Using Bayesian Networks as Metamodels for Predicting Uncertain Fusion Economics in Spherical Tokamaks

This study introduces a proof-of-concept methodology for utilizing Bayesian Networks to reason over uncertain fusion economics. Using Bayesian networks as a surrogate of a forward model facilitates bidirectional predictions because Bayesian networks do not distinguish between inputs and outputs. The network acts as a surrogate model to PROCESS systems code, enabling the user to make multidirectional predictions under uncertainty for both inputs and outputs in a faster response time when compared to using simulations. Model inputs are probability distributions of important fusion spherical tokamak (ST) parameters that impact economics, such as
$\beta$
, and outputs are probability distributions of cost parameters, such as power plant capital cost. An evaluation of the network’s efficacy in performing both forward and reverse inference underscores its ability to align with input ranges associated with both low and high capital costs. The results emphasize the paramount influence of optimizing physics and reactor geometries through parameters like
$\beta$
and
$A$
on cost reduction compared to engineering efficiencies, elevating the significance of physics parameters in fusion economics. Armed with this knowledge, fusion developers gain a probabilistic understanding of potential capital cost ranges within their uncertain design domains, with the potential to apply this methodology to other uncertain design spaces.