Traditionally, the field of product design and engineering has found itself behind other industries in incorporating artificial intelligence (AI) and machine learning (ML) due to the sheer volume and nature of the data required.
For industries like banking and retail, the data required is often two-dimensional and lightweight. Moreover, due to the regulations in place required for audit, these industries already have a wealth of historical data needed that ML tools can readily tap into.
For industries like architecture and engineering however, the design of three-dimensional products often requires highly complex 3D CAD models, Finite Element Analysis (FEA) meshes consisting of thousands or millions of elements, high-fidelity simulation of multiple physics, and optimization runs exploring multiple variants of a design. This all adds up to a huge amount of data and more often than not, businesses get stuck on how to utilize, share, and save these huge volumes of data.
Recent advancements in the fields of AI and ML, combined with the increased availability of robust simulation, testing, and field data sets has made engineering data science a critical component of the modern product development lifecycle.
There are already tools on the market today that blend modern AI and ML driven predictive analytics with traditional engineering capabilities. By making both current and historic simulation and analysis data directly available to engineers for design improvements in the early stages, these tools let engineering companies bring new product ideas to market faster, more efficiently, and at lower cost.
To extract maximum value from these exciting tools, however, companies need a plan to store, manage and utilize their data efficiently. In short, they need data discipline.
Get ready your data
Due to advancing sensor and bandwidth technology, increased storage capacity and computation power, manufacturers have begun to use ML for product design. Continuous method development and access to open-source codes have opened the doors for companies to experiment with ML and its capabilities.
However, some challenges remain:
- Data on demand: In cases where manufacturers do not have enough data to run an ML model, they often produce artificial data sets through physics-based simulation. This method is often time and resource intensive.
- Data in hand: Often data has been collected over a long period of time − sometimes decades − and stored in different places and formats created by different software versions. Reliability may be a concern. Additionally, this data needs to be converted to metadata for it to be used by an ML model.
- Data in flight: IoT sensors are typically responsible for collecting and producing large amounts of fast data from operations such as live telemetry data from F1 cars during a race. The volume and quality of data can be an issue in implementing an effective ML model.
Augmented simulation in product design
To speed up design convergence and achieve greater product innovation, machine learning algorithms can be used to augment physics-based simulation.
In ML-based AI-powered design, engineers can leverage current data without needing to make assumptions on the physics and material properties of products, enabling them to make wide design explorations.
On the other hand, in a physics-based simulation-driven environment, engineers start from what they already know about physics and material properties, to make assumptions to simplify the process. But traditionally, to achieve product designs with high accuracy and fidelity from this method, engineers need to provide very specific inputs to the modelling process. This is often difficult.
A solution emerges when manufacturers combine physics-based simulation with data-driven knowledge. Since there already exists a pool of data, we can train the neural nets and implement these trained neural nets into simulation tools.
The successful combination of both models will cost effectively accelerate engineering processes without compromising on the robustness of the products designed.
Rolls-Royce, for instance, is using the technology on vast amounts of historical product and in-service data from disparate sources, to carry out and optimise structural analysis and testing of aircraft engines.
By dramatically reducing the number of sensors to be deployed on prototype engines, the manufacturer can potentially save millions of US dollars in recurring costs. Rolls-Royce further expects to pioneer new AI-driven engineering use cases that can drive greater business value.
Closer to home, Serba Dinamik, a services and power generation systems provider to
the oil and gas industry, is using converged data analytics, AI and ML to develop more reliable turbines, and implement advanced predictive maintenance on its gas turbine generators. These generators are installed on offshore oil rigs to provide electricity to operations on the rig.
Improving turbine reliability and accurately predicting failures before they occur help Serba Dinamik customers like Shell, Exxon Mobil and Petronas save rig downtime costs that can run into millions of US dollars each day.
An exciting future awaits
The manufacturing industry is on the cusp of benefitting from the power that comes from ML-based, AI-powered design combined with physics-based, simulation-driven design leveraging the latest in high-performance computing.
At the same time, predictive data analytics techniques long associated with business-centric data are being aggressively deployed on asset-centric data to enhance manufacturing, warranty, and testing performance.
Truly, manufacturers today have unparalleled opportunity to harness technology to become globally competitive.
