Combined dataset of observations and neural network predictions of downward vertical turbulent heat flux at 0°N, 140°W used in "Modulation of global-scale warming of the atmosphere and subsurface ocean by diathermal heat transport in the equatorial Pacific's cold tongue"

This contains the composite observation-neural network predicted dataset of downward vertical turbulent heat flux (Jq) at 0N, 140W used in the following manuscript. Also included is the neural network used to make predictions from state variables, vertical gradients, and surface parameters. Please see the main manuscript for a detailed description, and cite the main manuscript if using the dataset or neural network:

Iyer, S. and J.N. Moum. Modulation of global-scale warming of the atmosphere and subsurface ocean by diathermal heat transport in the equatorial Pacific's cold tongue. Submitted to Journal of Physical Oceanography.

Contents:

"Jq_combined.nc": This contains the composite observational-neural network predicted dataset of Jq. Coverage is from May 1990 to March 2020 and 30 to 90 m depth with 4 hour temporal resolution and 5 m vertical resolution. The variable "Jq" contains time-depth gridded values of Jq (W m^{-2}; positive downward). The variable "Jq_flag" denotes whether data are from observations or neural network predictions. Observations are smoothed from higher-resolution data available on the NOAA/NWS NDBC website at "https://tao.ndbc.noaa.gov" and mirrored on the NOAA/PMEL GTMBA website "https://www.pmel.noaa.gov/tao/drupal/disdel/". Please cite those sources for use of the observational data. The authors strongly recommend that the neural network predictions only be used in the context of analyzing long-term variability (i.e., averaged over a minimum of 30 days, or across seasonal-annual timescales for greater confidence). This is because point-to-point predictability is weaker on shorter timescales and the neural network predictions include some peculiarities regarding the prediction of low values of Jq on these shorter timescales (please see the manuscript for more details).

"neural_network.nc": This contains the neural network specifications (weights, bias terms, mean/standard deviations of inputs used to normalize) used to calculate the predictions provided in "Jq_combined.nc".

"trained_neural_network.mat": This contains the trained neural network (weights, biases, other specifications) as output automatically during training by MATLAB. This can be used to make predictions of Jq using the same neural network with MATLAB with minimal additional coding. The specifications are also included in the netCDF file "neural_network.nc" so that predictions can be made with other software as desired. 

Contact:
Suneil Iyer
Oregon State University
iyersu@oregonstate.edu

This version: 
Version 1.2 (updated 13 March 2026)

All versions:
Version 1.2 - Minor edits to readme. Data files have not been changed and are identical to Version 1.0 and 1.1 (13 March 2026)
Version 1.1 - Minor edits to readme. Data files have not been changed and are identical to Version 1.0 (10 November 2025)
Version 1.0 - Original version (14 August 2025)

DOI:
10.5281/zenodo.19007876