How Can We Rely on AI in Fusion?

The challenge now is the transition from research systems to commercial use. Along the way, fusion technologies are becoming progressively more complex, more integrated, and more reliant on advanced software and data-driven modelling.

Artificial intelligence is already part of that picture. It is being used to interpret large volumes of data, anticipate behaviour, and support control in environments unlike anything we have built before.

But using AI in practice raises an awkward question: not just how well it works, but whether it can be trusted.

From Performance to Trust

In many fields, Artificial Intelligence is judged on performance: does it make accurate predictions, improve outcomes, or outperform existing approaches? But through our work embedding AI into real-world fusion systems, we have found that this is not enough.

A model may perform well under test conditions, yet behave unpredictably when deployed, encountering situations it has never seen before, or operating on data that is incomplete, noisy, or unstable.

It is no longer, therefore, a question of whether a system works well on paper, but whether it is justified to rely on it in operation.

Making Trust Measurable

To make that judgement concrete, we translate trust into three ongoing operational questions:

Is the system still fit for purpose? How confident are we in its outputs? And what should happen when that confidence begins to change?

Making confidence explicit allows us to understand when a system is operating within its expected range, and when it is not. This measure of uncertainty becomes a critical signal for decision-makers, prompting closer monitoring, human intervention, or a shift to more conservative control strategies.

This is what trust looks like in practice. Not a fixed checklist or a single validation step, but a way of establishing and maintaining confidence in a system over time, and understanding what happens when that confidence changes.

At digiLab, this focus on uncertainty sits at the core of our work. Through our collaboration with industry partners, we are applying these ideas in real fusion systems.

A Shared Challenge 

These questions are not unique to fusion.

In aerospace, an aircraft control system must perform reliably in conditions it has never physically experienced, from extreme weather to rare failures that can only be simulated. In defence, systems are constantly under attack and subject to manipulation and deception, where the true state of the world must be inferred rather than observed directly. In nuclear applications, even minor changes to a system must be painstakingly controlled and justified, because small errors can compound into catastrophic consequences.

In these settings, uncertainty is not something that can be removed, but something that must be understood, tracked, and worked with.

Trust is built not by assuming systems are correct, but by making their limits explicit: understanding where they are reliable, where they are not, and ensuring that those boundaries are respected in operation.

Why Fusion Is Different

Fusion brings many of these challenges together, but in a more extreme form.

Fundamentally, there is no stable data foundation to build on. It is not just the engineering that is new, but the physics itself, both co-evolving with the AI used to interpret and control it. As fast as our understanding improves, the goalposts move.

Also, you can’t just stick a thermometer in plasma. Merely observing the system is difficult, so we must often infer its state from indirect signals. It is not always clear what “good” data looks like, or how best to use it.

In this context, uncertainty runs all the way through the system - data, model, governance and everything in between. Inference under uncertainty is not a workaround, it is the best possible approach.

This stands in contrast to more mature industries, where trust is built on stable physics, long operational history, and established practice. In fusion, that foundation does not yet exist.

Trust must therefore be earned differently, built on evidence, qualification, and a clear understanding of uncertainty.

Humans in the Loop

Fusion is moving from research into real-world application, and Artificial Intelligence is helping to accelerate that transition. But raw capability is not enough. For AI to play a meaningful role in fusion, it must be trusted by operators.

That trust is not built on performance alone, but on a clear understanding of uncertainty and the ability to act on it. In fusion, that is not a limitation; it is the foundation for turning innovative experiments into a reliable, long-term energy source.