Presentation Information
[GW9BIT7e-05]Causal Operator Learning for Marine Heatwave Prediction
*Dong Eun Lee (1. Chungnam National University (Korea))
Predicting Marine Heatwaves (MHWs) remains challenging for data-driven
approaches, because they are rare, often unprecedentedly intense,
persistent, and shaped by nonlinear upper-ocean processes. A
prediction model trained solely on observational realizations may not
capture the underlying physics unless explicitly constrained, limiting
its ability to anticipate extreme MHW conditions not represented in
the observational record. Here, using a Laplace Neural Operator (LNO)
surrogate of a numerical model, we show that training on physically
contrasting sensitivity experiments promotes generalization to
unobserved MHWs.
An LNO is trained to emulate the numerically integrated upper-ocean
temperature profile in a controllable virtual environment governed by
surface heat flux, wind, and precipitation forcing through nonlinear,
Richardson-number-dependent mixing. The optimal hyperparameter
configuration is selected through systematic grid search. In 32-day
hindcasts, the surrogate outperforms a damped-persistence baseline.
We then use controlled forcing experiments to test whether it learns
physical response operators rather than experiment-specific Sea
Surface Temperature (SST) patterns. Models trained separately on
control and modified-forcing simulations exhibit limited
cross-experiment generalization, indicating that conventional training
primarily captures forcing-dependent statistical patterns. In
contrast, a surrogate trained on a balanced combination of control and
weakened-wind simulations retains high skill in both environments. An
SST-distribution-matched dataset randomly sampled across sensitivity
experiments yields substantially lower predictive skill, despite
broader coverage of temperature outcomes.
These results show that surrogate training on physically controlled
simulations can promote learning of forcing-dependent oceanic response
sensitivities, providing a pathway toward predicting MHW extremes
beyond those observed historically.
approaches, because they are rare, often unprecedentedly intense,
persistent, and shaped by nonlinear upper-ocean processes. A
prediction model trained solely on observational realizations may not
capture the underlying physics unless explicitly constrained, limiting
its ability to anticipate extreme MHW conditions not represented in
the observational record. Here, using a Laplace Neural Operator (LNO)
surrogate of a numerical model, we show that training on physically
contrasting sensitivity experiments promotes generalization to
unobserved MHWs.
An LNO is trained to emulate the numerically integrated upper-ocean
temperature profile in a controllable virtual environment governed by
surface heat flux, wind, and precipitation forcing through nonlinear,
Richardson-number-dependent mixing. The optimal hyperparameter
configuration is selected through systematic grid search. In 32-day
hindcasts, the surrogate outperforms a damped-persistence baseline.
We then use controlled forcing experiments to test whether it learns
physical response operators rather than experiment-specific Sea
Surface Temperature (SST) patterns. Models trained separately on
control and modified-forcing simulations exhibit limited
cross-experiment generalization, indicating that conventional training
primarily captures forcing-dependent statistical patterns. In
contrast, a surrogate trained on a balanced combination of control and
weakened-wind simulations retains high skill in both environments. An
SST-distribution-matched dataset randomly sampled across sensitivity
experiments yields substantially lower predictive skill, despite
broader coverage of temperature outcomes.
These results show that surrogate training on physically controlled
simulations can promote learning of forcing-dependent oceanic response
sensitivities, providing a pathway toward predicting MHW extremes
beyond those observed historically.
Comment
To browse or post comments, you must log in.Log in