• Richard Buckley

Meet my Digital Twin...

Amongst the hype of all things digital, an eerie apparition is being discussed: the digital twin! But what exactly is a digital twin, how do you create one, and more importantly why would you want one?

Imagery used by industry and media, characteristically drawn in blue wireframe, suggests a digital twin is the virtual avatar or doppelgänger of an entity in the physical world. The reality is much less spectacular – or sinister.

A digital twin is simply a mathematical representation of a physical system, process or characteristic. It is true that digital twins can take visual form if that is their primary purpose (e.g. in architecture), but in most industries, digital twins are invisible sets of algorithms and calculations, running on large computer clusters, and being fed data from the real world 24/7.

Imagine the simplest system: a wall switch and a lightbulb. If the switch is on, the light is on (and vice versa). This simple system’s digital twin could be an Excel spreadsheet describing the status of a wall switch and predicting the status of the light bulb. This system can be perfectly modelled and predicted. At some point in the future the bulb will fail, but we need more data to predict when. Let’s assume historic data tells us the failure can be predicted based on two factors: the number of times a bulb is switched on and off, and the total number of hours it has been on. The spreadsheet would thus need to be expanded to capture running hours and a count of the number of switch cycles.

Of course, the use of models to represent reality is hardly new, so you might ask what is added by going digital. After all, for years the automotive and aviation industries have built physical scale models, using wind tunnels to predict and improve aerodynamic characteristics of designs before full-scale construction. Ship builders have followed suit, creating scale models of vessels that are tested in dedicated tanks where wind and waves can be simulated. This is shipping’s ‘wooden twin’, providing an early prediction of the performance of the life-size vessel, but it has its limitations.

For automotive and aviation, the model and wind tunnel approach is highly accurate and scales quite successfully. For shipping, however, the many external factors that influence vessels in open seas make this much harder: sea currents; wind speed and direction; the angle, amplitude and frequency of waves; the draught of the vessel and depth below the keel; the variable wind resistance due to a ship’s cargo profile; the condition of the hull and propellor; and even the salinity and temperature of the water all affect the performance of a vessel. It is a chaotic system, and as such, scaling up from a model isn’t that reliable.

A few years ago, I spoke with the operator of one such ‘tank test’ facility in Southampton. He told me, “We build and measure from the scale model, but once the sea trails have given us real data, we apply a fudge factor to calibrate the models and make them reflect reality!”

Essentially, digital twins reduce the reliance on wooden models and start to use the past performance of the actual vessel as the basis for predictions. They then add their own, carefully calculated ‘fudge factors’ for the exact draft, power and weather of the vessel. Data science techniques can be used to capture, model and subsequently predict the behaviour of complex systems with multiple uncertainties like those involved in shipping. A combination of increased computer power at non-prohibitive costs and data now available from vessels through the Internet of Things, as well as better environmental data, means digital twins can now take modelling for shipping to the next level. Modern digital twins allow the principles of physics, sea trial models and the dynamics of real-world observed characteristics all to be combined to develop a more accurate model.

Most digital twins of modern vessels combine physical principles of how vessels behave, expressed as mathematical equations, and historic data of how this specific vessel has actually behaved in the past. These equations contain lots of vessel specific numbers that control how the equations behave. Machine learning is used to allow the system to learn how best to manipulate these equations so they most accurately reflect the way this vessel has behaved in the past, allowing them to then start predicting the future. The more historic data they are given to learn from, and the more accurate that data is, the better the system can learn how to predict the vessel’s performance. Some models then combine these techniques with deep-learning neural networks to further increase accuracy, but they tend to require even more data to train on.

So now you have a digital twin, what is it used for?

In most commoditised industries, the key to increasing profitability is lowering costs through economies of scale and optimising operations. Optimisation is a journey rather than a destination, however, and becomes increasingly harder once the lowest hanging fruit has been addressed. More complex optimisation requires businesses to consider multiple future scenarios and predict their outcomes to select the most optimal approach. This is the key role for digital twins – they can run, test and validate thousands of scenarios in parallel to provide choices – and answer key business “what if” questions.

Imagine evaluating a number of possible routes for a vessel crossing the Atlantic, each involving different external conditions. It would not be feasible or economical to keep running tank tests for every possible scenario, and applying simple speed consumption curves to each route would not capture the effects of weather or Emission Control Areas on fuel consumption. While weather routing is a useful service, not all vessels respond the same way to winds or wave, so you might be missing a possible better solution. What if instead you could evaluate each route taking speed, weather and local restrictions into account for just this vessel? Or even try different speed profiles for each route to find the most cost and fuel efficient way to reach your destination at the right time?

A real example of the power of digital twins comes from a visit I made to the motor racing team McLaren. They built a digital twin to augment the scale models and wind tunnels used by the engineering team and add realism to their driver’s simulator. With initial versions of the digital twin, drivers felt the simulator was a poor proxy of reality, but the management asked them to persevere and provide feedback to the developers. At the same time, the team used real-life measurements from the cars and racetracks to improve the digital twin continuously.

Over time, the quality of the simulation improved dramatically. The drivers loved it! They could re-run qualification races, making incremental (digital) changes to the configuration of the cars and testing the impact. McLaren’s digital twins are now so accurate that all their cars are designed, evaluated, iterated and even tuned digitally before a single component of a car is built.

Today, digital twins are all around us and impact our daily lives, not just who wins at Monaco. They help us predict the weather, tell us the quickest route to work and even tell us when the spread of Covid might abate. The EU recently announced Destination Earth (DestinE), a proposal to combine 20,000 GPUs into a supercomputer with the aim of studying climate change using a digital twin of planet Earth!

So digital twins are far from being the stuff of science fiction. These mathematical models are changing the world for the better. They give a competitive advantage to every business that deals with complex systems and has the foresight to embrace them.