REVEALING UNCERTAINTY | FUSION
probabilistic models well suited to engineering problems with sparse data. A Gaussian process provides a mean prediction and an associated uncertainty that varies with the inputs. This allows engineers to identify regions where the model is well supported by data but also regions where predictions are extrapolative and therefore less reliable. The practical implication is a shift in how model outputs are used. “Instead of the computer telling you ‘A, B, or C,’ you might see that there is an 80% probability of B and a 20% probability of C,” Niedfeldt says. “That allows an engineer to decide whether more testing is required or whether the level of confidence is sufficient for the decision at hand.” This probabilistic framing aligns closely with existing engineering practice, where margins, safety factors, and confidence intervals are already widely used. However, Niedfeldt emphasises that digiLab tools are intended to support, not supplant, human expertise, saying “In the sectors we work in, machine learning is a tool to inform decision-making, not replace it.”
Embedding physical knowledge A recurring criticism of some AI approaches is that they lack a basis of real-world physical understanding. By embedding prior knowledge, constraints, and physical relationships into its models, digiLab explicitly addresses this issue, as Niedfeldt explains: “In a Bayesian way of thinking, you start with a prior. A scientist or engineer already knows something about the world. When they study materials, for example, they have expectations about physical properties and those constraints and parameters can be built into the model.” This approach contrasts with purely data-driven methods that attempt to infer relationships without reference to underlying physics. While such methods can perform well in data-rich environments, they are poorly suited to nuclear applications, where data may be limited, expensive to obtain, or entirely unavailable for new designs. “The majority of models don’t inherently understand
physics,” Niedfeldt observes. “That’s why we focus on physics-based, engineering, and scientific applications, where expert knowledge is essential.”
Sparse data and early-stage nuclear systems One of the most persistent challenges in nuclear engineering, particularly in emerging areas such as fusion and advanced reactors, is the lack of operational data. In some cases, no full-scale system has ever been built. Niedfeldt argues that this is precisely where uncertainty quantification offers the greatest value: “One advantage of our approach is that you can get started with sparse and limited data. If you only have three or four material samples, or you are just beginning to build a simulation, you can still use these methods to understand what you know and, importantly, what you don’t know.”
These methods are designed to operate across the entire facility lifecycle, from early-stage design and optimisation through to operation and eventual decommissioning. In the design phase, probabilistic models can be built around computationally expensive simulations. These models approximate the behaviour
of the full model while explicitly tracking uncertainty. “If I build a simulation of a reactor subsystem, I know it is wrong to some degree,” Niedfeldt explains, “the question is how wrong it is and where that matters.” She continues: “That allows you to decide where to invest your limited budget. You might find that uncertainty in one model has little impact on the overall outcome, while uncertainty in another is critical. That’s where additional testing or simulation can be focused.” This approach contrasts with traditional conservative design strategies, which often apply large safety margins uniformly. While conservative margins remain essential, digiLab argues that a more informed allocation of effort can reduce unnecessary conservatism without compromising safety. “If scientists are allowed to work indefinitely, they
will keep refining until they are extremely confident,” Niedfeldt notes. “But if we waited till we have perfect simulations we won’t have working reactors in the next 30 years. Engineering has always been about making informed decisions under uncertainty.” The nuclear sector’s embedded safety culture
places a premium on traceability and justification. Any analytical tool used to support design or operation must be justifiable to the relevant regulatory authorities. This requirement has historically limited the adoption of opaque algorithms. “The linchpin is trust,” Niedfeldt says. “It’s not just ‘what is the answer?’ but ‘how did you get that answer, and can I rely on it?’” Recent high-profile failures of poorly governed AI
systems in other sectors have reinforced these concerns. While such incidents may have limited consequences elsewhere, in nuclear engineering the stakes are far higher. “In nuclear, you’re signing off on decisions you’re liable for,” Niedfeldt observes. “You need to be confident in the output.” By producing explicit uncertainty estimates and maintaining clear links between data, model assumptions, and results, probabilistic approaches
www.neimagazine.com | February 2026 | 21
Above: digiLab tools are used to determine where additional simulations of plasma physics will provide the most information. Source: STEP Fusion
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53
