AI and simulation: a powerful combination for enhanced performance

To understand the central role played by simulation, consider an air defence system whose radar must detect and track airborne targets so that they can be intercepted. How can we be sure that it will perform effectively against threats — missiles, drones, fighter aircraft and others — approaching from any direction, travelling at different speeds and following very different flight dynamics?

The answer starts at the design stage. The equipment is exposed to as wide a range of simulated attack scenarios as possible in order to assess its performance and identify its limits.

But multiplying the number of scenarios and reproducing threats realistically requires considerable computing power. This is one area where AI can provide a decisive advantage.

“From the engineering phase onwards, we make architectural choices, so we need simulation to ensure that the system will deliver the expected level of performance,” explains Frédéric Barbaresco, Senior Expert in the AI for Sensors segment at Thales and co-author of the AI and Simulation position paper.

“At this stage, AI helps make attack scenarios more realistic. But its role does not end with simulation: it can also be used in operational conditions, particularly through hybrid digital twins, to help operators determine the most appropriate radar operating modes". Based on information about an identified threat — the targets involved and the jamming environment, for example — the radar can develop surveillance strategies and associated waveforms to maximise performance. AI therefore makes it possible to exploit the sensor’s many degrees of freedom more effectively and turn them into a tangible operational advantage.

More realistic training, more precise insights

These improvements are not limited to engineering and manufacturing. AI can also enhance the training of the people who will operate the equipment. Whether in a cockpit simulator or during a command exercise, the effectiveness of training depends on how realistic the simulation is.

“If you are building a tank simulator, you will create attack scenarios involving intelligent enemy tanks,” says Frédéric Barbaresco. “We can use artificial intelligence to control the attacker so that it behaves more realistically, reducing the reality gap⁽¹⁾ in our simulations. AI also enables us to perform certain tasks in real time that were previously impossible, because it can achieve the same result with less computing power.”

^{(1) The “reality gap” refers to the difference between a simulated environment and the real world.}

AI can improve not only the quality of training, but also the feedback generated as a result. This is the case with GLADIATOR, a system developed by Thales that allows armed forces to train in a real-world environment using real equipment, but where weapons fire and its effects are simulated. The huge volumes of data generated by GLADIATOR during training exercises are simply too large for instructors to process manually.

Thales has therefore developed an AI-enhanced debriefing tool capable of automatically analysing this data and identifying key indicators during the exercise, enabling instructors to provide trainees with more valuable and precise feedback.

The same principle is used in the Combat Staff Trainer, a system designed for command training. Here, what is known as “behavioural AI” controls all the units (friendly, hostile or neutral) that are not directly operated by the participants, reproducing their tactical behaviour and their responses to events. It can even use language to communicate with users, either verbally or in writing.

Simulation as a training ground for AI

As detailed in the AI and Simulation position paper, while artificial intelligence brings significant benefits to simulation, the reverse is also true: simulated environments provide an invaluable training ground for algorithms, particularly for reinforcement learning techniques.

Using data generated through simulation, Thales teams can, for example, teach a drone to identify the best trajectories for intercepting another drone. The algorithm explores a range of strategies within a simulated environment. Flight tests are then used to verify that the capabilities acquired in simulation translate effectively into real-world conditions, taking account of wind, turbulence, drag and friction.

As Frédéric Barbaresco puts it: “We try to reproduce in real-world conditions what we have simulated, in order to assess its robustness against the different uncertainties it may encounter.”

Simulation also addresses another challenge: obtaining enough images to train object-recognition models. Some types of military equipment are rarely photographed or may not have been photographed at all. Thales therefore combines real and synthetic imagery to enrich its datasets.

“Whether we are looking at frigates, tanks or certain transport aircraft, we sometimes simply do not have enough image data,” explains Jérôme Bernier, an engineer at Thales Alenia Space. “In those cases, we use simulation to supplement the image datasets used to train AI models. Without the right data, you cannot build the right models.”

A virtual vehicle can, for example, be inserted into a satellite image while accurately reproducing the lighting and shadows of the original scene. Weather conditions in the areas of interest can also be incorporated. The objective is to ensure that these synthetic images are representative of what the sensors will subsequently observe in the real world.

From virtual prototypes to decision support

AI models are transforming the world of digital twins — virtual replicas of an object, a process or even an entire system, such as an engine, cockpit, radar or building. Synchronised with its physical counterpart, a digital twin is continuously fed with new real-time data. It can support decision-making and help anticipate potentially high-risk scenarios.

TwinFire, the digital twin of the SF500 naval radar, is one example. By integrating AI, Thales can validate the radar more rapidly, assess its performance when physical testing is complex or costly to organise, and verify upgrades while conducting non-regression testing.

Another striking example illustrates just how far AI is extending the possibilities of simulation. As part of the DesCartes programme, conducted with the CNRS in Singapore, Thales is using tools that simulate urban wind flows to make drone trajectories safer. In this case, the hybrid digital twin is an entire urban district.

Starting with a digital representation of the buildings, physics-based simulations are used to train an AI model. The model can then estimate turbulence between buildings in real time, making it possible to establish no-fly zones and adapt drone trajectories to maximise endurance — for example, by taking advantage of favourable winds.

Trust depends on bridging the gap to the real world

All these examples raise the same critical question for defence companies and their customers: if an artificial intelligence system performs well in simulation, will it perform equally well in real-world conditions?

To answer that question with confidence, Thales draws on its extensive experience in critical systems and its expertise in integrating AI into operational equipment.

Enhanced by artificial intelligence, simulation is becoming an essential tool for exploring possible scenarios, testing the limits of military systems and preparing them more effectively for the realities of the operational environment.