From simulation to prediction: the emerging role of AI in fire engineering

Machine learning could allow fire engineers to test more scenarios, explore designs earlier and optimise solutions faster. But in safety-critical engineering, capability alone is not enough – the industry needs to gain confidence in AI-powered solutions too, explains Mott MacDonald’s Tom Lewis

Fire engineers are increasingly being asked to help project teams answer complex design questions while designs are still changing and there remains an opportunity to influence the outcome. Detailed computational modelling can provide those answers, but the time required to develop and run simulations inevitably limits how many scenarios engineers can practically explore.

AI and machine learning could change that, potentially turning modelling from something primarily used to validate a design into a tool that can help inform it. That opportunity is particularly interesting in fire engineering because determining whether a design is safe often involves understanding the interaction between many different variables.

AI creates the potential to explore far more of those combinations and identify where the critical cases lie. But this is safety-critical engineering, so proving that the technology produces reliable and traceable results matters every bit as much as what it can do.

Computational fluid dynamics (CFD) is already an established tool for understanding how heat and smoke are likely to behave in buildings where prescriptive fire engineering guidance alone cannot provide the answer. For complex buildings, CFD might be used to establish whether occupants have sufficient available safe escape time before conditions become untenable. The difficulty is that detailed modelling is computationally expensive, with an individual simulation potentially taking up to a couple of weeks to complete.

That means engineers typically use their experience to determine which combinations of fire size, location, ventilation and other parameters are most likely to produce the critical scenarios. Machine learning offers a different approach because it could allow us to interrogate a much broader range of possibilities from the simulations we undertake. Rather than replacing engineering judgement, its immediate value may be in giving that judgement a much wider evidence base.

This potential has recently been explored through collaborative work with a safety-critical government organisation to connect machine learning models with Fire Dynamics Simulator. The technology allows engineers to automate simulations within conditions they define, then use the resulting data to make predictions across a much broader range. Engineers still determine those safety-critical conditions, which is an important distinction when considering where responsibility for the resulting design decisions sits.

As an example of how this might work in practice, the model might initially consider fires between 1MW and 5MW and use those results to understand the relationship between increasing fire size and conditions within the building. Machine learning can then help predict what might happen beyond the cases already simulated, without requiring another complete simulation for every question that needs to be asked. Those predictions include uncertainty measures, which means engineers can assess their confidence in an outcome and determine where further detailed modelling is necessary.

The Mott MacDonald team tested this approach using a model for a 20,000-seat public arena, where fire engineering must account for large numbers of people using seating areas, concourses and concessions. The conventional modelling approach could demonstrate whether a particular concession location and fire size allowed occupants sufficient time to escape safely. Running enough detailed simulations to investigate numerous alternative locations, layouts and fire loads, however, could rapidly become impractical during a live design process.

The trial of the machine learning approach enabled those possibilities to be examined more quickly, identifying additional locations where concessions could safely be positioned. Predictive modelling could then be used to understand the areas between the individual simulations and map them across areas that had not been subject to direct simulations. Instead of identifying several points where modelling demonstrated a concession could be located, the approach could indicate a broader zone within which it could potentially operate safely. That creates useful flexibility, not only during design, but also potentially when spaces are changed or repurposed during the building’s operational life.

The same predictive capability also makes the modelling more adaptable when a client’s requirements or the design changes. If, for example, additional concessions or greater fire loads are proposed after the initial simulations have been completed, engineers can use the existing results to assess the likely effect without immediately re-running every scenario. This provides a quicker indication of what may be possible, together with the level of uncertainty involved, before targeted simulations are undertaken to validate the revised design.

The opportunity becomes even more interesting when optimisation is added because many fire engineering problems involve multiple competing design parameters rather than one question with a simple answer. A scenario where ventilation equipment within the arena could no longer be installed in its originally proposed location was also examined using the technique. By that stage, other elements of the design were already established, so the available options for moving the equipment were constrained.

The optimisation process tests different combinations within those constraints to seek a solution that maintains the required safe escape conditions. In the arena scenario, this allowed the engineer to understand relatively quickly whether the proposed change appeared viable before committing significant computational resources to fully validating it. Crucially, that does not remove the need for validation, but it could make the process of identifying the solution that needs validating considerably more efficient.

There is also a potentially important consequence for when fire modelling becomes valuable within the wider design process. Detailed CFD analysis traditionally requires sufficient design certainty to justify the investment, because there is little value in spending considerable time developing a model if fundamental aspects subsequently change. Being able to examine more scenarios efficiently could move some modelling earlier, giving project teams an indication of what appears possible while they still have greater freedom to respond.

Later in the design, the emphasis could instead move towards optimisation as the available choices narrow and different engineering requirements need to be reconciled. This could make fire modelling more useful at both ends of the process: initially helping teams understand the boundaries of the problem, then supporting more detailed decisions as the design develops. In both cases, the advantage lies in exploring more possibilities rather than asking AI to make the engineering decision.

That distinction between AI capability and confidence in the AI-powered solution becomes particularly important when considering what needs to happen before these approaches become routine. As with many AI solutions, capability and confidence in its use in fire engineering are developing at different rates. Any AI-supported analysis must be referenceable, factually correct and based on appropriate information, with engineers able to understand the assumptions and limitations behind its output. In a discipline governed by codes, standards and safety requirements, an answer that looks convincing but cannot be substantiated is of little practical value.

The same applies when machine learning is combined with computational modelling, where predicted outcomes need clearly understood uncertainty and appropriate validation against conventional simulations. Someone still has to take professional responsibility for determining that the resulting design is safe and that responsibility cannot simply be transferred to an algorithm. Human expertise, checking and an auditable engineering process therefore remain fundamental to how these technologies should currently be used.

In the longer term, the development of parametric models could take this much further by using sufficiently large datasets of previous simulations to predict outcomes without running any CFD simulations. Such an approach could allow non-CFD specialists to interrogate design options much more rapidly and potentially make sophisticated modelling useful across a much wider range of projects. It is an exciting direction, but those predictions would still need rigorous validation before they could be relied upon for safety-critical decisions.

Exactly how long it will take for confidence to catch up with capability is difficult to predict, particularly given the speed at which AI technologies themselves continue to develop. However, what is certain is that AI and the potential to use parametric modelling will open up broader opportunities for building design in the longer term.

Tom Lewis is technical principal for fire engineering at Mott MacDonald.