1 of 1

Parametric Sensitivities of Oceans Using Neural Surrogates�

Scientific Achievement

Developed neural network surrogates for an idealized wind-driven ocean model to investigate its parametric sensitivity. Focused on four parameters and their impact on the prognostic variables using reverse-mode auto-differentiation of the surrogate models.

Significance and Impact

Oceans are essential in mitigating climate change by transporting heat and carbon dioxide. Understanding the impact of parameters, such as salinity, is crucial for accurate forecasting of the ocean dynamics. Our work bypasses the expensive, non-differentiable first principle-based simulations and leverages the expressivity and differentiability of neural networks to calculate the parametric sensitivities.

Technical Approach

- A perturbed parameter ensemble of Simulating Ocean Mesoscale Activity (SOMA) generates data for surrogate training and evaluation.

- Fourier Neural Operator-based learners are trained to forecast one-step forward dynamics with varying ocean parameters.

- Adjoint sensitivities of the parameters are obtained via reverse-mode auto-differentiation of the trained surrogates.

PI(s)/Facility Lead(s): Sri Hari Krishna Narayanan

Collaborating Institutions: Argonne National Lab, Los Alamos National Lab, University of Texas, Austin.

ASCR Program: SciDAC Partnership

ASCR PM: Lali Chatterjee, Kalyan Perumalla

Publication(s): Sun, Y., et al, "Parametric Sensitivities of a Wind-driven Baroclinic Ocean using Neural Surrogates," in the Platform for Advanced Scientific Computing Conference, 2024, in press.

Fig 2: One-step forward predictions from the trained FNO for all five prognostic variables at a depth of approximately 43 m in the circular basin of SOMA and at timestep 15 (days).

Sun, Y., et al, "Parametric Sensitivities of a Wind-driven Baroclinic Ocean using Neural Surrogates," in the Platform for Advanced Scientific Computing Conference, 2024, in press.�

Fig 1: Fourier Neural Operator for single time-stepping forecast. The input consists of the prognostic variables at the current time and the physical model parameter. The output is the same variables at the next time step.

With ImPACTS Partnership (BES)

LOCAL POC Yixuan Sun, Sandeep Madireddy, Sri Hari Krishna Narayanan (ANL)

TALKING POINTS:

This work seeks to understand the sensitivities of ocean model output to its input parameters, such as salinity.
Instead of differentiating a hard-to-differentiate ocean model written in Fortran, this work created a neural network surrogate model in the PyTorch machine learning framework and differentiated it using PyTorch’s internal differentiation capabilities.
This approach was tested on an idealized wind-driven baroclinic ocean model by creating perturbed parameter ensemble datasets.
The data was used to train accurate neural surrogate models for single time-stepping forecasting of five prognostic variables with varying physical model parameters.
The differentiated neural surrogates computed the the adjoint sensitivity of the prognostic variable with respect to the physical model parameters.

METADATA:

Name of the associated awarded project: Improving Projections of AMOC and its Collapse Through advanced Simulations (ImPACTS)

PI name(s): Luke Van Roekel (LANL), Sarat Sreepathi (ORNL), Shinjae Yoo (BNL), Sri Hari Krishna Narayanan (ANL), Gokhan Danabasoglu (NCAR), Zhengyu Liu (OSU), Julie McClean (SIO), Patrick Heimbach (UT)

Name of the program manager: Lali Chatterjee, Kalyan Perumalla

CITATIONS:

Yixuan Sun, Elizabeth Cucuzzella, Steven Brus, Sri Hari Krishna Narayanan, Balu Nadiga, Luke Van Roekel, Jan Hückelheim, and Sandeep Madireddy. "Surrogate Neural Networks to Estimate Parametric Sensitivity of Ocean Models." In NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning, 2023. https://arxiv.org/pdf/2311.08421.pdf

Yixuan Sun, Elizabeth Cucuzzella, Steven Brus, Sri Hari Krishna Narayanan, Balasubramanya Nadiga, Luke Van Roekel, Jan Hückelheim, Patrick Heimbach and Sandeep Madireddy "Parametric Sensitivities of a Wind-driven Baroclinic Ocean using Neural Surrogates," in the Platform for Advanced Scientific Computing Conference, 2024, in press.

AWARDS:

N/A

REPRODUCIBILITY:

Repo not made public yet

BACKGROUND: