Deep learning enters the engineering design optimization workflow, reducing time-to-market and increasing product quality.
Neural Concept has sponsored this post.
In the spirit of ‘better and faster,’ deep learning is entering the engineering simulation and design optimization workflow to reduce time-to-market and increase product quality. This introduction of data-driven design parallels the introduction of simulation—decades ago—which shortcut development times by reducing the need for physical prototyping and testing.

Illustrating this trend, said Pierre Baqué, CEO of Neural Concept, are “Global players [who] have already adopted deep learning technology in engineering design, such as Bosch, Airbus and Safran.”
Baqué is an expert in the field of AI image recognition. When this technology started identifying stop signs on the road, Baqué wondered if it could be used to improve engineering workflows. After all, if deep computer vision could recognize the shape of a stop sign and associate it with a warning, why couldn’t it recognize the shape of a CAD model and associate it with a fluid flow?
“You just have to connect geometry files to simulation output and run it through one neural network that will learn from all this data,” said Baqué. “It’s more a question of having the right type of data and the right type of problem. It’s quite interesting to see. This technology is versatile and can adapt to many types of problems.”
He adds that once a neural network is trained, Neural Concept can successfully predict simulation outputs from CAD inputs, within milliseconds. Some of the scenarios they have been able to study include:
- Aerodynamics
- Hydrodynamics
- Internal flows
- External flows
- Crash
- Stress analysis
- Fatigue
- Electromagnetics
Baqué’s team sought to use neural networks to shortcut multidisciplinary design optimization, ease the workload on simulation experts and make it better and faster to optimize product designs.
Deep Learning in Multidisciplinary Design Optimization Frontloads High-Value CAE
The aim of Neural Concept is to frontload the design cycle with neural networks, that copy high-value CAE, to speed up and improve the design process.
“The benefit is that the designers can run simulations independently. Which means that they are saving a lot of time as they don’t have to go to the simulation engineer,” said Baqué. By shortcutting this process, he notes that companies will benefit from:
- Shorter design cycles
- Faster request for quotes
- Faster time to market
- Better optimized products
He added, “Instead of testing 10 different configurations, the designer has run scenarios on thousands of configurations taking into consideration different boundary conditions. In the end, this makes the product better.”

The quick simulation turnaround of data-driven design optimization enables innovators to test out more ideas without putting innovative theories on hold. When it takes days or weeks to hear back for results from traditional simulation tools, designers can lose their inspirational thoughts.
The concept of simulation in design isn’t new; simulation tools have been added to various CAD solutions. They promise many of the same workflow benefits as data-driven design optimization. Baqué explained that what sets them apart is the quality of their output.
“There are tools today that exist that are doing CAE for designers, directly inside the design tools. However, these tools are simplified and idealized solutions because they must run very easily and fast,” said Baqué. “They are not adapted or faithful at reproducing high-end simulations that have been validated by the simulation engineers.”
Conversely, he added, “The neural network can be trained to learn how to go directly from geometry CAD files to the post-processed simulation results. Which means we are completely skipping the pre-processing, geometry cleaning, meshing and simulation steps. And we have a tool that can run extremely fast.”
Make Multidisciplinary Design Optimization Available to a Wider Audience with Deep Learning
It’s rare for simulation-savvy companies to only perform one type of analysis during the design process. Typically, they utilize a complex simulation workflow that studies multiphysics and their interactions.

