Robust
Designs
The Role of Variability (UQ) in Computational Design Optimization
Interview with André
Wilmes, CEO & Founder of Rafinex
Could you start by outlining the focus of your presentation at the upcoming CDFAM Symposium in Berlin?
In Berlin, Rafinex’ presentation will primarily showcase how to scale up stochastic topology optimization to large mesh sizes and assemblies from both a technological and business case interest point-of-view.
Regarding technology, I will be discussing both solver aspects as well as advantages that arise from cloud-first HPC algorithm that can be leveraged compared to classical on-machine algorithms and software. These advantages in turn enable novel business cases, first and foremost the ability to scale stochastic topology to entire vehicle assemblies at once – enabling a holistic robust optimization of systems rather than merely individual components.
Finally, I will briefly touch on special algorithm developments and application engineering cases, ranging from large-scale anisotropic carbon fibre or ceramic AM designs, thus showcasing the power of advanced computational design can achieve in real-life use cases.
Data Integration into Möbius Beyond CAD Geometry
Beyond geometry data from CAD, what additional types of data do engineers need to import into Möbius to effectively process full assemblies?
Stochastic topology optimization requires both CAD inputs for the available design spaces as well as adequate representations of the to-be-expected load conditions that the future structure is to withhold.
For the former, the CAD volumes can be rather rudimentary, while for the latter, the availability of quality boundary conditions and know-how of how to apply them to the topology optimization is a crucial ingredient and differentiator between achieving outstanding designs rather than academic or simplistic solutions.
While the ability to apply stochastic boundary conditions can mitigate the level of certitude that one may have with regard to the boundary conditions, it cannot alleviate for a total absence of this information.
It is of note that Rafinex has rolled out advanced modelling techniques, including couplings and springs, to assist with the application of realistic and representative boundary conditions. Furthermore, the latter enable the modelling of assemblies and systems, which is insofar helpful because boundary conditions and load inputs are often known at locations in the structure afar from the component-to-be-optimized.
For example, for suspension systems in vehicles, the forces are typically known for the contact patch with the street and the connection points into the white body, and Möbius can now represent this sub-system all-at-once, even if only one or more individual components within this assembly as subject to being optimized. For this specific example, we have in fact demonstrated that these load conditions may also be sourced from a Multi-Body-System (MBS) simulation, opening the gateway to coupling MBS and Topology Optimization at scale across all components in a vehicle.
Does Möbius use external solvers or have you developed your own, and how do these scale for very large parts and/or assemblies?
At present, our in-house record for full in-core stochastic topology optimization stands at around 25m elements in adaptive meshing during optimization, but we are already working on breaking through this current ceiling and achieve a step change in order of number of elements we may be able to solve by the end of the year.
However, we do not do this scalability efforts for the sake of technology developments, we do them because customers in aerospace, automotive and defence applications have demanding applications of large-scale components and/or assemblies while also requiring a high-resolution of small features at the same time. Namely, these capabilities allow for meter-sized components to be topology optimized while creating geometric features of the order of a few millimeters only.
With the exception of the embedded CAD kernel and meshing algorithms, Möbius uses no external components. The stochastic topology optimization kernel as well as the FEA validation kernel are developed in-house using proprietary and open-source components.
Considerable attention and work is dedicated to stringent Q&A testing, with dedicated test pipelines running unit tests to end-to-end test several times a week. It is of note that part of Rafinex know-how is a dedicated tuning of efficient coding, attention to detail in compiler settings while also ensuring the latter is tuned precisely to the cloud-HPC hardware and CPU infrastructure in question.
For instance, Rafinex regularly collaborates with early-access programs and tests newly available hardware and clusters such as AWS Graviton or the Luxembourg Meluxina systems to investigate the scalability of computational physics codes in these environments. While technologically interesting and challenging in itself, Rafinex does this work both in the customers’ interest as much as its own.
With the chosen business model of Rafinex, which does not include a variable compute pricing component and which sees Rafinex pay for compute resources on its end, the interest between partners and ourselves is aligned in that the developers have every interest in creating highly time- & cost-efficient code at runtime.
Real-world Applications and Client Use Cases
Can you share some use cases where your clients have successfully applied Möbius to optimize their designs and assemblies?
Public showcases published to date remain limited, unfortunately, given the confidentiality in the sectors we operate in. However, one outstanding reference case is published by Celanese, formerly Dupont, which used Möbius to optimize a polymer-based injection-moulded engine mount bracket for a large German automotive OEM. Celanese in Geneva managed to reduce the component’s weight by 25%, while equally reducing the strain energy in the part by 15%.
We have also successfully demonstrated & published how to optimize a robot gripper head for a range of highly variable and unknown conditions because it needs to pick up heavy metal preform parts which are orientated randomly in metallic cage box. This gripper is now in operations in a factory in northern Germany, most of our other further work is either confidential or classified.
That said, throughout 2024 and 2025, I am working hard with partners to design and build a suite of several computationally designed vehicles, be it ground-based drones, air-based drones or space-based systems, in order to contribute to more & better end-to-end benefit showcases for the industry at large. In Berlin, I will be able to discuss some of these ongoing preliminary works, particularly in carbon fibre composite AM and high-precision ceramic AM applications, so let’s keep the suspense until we meet in-person.
Addressing Uncertainty in Design for Extreme Environments
In conversations with engineers at organizations such as NASA, a recurring challenge is the difficulty in anticipating all potential boundary conditions and stresses their designs might encounter, especially in extreme environments like space. How does Rafinex’s stochastic optimization approach help engineers design for such uncertainties?
