Description
We present a reduced-order, data-driven surrogate model
of MAST plasma evolution, targeting a compact performance
state xt = [Ip(t),Wn(t)], considering the internal plasma cur-
rent Ip, where neutron emission rates Wn serve as a global
reactivity proxy. Leveraging experimental data from MAST
campaigns 8 and 9, the model is conditioned on actuator sig-
nals, including poloidal field (PF) coil voltages, gas injection,
and neutral beam injection (NBI) power, with line-averaged
density treated as an exogenous input. To focus on trans-
port and confinement dynamics, the framework targets quasi-
stationary phases, ensuring that the learnt dynamics represent
stable plasma evolution rather than disruptive transients. The
dynamics are learnt by Selective Dense State Space Models
(SD-SSMs), which utilise an input-driven selector mechanism
to interpolate among a dictionary of locally linear transition
operators. We demonstrate that the temporal trajectory of
the selector weights provides a structured segmentation of the
discharge, identifying transitions between operational regimes
such as Ohmic ramp-up and NBI-heated phases. Through
eigenvalue analysis of the learnt transition operators, we ex-
tract characteristic system timescales directly from the data.
Moreover, we draw insights into the underlying physics by
leveraging this data-driven approach, which allows us to un-
cover patterns not immediately evident through traditional an-
alytical methods. This framework demonstrates how high ca-
pacity sequence modelling can be reconciled with transparent
system identification, offering a pathway toward data-driven
scenario planning and trajectory evaluation for spherical toka-
maks.