Why Design Optimizations (Structural, Material, etc.) by FEA Do Not Apply to Every Case?
Despite of their tempting concept and attractive virtual visuals, the optimized designs functionality and integrity are more questionable. The age of generative design and AI based design exploration tools shows its unreliable face once it is employed for unsupervised engineering decision makings. A supervision that can only be done by expert human mind capable of provisioning things beyond binaries. When it comes to fully automatic design optimization via autonomous computational efforts, there are two technological challenges that computational tools are either yet to tackle or may never solve.
1. Impractical handling of large and complex models
2. Lack of inclusive design intuition and absence of engineering sense
Although this short and experience based opinion may seem anti-robotic and pro-human work force, it is intentionally made to share some expert customer clues to help design engineers focus on and empower their true values and guide the developers of commercial FEA packages to find the right questions to address.
Limitation of computational power in design optimization and exploration of mega-models is not only a problem for average users but also the main challenge of richest R&D departments around the globe. This further intensifies once design engineers are in the final phase of the design review just before the manufacturing. Spending an average of 2/3 of the whole product lifecycle expenses only in the design process of advanced machines but delivering weakly optimized “engineering marvels” to the real world is typical. This lack of competency for idealizing the end-solution for market and research sectors is mainly stemmed from design engineers illusive expectations from computers. For example, dealing with a model having 3 million finite elements with more than 200 contacts exposed to a short-cycle fatigue scenario, given the current and near future capability of FEM based software packages, it is simply impossible to optimize the whole system even with the simplest optimization roadmaps. Let alone multi-purpose multiphysics optimizations of advanced nature for multiple scenarios. In these cases having a work-force of different disciplines and some eagle eyes for categorizing and breaking down the entire systems into appropriate sub-systems and sub-functionality is quite vital for a simple optimization of the system. The age of overdesigning and blind trust on automation long passed and those who ignore this are joining the history of the industrial failure and market disaster one after another. The surviving path is to trust third-party experts who know by knowledge and experience how to approach and decompose your massive design models into different portions and relate them back together considering its geometrical and kinematic assembly, background physics, working environment and target functionalities.
Design intuition and engineering sense on the other hand, very similar to human consciousness, is not yet scientifically defined thus improbable to replicate and implement via AI. That is the main reason that engineering consultancies not only continue to exist today but also gain more reputation and values as the market demands more progressive solutions. Elaborating on the computer aided calculation is what we call Computer Aided Engineering or CAE (despite many mistaking it for a term referring merely to a certain family of commercial softwares). This elaboration is best done by engineers who gained practical and theoretical insights into the basic and fundamental matters of mechanical engineering which forms the establishment of any physical design. From micro concepts like considering the difference between contact pressure and bearing stress where needed, to macro decisions such as following the most sensible path from internal model verification to external final validations, etc. these engineers are experienced and trained to perform flexibly for the given problem according to its specific needs and unique characteristic. The latter is extremely out of scope, if not impossible, to achieve by pure automatic computational algorithms as they need to be at least tuned every time to the new demand with unknown parameters. And these parameters only become clear as the analysis process goes on with attaining different personal based engineering qualitative reliability factors which are not defined in the digitalized world.
Both explained issues hit their most complex level when one is dealing with a range of uncertainties in their calculations input. Example of such doubtful inputs can be found in aerospace structures and devices working under extreme conditions, fusion reactors in-vessel components facing hundreds of million degree centigrade plasma and its disruption events, off-road heavy duty and racing vehicles, etc. where the predicted working environment and to be faced events can be considerably different from what was anticipated in prototyping phase. Doing extra iterations and/or dismissing some analysis in the design optimization process, toward a more efficient design optimization with better reliability, is something that computers are not able of. A computer optimizer needs to spend a month fully calculating and proving by a number that a design fails to be optimized in a certain way while an expert can tell you that in one day guesstimating the odds and exploring the others experience. For instance feeding ASME standard into an AI based optimizer can result in designing a component failing in the normal operation service only due to the fact that blindly standardizing the design by the AI resulted in false consideration of pre-loaded stress as non-primary damaging stress in permanently affected components. A non-standard yet useful consideration that only expert engineers have.
In short, design optimization via FEA can not confidently be trusted since its backgrounds, FEM and computational algorithms made for parametric changes, are immaturely developed based on perfect mathematics rather than engineering observations.