There are generally two sources of uncertainty sources that we regularly encounter at Rafinex. The first source is, like you describe, uncertainty stemming from real-life variabilities.
For this type of uncertainty, uncertainty can generally be approximated even if the uncertainty spread is quite large. We put several probability distributions at the disposal of the users to define their load condition variabilities in terms of likelihoods. However, we see the biggest changes in optimized design outputs when going from deterministic to stochastic inputs in the first place. This is when novel formally robust design appear. Increasing uncertainty levels beyond that typically only lead to incremental morphing changes of the overall robust macro-structure. In the validation stages however, confidence in your uncertainty ranges is required for a quantitative evaluation of the stochastic stress distributions.
Finally, it is of note that stochastic topology optimization can target several probability metrics of the target function (e.g. stiffness, strength, …) such as optimizing for the mean of the functional, but should additional safety be required metrics such as conditional value at risk ensure that optimization targets the worst case scenarios first and foremost.
The second type of uncertainty source we see often is actually stemming more from project management and requirements settings phase.
Often during the first phase of a project, when optimization can have highest impact at lowest effort, load conditions are sparingly or sparsely known/defined.
This is also something that stochastic topology optimization can overcome – Ultimately, the technology is how to perform optimization under uncertainty, regardless of what the source of the latter is.
Building Robustness into Engineering Designs
There is a tendency for optimized supply chains to sacrifice robustness for speed, resulting in fragile networks. How can engineers leverage Rafinex’s tools to avoid similar pitfalls in product design and manufacturing processes?
There is indeed a massive challenge in recent economic years with supply chain disruptions, macro-economic changes and raw material price pressures that have led people to reconsider near-shoring or re-shoring manufacturing supply chains more locally. For these type of applications, we have demonstrated that formally robust designs are in fact more “forgiving” against raw material property fluctuations. This in turn, opens the possibilities for design owners to i) consider a wider range of material sources and manufacturing suppliers and/or ii) consider more sustainable material sources which often come at the price of increased property variabilities.
Regarding the trend for near-instant simulation results in design itself, we at Rafinex put quality, safety and reliability above computational time.
Our customer base is looking for safety-critical design optimization and thus compute times ranging reasonable from a few minutes to a few hours are entirely acceptable, given the engineering design warrants this time.
Particularly with cloud-first algorithms, which work asynchronously with respect to the engineer’s time, we can parallelize many optimizations at once using server scalabilities.
Integrating Experienced Engineers’ Expertise into Optimization Tools
The acceleration of education around optimization tools is crucial, but so is capturing the tacit knowledge of experienced engineers. How does Rafinex propose to embed this expertise within software workflows to enhance design outcomes?
Education and user know-how is a crucial ingredient to obtaining a successful computational design, and this expertise may well be the Achilles heel of our entire community, today and in the future, particularly with a lot of experience leaving the workflow momentarily.
Hence, capturing know-how and encoding it, is on everybody’s mind.
The approaches to this challenges vary however. From what we see at Rafinex and its partners, a total AI-blackboxing magic approach does not work.
Even if it works from a purely technological point-of-view, if engineers have no insight into how a design was generated, nor how it functions in real-life, then we have no trust nor safety margin. Such a situation is asking for a disaster to happen, and as far as Rafinex’ customers in aerospace, defence and automotive are concerned, that is a non-starter situation.
Hence, it is not surprising to say that I have my reservations about entirely black-boxing know-how into a magical AI box. That said, I am not dogmatically opposed to new technologies either, so what I can see is an AI repertoire in conjuncture with a library of vetted boundary condition templates, from which proposals / templates of boundary conditions are made from by an AI, but always keeping the human-in-the-loop. In this case we are speaking about effectively advanced template retrieval and workflow acceleration / work speed ergonomics of the human user in question. The latter does need to have a know-how in FEA and optimization though regardless.
Key Takeaways for CDFAM Berlin Audience
What are the primary insights and takeaways you hope attendees will gain from your presentation at CDFAM Berlin?
Personally, I aim to convey the value & impact that good computational design can have for companies.
With 70-80% of a product’s lifetime costs and environmental impact being locked in during the design phase, I will want to highlight the that cost of Good Design is minuscule compared to the Cost of Bad design.
Furthermore, I will want to contribute to clarity and transparency of today’s technologies capabilities, strengths as well as limitation.
Being truthful with what can and, equally important, what cannot be achieved as well as clearing up hyper-inflationary terminology / taxonomy is in my opinion to the benefit of the entire computational design community.
If we are not careful, marketing-driven terminology and promises will only lead to mismatches in expectations, bad designs and potentially engineering failures – thus risking to send the computational design field into a “winter period” of its own – I will absolutely want to avoid the cyclical winters that our wider AI field is regularly subject to because their hype got out of hand (again).
Objectives for Participation in CDFAM Berlin
Beyond sharing your expertise, what do you aim to learn or whom do you hope to connect with during the symposium?
CDFAM is an outstanding, dynamic and high-quality community focused around design, optimization and manufacturing. I hope to connect with as many computational design enthusiast and industry leader as possible.
My overall aim is to contribute to the community as much as possible and work together with everyone to advance the adoption of computational design in mainstream industries, for the benefit of everyone in the community. To this effect, if you are interested in discussing research collaborations as well as demonstrator design showcase endeavors, please do not hesitate to talk to me.
Register to attend CDFAM to connect with Ańdre and other experts in topology optimisation, simulation driven and computational design.
