Could generative design signal the death of the engineer?

It was engineers from Sony Computer Science Laboratories in France who developed a computer called DeepBach that could learn and replicate the style of JS Bach’s chorales. These arrangements consist of four vocal parts singing in harmony and DeepBach was able to generate new pieces that so closely mimicked Bach’s style that a test group including professional musicians were in many cases not able to distinguish Bach’s work from the computer-generated impostors.

In a similar vein, researchers at the University of Warwick recently created an algorithm that can rate the beauty of scenic images on a scale from one to ten. It appears beauty is no longer in the eye of the beholder – it has a digitally determined place on a sliding scale. And should Bach still be alive and depending on his musical prowess for an honest income he may be concerned that a digital usurper would do him out of a job.

These developments raise an uncomfortable question – are similarly gifted computers now obviating the need for engineers and designers?

It sounds somewhat fatalistic and not all that realistic, but there can be no doubt that computer aided design software, in its broadest sense, has come on in leaps and bounds. A roomful of draughtspeople, each labouring over their own part, have now been replaced by a handful of CAD engineers often with a more model-wide perspective of the project.

But the shift from drawing board to CAD workstation was initially about digitising the process and then adding all the features for speed and convenience that the computer environment facilitated. It was still the engineer who was doing the engineering.

Topology optimisation, first introduced a couple of decades ago, is based on finite element analysis (FEA). It essentially involves defining the envelope for a design – basic dimensions, material, key loading criteria – and producing an optimised model. As it has been around for some time, the abilities have become more advanced. There are more complex assemblies, multiple loading attributes and different manufacturing methods for example.

GRM is a company that is both a provider of optimisation software and also a user in its role as an engineering consultancy. Managing director Martin Gambling says topology optimisation is an important part of what the firm does, but it is not always the answer: “The key learning we aim to teach our consulting engineers is to know when to apply optimisation and when not to. We use topology and other optimisation methods at the start and also through a development programme. Specifically we use it to guide us on initial design concepts, provide insights into ways to improve a design that isn’t meeting performance, and remove the labour-intensive iteration steps of an FEA analyst’s work.”

So, given that these tools are well known and used in aerospace and automotive industries, why doesn’t everyone use them? Partly it is because optimisation is ideal for minimising structures – taking out the weight. The amount of fuel saved over a lifetime just by taking a few grams out of an aircraft or car is considerable, while it is not such an issue in most other industries. However, Gambling points his finger elsewhere: “Awareness is probably the main reason.”

Awareness is about to change, albeit under a new name. Generative design, according to Gambling, is the same as topology optimisation but with a more designer-friendly name. He observes: “The new name is quite nice, as the word ‘optimisation’ often puts people off considering the technology. The only extension to the technology is that some of the CAD vendors are developing very powerful tools for converting the results from the optimisation into solid geometry that can then, in theory, be directly manufactured.”

This new technology is one that all the major CAD vendors are circling around and in some cases is being filtered into latest versions of their software. Paul Brown, global marketing lead for Siemens NX, believes the evolution is more dramatic: “Topology optimisation has been around for years. The thing is that the end result has been, at best, good advice for an engineer, but what tends to come out are some weird and wacky shapes.

“What we are now seeing is a breed of topology optimisation tools coming out which give you geometry you can start working with. The algorithms are coming out with much more logical and much more usable shapes.”

Brown concedes that Siemens’ first foray into this technology was not a resounding success, but it did start to show areas where material could be removed from a design.

Part of the problem was that all such models were made up of facets, little triangles, that were not always that useful when it came to establishing a base model to work on.

As Gambling says, Siemens is developing conversion modelling for its CAD packages. Brown says: “Conversion modelling allows you to start working with this facet data alongside regular proper solid data, the mathematically accurate surfaces in CAD. Conversion modelling melds those two together and we have embedded that in both NX and Solid Edge to try and start moving towards this whole generative design environment.”

Such software allows intervention at any stage of the design process, producing potentially thousands of optimised designs. Brown continued: “You could decide, I like this one, but this other one may meet the criteria better, but it looks ugly or it looks stupid – there’s a whole balancing act.

“We are not at the fully autonomous designing stage yet, luckily, otherwise the world would get a scary place. But we are at a point where we can help the designer with getting lots of different iterations. They can then decide which one to take forward and start building on with regular CAD modelling, to expand on what’s been given back by the system.”

Another of the main players in the CAD world, Autodesk, has been developing its own solution, Dreamcatcher. This is not a product in its own right; it is an application that is being introduced into the company’s design to additive manufacturing solution, NetFabb, and will appear in its CAD environments Fusion 360, Inventor and AutoCAD as time goes on.

Benjamin Schrauwen, director, additive products and manufacturing platform at Autodesk, believes: “This is a paradigm shift in engineering design in the same way as CAD was when it was first introduced.” He claims that topology optimisation has not taken off in general product design because you needed to be an aeronautical engineer or automotive engineer to understand the techniques – they are just too complicated.

“You need to be an expert in meshing, complexities of using a state-of-the-art simulation to use topology. But now generative design is a divergence of the design paradigm – it helps engineers explore design alternatives that they wouldn’t do by themselves.”

