Emily Osborne ^1* Yuanyuan Xu ^1,2 Madison Soden ^1,2 Jennifer McWhorter ¹ Leticia Barbero ^1,2 Rik Wanninkhof ¹

• ¹ Atlantic Oceanographic and Meteorological Laboratory (NOAA), Miami, United States
• ² Cooperative Institute for Marine and Atmospheric Studies, Rosenstiel School of Marine and Atmospheric Science, University of Miami, Miami, Florida, United States

Within the Gulf of Mexico (GOM), measurements of ocean pH have been limited across space and time. This has hindered our ability to robustly study and monitor anthropogenic CO2 storage, ocean acidification, and general carbon dynamics over this biogeochemically variable marginal sea. The 2021 GOM launch of five-sensor Biogeochemical-Argo (BGC-Argo) ocean profiling floats represented the entry of this ocean observing technology in this region. The resulting near real-time, freely available BGC-Argo observations have vastly increased the number of profile observations, including sensor pH profiles, available within the under-observed "open GOM" region (>2,000 m water column depth). To circumvent a set of uncertainties associated with the collected float pH data, four neural network algorithms trained with high quality GOM shipboard bottle data were generated to skillfully predict pH (total scale, in situ temperature and pressure) in the region. The GOM neural network pH (GOM-NNpH) algorithms were trained using CTD and bottle data (temperature, salinity, oxygen, nitrate, pressure, and location) collected at deep (>1,000 m water depth) stations during three of NOAA's Gulf of Mexico Ecosystems and Carbon Cruises (GOMECC;2012, 2017, 2021). The GOMECC dataset represents the largest climate-quality (analytical uncertainty of 0.003 in pH), publicly available GOM pH dataset collected to date. Using a combination of concurrent seawater property measurements, we compare the performance of the GOM-NNpH algorithm relative to two widely used globally trained neural network algorithms. Results demonstrate the advanced skill of regionally trained neural network in reproducing ocean pH profiles within the region. We applied one of the GOM-NNpH algorithms (Equation 2) to our BGC-Argo float dataset to assess sensor performance and to generate a useable pH dataset based on the collected float profiles to date. While these algorithms are ideal for intercomparison to BGC-Argo float sensor datasets, they can be applied by various users seeking to estimate pH values in the open GOM in the absence of direct pH observations.