In tokamak fusion devices, a plasma disruption is a sudden loss of
plasma confinement and current. In the next-step fusion device ITER,
the resulting fast release of thermal and magnetic energy may lead to
severe heat loads and mechanical stresses. Crucial for successful
mitigation is prediction of an upcoming disruption sufficiently ahead.
However, the mechanisms of a disruption are complex and current
physical understanding is not sufficiently advanced to reliably predict
it. In this regard, “data-driven” techniques are being studied to
classify the plasma state into disruptive or healthy state. Due to the
diversity of disruption causes and differences from one device to
another, these classification models require an unpractical training
phase and yet do not generalize well, leading to high rates of false
positives and false negatives. To address these issues, I will focus
my research on developing machine learning techniques such as
“anomaly detection” to predict disruptions in real time, based on
measurable quantities with a clear physical interpretation and
normalized to guarantee machine independence. This will ensure
transfer-ability of the system to an unseen device (e.g., ITER). By
combining physical insight with state-of-the-art anomaly detection, a
novel, robust disruption predictor is envisaged that is able to deal
with measurement error and uncertainties in physical models of
today, with the required universality for the fusion devices of
tomorrow.