Generative design is currently only really appropriate for single components. Further down the line it will be able to do more complicated components and entire assemblies. Could it ever design a complete engine or car? “That really is too far down the line to predict,” says Schrauwen, “but I can certainly see it being used in shaping the chassis and optimising body parts.”

The key thing here is that it will fit into the linear design workflow of the typical engineer – not replacing the engineering function, just speeding it up. Schrauwen adds: “The whole design process could be shorter and more efficient, which can be a real driver in today’s quest to minimise time to market.”

There are a couple of factors that come under the category of ‘enabling technology’ as far as generative design is concerned. First is cloud computing. With its process of simulating and iterating, the computer processing power required for generative design is going to be beyond the average workstation. Using the unlimited resources of the cloud gives open access to generative design, even if further CAD work is taken out of the cloud and back down to earth.

Secondly, the rapid developments in additive manufacturing techniques and materials have made it the perfect production partner for generative design. Previously the problem with the models that came out of the topology optimisation process was that they may have been ideal load-bearing parts from an engineering perspective, but that is of little use if they can’t be manufactured. The Gaudi-esque creations were often difficult to manufacture using traditional machine or moulding processes.

Such limitations don’t apply with additive manufacturing – it can make just about anything, no matter how weird and wonderful the shapes are. As long as production runs are not going to be prohibitively long, as additive manufacturing is still a relatively slow process, and it is not intended to migrate manufacturing to another process, additive has given product designers more flexibility to move away from the restrictions imposed by the manufacturing method.

And additive manufacturing is not staying static itself. The leading vendors are developing machines that can print, in a single run, a model with numerous colours and material properties. HP has even demonstrated how it could use its Multi Jet Fusion platform to print chain links that have embedded strain gauges.

“If you could print parts which have multiple properties, multiple behaviours, that opens up a whole new world for design,” said Brown. “In fact, it opens up a whole new world of being able to do electromechanical parts in a different way, in which case generative could cross all the disciplines. I think as technology emerges in some of the supporting technologies, like additive, that’s going to open up even more capabilities around generative design and people are going to start exploring new avenues for using it.”

If the job of the software is to optimise, does that then mean that all designs that have used generative design software start to look the same – the vanilla effect?

“Yes I think to start with that could be the case,” says Schrauwen. “It was funny when we introduced new architectural capabilities into AutoCAD; you could tell when certain buildings were designed by some of their features. I think when people can use generative design to do a complete design then there is probability that there will be similarities for a while, but that will change when designers become more familiar with the software and certainly in more complex designs – particularly when the aesthetics of the product are important.”

Brown concurs about where the human will remain the creative key: “Anything that’s got to be aesthetic, because how do you teach something to appreciate aesthetics? We have enough problems teaching designers what is aesthetically pleasing and what isn’t.”

The team at the University of Warwick is addressing this with its scenery appreciation algorithm, but it is a more difficult job translating attractive design to suit all products. Brown gave an insight into one of the directions that generative design development is heading. “I think every vendor that is looking at generative design is currently investigating stuff around machine learning.  I’m not going to tell you that we have solutions in NX or Solid Edge now for machine learning, but it’s part of our vision.

“As we do designs and you come up with one that you believe is a good idea, then the system should understand the decision process you’ve been through, learn from that and actually apply it. We have a number of research projects under way into this whole topic of machine learning. It’s moving the analyses away from just geometry and topology, it also involves topics like costs, materials selection, aesthetics, but also it learns and it gets smarter as we go forward.”

There are some increasingly capable early-stage design packages being introduced to the market, made more beneficial by the instinctive nature of touchscreens. Ideas can be sketched on a tablet and these concepts can be turned into workable CAD models for the engineer to develop. But what if generative design was used to pick up the model created from the concept drawing? An idea then becomes a finished, optimised product without an engineer dirtying his or her hands on it. Is that realistic?

“Yes!” says Schrauwen. “But we are a long way from that and even then it would only be realistic for simpler products.”

Gambling has the further concern that younger engineers may rely on the tools without thinking about why a structure has evolved from generative design. A lack of understanding of engineering fundamentals could be papered over by using the capabilities of these new design tools. “We try to teach our engineers to think about the results that are presented to them. It is almost a responsibility of those developing these tools to educate that they help good engineers, not replace them, although that will be a difficult topic,” comments Gambling.

But despite these fears that generative design could replace some engineers’ knowledge, there seems little danger of it replacing engineers. Brown concludes: “Generative design is the next step in engineering, but is only as good as the constraints you give it. Engineers will always have a place – especially for complex and high precision parts.

“There is a long way to go before computer software can replace the skill set of the engineer, but generative design is a great tool to make that engineer’s life easier.”

Case study: application of generative design

Airbus cabin partition

Autodesk’s generative design was used by Airbus to design a cabin partition that could appear in its A320 aircraft from next year onwards.

(1) The design that was arrived at by using an algorithm based around mammal bone growth.

(2) Stress testing compared to existing partitions.

(3) The 3D-printed parts being cleaned up.

(4) The resulting partition, which is 45 per cent lighter than the existing panel.