[[Audio 1]] [[DownloadsSidebar]]

For decades, digital technology has been quietly revolutionizing the world of engineering design. Three-dimensional digital models have supplanted drawings, and the development of simulation software has allowed engineers to replace many physical tests with faster, cheaper virtual ones. Engineering companies have invested hundreds of millions to make these computationally and data-intensive solutions more efficient.

As computers have become more powerful, engineering teams have been able to develop ever-more-detailed digital models that replicate more of a product’s characteristics and expected behaviors. Today, digital twins are changing the way products are designed, operated, and maintained in a host of fields, from industrial machines to medical devices.

Digitization has also allowed engineers to give computers a more active role in the engineering process. Generative design and related optimization approaches work by programming a computer to run hundreds or thousands of simulations, tweaking the design between each run until it finds the best solution it can. The resulting geometries can outperform the work of the most experienced human designers.

Not yet optimal

[[MostPopularArticles 5]]

Despite its clear potential, digital design optimization also has some significant limitations. Simulating the performance of structures or fluid flows relies upon differential-equation solvers that are computationally intensive, so design optimizations can only focus on a few simultaneous parameters. As a result, today’s systems only explore a small part of the design space and tend to deliver only incremental improvements.

For the same reasons, companies still rely heavily on human engineering expertise to select and optimize the right parameters for their projects. That’s difficult when talent is scarce, and it increases the risk that human bias will lead to suboptimal designs. Worse still, simulation systems have become so complex that companies increasingly rely on external providers to run their simulations, eroding in-house knowledge and competence.

In a world where competitive advantage depends upon both speed and innovation, the traditional simulation approach may no longer be enough. Engineering companies may need to think differently about their approach to design optimization and the technologies they use to achieve it.

A faster route

Today, some leading companies are exploring an alternative that promises to increase the speed and effectiveness of automated design optimization, with the possibility of extending its application to larger and more complex engineering problems. This new approach is based upon artificial intelligence models like those that have been the key to many other difficult computing problems, from image recognition to mastering the game of Go. We call these models deep learning surrogates (DLS).

DLS technology reduces computing complexity and increases speed dramatically, with deep learning models running several orders of magnitude faster than traditional physic simulations. This, in turn, translates into many competitive advantages for organizations. It cuts time to market by reducing engineering process time. It cuts costs, too, by reducing the complexity and intensity of the engineering effort required. And deep learning simulations also allow companies to explore a much wider set of parameters for product design, discovering new optimizations unknown to expert engineers and yielding better product performance.

DLS explained

At its outset, a DLS process looks a lot like other digital design optimization approaches. The engineering team defines the constraints and desired performance characteristics of the product, and the computer runs multiple conventional simulations on different design options. That’s where the approaches diverge, however.

As those initial simulations are run, they are used to train a neural network, which is set up to take the same inputs and attempts to replicate the outputs of the simulation system. When training is complete, this deep learning model will work just like the conventional simulation, but much, much faster. In real-world projects, deep learning simulation models can run orders of magnitude more quickly than their conventional counterparts (exhibit).

[[Exhibit 1]]

That speed increase is a game changer. Using DLS, the machine can explore much more of the design space than with conventional simulation, helping it find the best possible solution. And the efficiency of this process is further enhanced with advanced search algorithms derived from the work of the wider AI research community. Ultra-high speed also means that complexity is less of a barrier: engineering teams can tackle larger, more elaborate components and systems, and optimize simultaneously across multiple domains.

Using deep learning in product development has important implications for companies that go beyond their technology choices, however. It will require engineering departments to adapt their organizations and allocate resources in new ways. That might include the acquisition of new talent, especially data scientists, data engineers, and machine learning specialists, and access to new computing resources, either in-house or in the cloud.

DLS in the real world

Deep learning surrogates are already proving their worth in some notoriously challenging engineering environments. One company in the power-generation sector, for example, used the approach to optimize the design of large turbines for hydroelectric plants. These machines must be individually configured for the specific operating conditions of each installation, a process that can take thousands of engineering hours and up to a year to complete.

In a pilot project, the company’s engineering team partnered with external specialists to create a deep learning model that could simulate the performance of the four major components in its turbines. The model was designed to accept the desired operating point as an input, and consider a host of different constraints, from the acceptable mass and strength of each part to fluid-flow problems such as cavitation or pressure pulsation.

Developing the model took the project partners six months, but its impact on the design process was immediate and profound. Using DLS, the company reduced the engineering hours required to create a new turbine design by 50 percent and cut the end-to-end design process by a quarter. Better still, the approach generated turbines that were up to 0.4 percentage points more efficient than conventional designs. For a typical hydroelectric dam, that would translate into an additional 13.5 gigawatt-hours (GWh) of energy production every year, enough energy to meet the needs of more than three thousand homes.

In the wind energy sector, another major player has applied the approach to optimize the control algorithms for large wind turbines. In high winds, these turbines must adjust their output, or even shut down altogether, to prevent damage. But those control decisions have significant implications for the operator: too conservative and the turbine misses out on generating valuable power; too aggressive and reliability might be compromised.

The company was already using conventional simulation techniques to develop site-specific control rules for each turbine. Those simulations, however, were time-consuming and expensive, taking eight hours or more to evaluate a single turbine configuration at a single site. Replacing the simulation system with a deep learning model cut that time to less than a second. With the new system, the company can now help its clients to optimize the design of new wind farm projects, evaluating many more turbine configurations and site layouts to find the best possible solution. The DLS system produces meaningful increases in total output too. The company’s data suggest that applying its approach to just 10 percent of Europe’s planned wind energy investments over the rest of this decade would increase annual generation by more than 210 GWh, enough to power more than 50,000 homes.

The future is deep

The application of deep learning in product development is just getting started. After the success of early pilots, leading companies are now building DLS into their standard engineering processes for multiple product categories. In addition to extending the approach to more industries and problem domains, researchers are also working on integrating the approach into the design process in radical new ways. One potential approach is to abandon conventional simulation altogether. Instead, researchers have proposed that deep learning models could be trained using real-world data from existing products in the field. That could transform the way companies improve their products, with design and optimization tools that learn automatically from the performance of previous product generations.