Parametric roll is a resonance: when encounter waves modulate a ship's restoring force at twice the natural roll frequency, roll amplitude grows exponentially until nonlinear effects saturate it. The onset is abrupt, the amplitudes are extreme, and the event is rare enough that most operational data contains no examples.
The paper on learning parametric roll episodes from wave-vessel time series (arXiv: 2603.24431) trains a stacked LSTM surrogate on paired wave-elevation and vessel-motion data generated by high-fidelity URANS simulations. The model learns the causal functional mapping: given the wave history, predict the ship motion forward in time.
The critical test is whether the surrogate captures the statistical shift associated with parametric roll — not just the time series shape but the change in the roll probability density function. In the most severe sea state, the LSTM tracks the onset and growth of large-amplitude roll and reproduces the corresponding fat-tailed distribution. The model doesn't just predict the trajectory; it predicts the regime change.
The authors compare loss functions and find that standard mean squared error optimizes average accuracy at the expense of tail fidelity. Amplitude-weighted and relative-entropy-based losses trade some average accuracy for better reproduction of extreme events — the events that matter for operability and risk assessment.
The through-claim: the loss function encodes what you think matters. MSE treats all errors equally, which means rare extreme events contribute negligibly to the gradient. A risk-aware loss function must weight the tail, because the tail is the risk. The surrogate that's best on average is worst at the thing you actually need to predict.
2603.24431. Naval architecture / parametric roll / LSTM surrogate / risk assessment / loss function design.