Blog
I used to optimize turbocharger turbines for a living for a while, at ABB Turbo Systems (now Accelleron Industries), with PSO (Particle Swarm Optimization) and NSGA II, libFANN models, clustering algorithms, etc.
-
Expectation: “the component (e.g., turbine) can be made sooooo much better!”
-
Reality: the potential optimization gains hide across the interfaces of mechanical systems, as we have been squeezing blood from a stone for way longer than one decade, when it comes to singular system components.
Hence multiphysics simulations, multiobjective optimization, etc.
In such endeavors:
-
The optimization yield is only as good as the quality of your simulation (in terms of model and physical accuracy).
-
The training/validation datasets are costly because of proprietary simulation-software licensing models.
-
Parametrization and variant (geometry) generation play a big role in how well you explore the design space.
-
Simulation discontinuities (e.g. stress at the edges of finite-element models) impact gradient descent behavior (therefore, PSO and Evolutionary Algorithms are less prone to getting stuck in local minima).
And finally, much like as in the case of using ChatGPT:
- It’s better to first go wide than deep (i.e., do not fall for the gambling allure of greedy search heuristics.)
Exploration at ALARP (As Low As Reasonable Practicable) cost and time is where it’s at. Progressive enhancement of good results through the progressive elimination of bad regions of the design space or those that look great… until you see their robustness due to real-life variation (manufacturing constraints, operating conditions).
The best results came not from optimizing in the vicinity of an alleged globally optimal design, but from exploring the design space extensively and at low cost (thanks to “metamodels”) to find regions that combine performance and the robustness required in the real world.
Once this mindset grabs hold of you, it never leaves you, regardless of whether you design processes, develop products, build teams, manage operations, run a business, etc.