“It’s impossible to assume that we would have front-loaded CAE that will be able to do this,” said Baqué. “Simulation teams don’t always approve and trust the simplified simulation that is implemented in CAD tools. [With neural network technology] simulation teams are in charge of building and validating the model and training it from the high-quality CAE data they have generated. Therefore, they stay in control of the simulation chain used in production.”
Baqué explained that the machine learning application built by Neural Concept and a company’s simulation team can get arbitrarily close to the results from high-quality CAE and complex multiphysics workflows.
“We have two types of end users,” explained Baqué. “We have the experts and the designers. And they are using two types of interfaces. The expert interface is built for advanced simulation engineers or data scientists. So, they don’t need an expertise in machine learning or deep learning. They need more of a scientific mind-set and comfort with simulation. The expert platform is made for an engineer to be able to train and validate a neural network model.”
How to Use Deep Learning to Make Multidisciplinary Design Optimization Accessible at Scale
Baqué noted that the biggest benefits from data-driven design optimization appear when designers and engineers need to evaluate many similar designs throughout many design cycles.
“Since it’s machine learning, you need data to train your algorithm. It’s an investment, it’s only worth it to train an algorithm with data if you are going to use it many times,” Baqué said. “It’s useful if you’re designing a certain type of part, with many different variants, many different times over a year. One situation where we see a big use, is suppliers for the automotive industry, where there is a big challenge in efficient and fast customization for different customers.”
For instance, say you’re designing battery cases that will be supplied to twenty different companies. If each company has three to four products, that means you need to design 60 to 80 different cases and hundreds of iterations. Not only does this give you a large data supply to train a model, it’s also a perfect opportunity to utilize that model to accelerate the process.
“Another situation where machine learning is helpful is when you have one prototype that needs a lot of optimization,” added Baqué. “Then you do a lot of design cycles for different concepts and variants. Now you have a lot of data to generate the algorithm, so you can use it for the design optimization process.”
As for how to get a company started using machine learning technologies, Baqué said, “Typically, the way we work is that we start with an onboarding project where [we and] the customer experts… run a project in parallel … in order to build the first model that works, is precise and validated.”
The simulation team then simplifies and customizes the application to the workflow of the designers. The functionality can be embedded into a CAD engine, on the cloud or used as a stand-alone UI that reads in STL files. The engineers can also provide some safety features in the application in the form of a confidence index. This would tell the end user how much they should trust the given simulation.
Baqué added, “There is a phase where this trained knowledge and the application is deployed to the designers and end users of the company. There is back and forth between the designer and the experts to improve that application or refine the model. Then the designer can be completely independent working with this application.”
Deep Learning in Multidisciplinary Design Optimization Offers Consolidated Applications
“One of the key elements that needs to be understood is that we don’t provide pretrained models for specific applications,” said Baqué. “What we do is empower experts to prepare applications that will then be deployed to end users.”

So, this means that simulation technology isn’t going anywhere, even if data-driven design optimization becomes ubiquitous. Engineers will still need simulation results to get the data needed to train models and provide applications with consolidated workflows.
“Neural Concept is here to support the experts on the customer side, but the expert will become more and more independent,” said Baqué. “In my experience this goes very fast, a good simulation engineer after a few months will have a few apps deployed. If we compare to some efforts made by companies that do this from scratch, typically it takes them years before having something for one specific application that’s close to what we do. Then they need to start from scratch for the next one.”
How Deep Learning Improves Early Design Optimization, Customization and Iterations
Clearly, deep-learning based design optimization has its advantages over other front-end simulation technology. It opens the door to new optimization strategies that go beyond parametric representations.
Baqué explained, “In traditional optimization and simulation pipelines, the optimization is done using a parametric representation of the shape. Maybe five to ten parameters and the software finds the best configuration within this parametric space. It’s very limited as the engineer needs to define the parameters and the optimums will only be in that parametric space.”
That is why Neural Concept has leaned in on machine learning. The philosophy is that parameterized optimizations are only the first step. The neural algorithm can learn from them, build a predictive model and move beyond.
“In our solution there is no limitations with parameters. It can start in a parametric space and then extrapolate outside of that space by adding more explorations without loosing the knowledge gained before,” said Baqué. “The neural network can explore new shapes and designs that are more and more different from the initial ones. Then it’s up to the engineer to stop the exploration at some point, or to tune how much the optimization process should go outside the box.”
To learn more about deep-learning–led optimization, visit Neural Concept.