass flow and rotor speed). Therefore the variation in design can be clustered far more closely around the target design point, with the hinterland of non-compliant designs being completely ignored. This approach (a) reduces the dimensionality of the design space - easing the effort required to explore it and (b) removes the wasted effort of producing and evaluating candidates which cannot even meet the basic performance requirements set out in the original specifications.
We also note that turbomachinery optimization is a multi-point discipline. Designs must deliver improved performance at a number of operating points. It is no good having outstanding performance at the design point if off-design performance is completely compromised. Most turbomachines operate at a range of flows and speed and undergo transient and part-load conditions as part of their normal operational requirements.
Figure 3: Typical pressure-flow curve for an axial fan. Widening the optimum range, which is a multi-point optimization challenge, can be as important as improving peak performance
Reactive Response Surface (RRS) is an extremely efficient method for conducting focussed searches in a multi-dimensional design space. It enables rapid convergence to a global optimum using a comparatively small dataset. Coupled to the powerful yet sparse blade parameterization, ADT and our clients have repeatedly demonstrated the practical application of RRS Machine Learning that delivers real multi-objective performance improvements in practical time-scales to the end product.
Ultimately, high-fidelity CFD is needed to evaluate performance, but should only be used for near-optimal candidates. ADT’s Reactive Response Surface + CAE optimizer algorithm is smart enough to use lower-fidelity methods for the majority of the design space search, exploiting the high efficiency of the Reactive Response surface method, only switching to high-fidelity methods to verify candidates near to the Pareto front. Moreover, any data generated in the CFD simulation is fed back to the optimizer to inform next steps. This process is seamless and invisible to the user within the TURBOdesign1 GUI, with the Reactive Response Surface optimizer automatically switching between low- and high-fidelity methods as required.
Towards expert systems for practical design cases
ADT have developed and assembled a set of unique technologies that enable a true Machine Learning System. Contained seamlessly within the TURBOdesign1 environment, this system is designed to operate in practical timescales on modest compute resources to discover and evaluate optimal multi-objective, multi-point turbomachinery blade designs.
Figure 4: The building blocks to a Turbomachinery Expert System
In part 2 of this blog, we will discuss how to apply the ADT Machine Learning system with the Reactive Response Surface + CAE method to the problem of improving the multi-point performance of an Axial Fan for an industrial application.