June 2025 | Oceanography
33
tidal forcing as the initialization states. One generator, GNT→T(·),
translated from the non-tidal to the tidal domain, and the other
generator, GT→NT(·), translated from the tidal to non-tidal
domain. To address the issue of the chaotic, turbulent nature of
the ocean, we considered the simulations to be unpaired (i.e., not
a direct translation between one state and the other). Instead,
the GAN used “cycle-consistency loss,” the mean-squared differ
ence between the original data sample and the doubly translated
data (Zhu et al., 2017). The cycle-consistency loss was combined
with the traditional GAN losses (i.e., the difference between the
generator and the discriminator output) to train the networks.
The Atlantic Ocean was used as a test-case region; one week of
hourly HYCOM data was split into 90% training data and 10%
validation data.
The GAN results retained the general structure of the tem
perature and salinity profiles from HYCOM while adding or
removing a semidiurnal tide (Figure 7). The GAN performed
well in the relatively quiescent region of the tropical mid-
Atlantic (Figure 7b). There, periodic signatures in HYCOM with
tides matched the periodicity of the outputs of GNT→T(·). The
semidiurnal signature was also removed in GT→NT(·) to match
the non-tidally forced HYCOM. It was more difficult to separate
the tidal structure from mesoscale variability in more energetic
regions, such as near the Gulf Stream (Figure 7c,d). For example,
just north of the Gulf Stream (Figure 7c), the GNT→T(·) repro
duced semidiurnal periodicity of the tidally forced HYCOM,
but there was also periodicity in the nontidal fields. In the Gulf
Stream extension (Figure 7d), the GAN imposed a periodicity to
make the sample like other tidally forced results, but this was a
region dominated by mesoscale variability.
Because the HYCOM output used to train the GAN was sam
pled from the same region of the globe during the same time of
year, no two samples were completely independent. This intro
duces the risk of overfitting. Using unpaired data made the
model more robust to overfitting but did not remove the risk
entirely. Additionally, the sound speed structure had a persistent
offset of about 5 m s–1 greater in the GAN-generated results than
the original HYCOM simulations (not shown). Thus, although
this work provides a good starting point, further work will help
revise this approach.
FIGURE 7. Temporal out
puts of the deep learn
ing GAN model at the
locations mapped in (a).
For each panel, the first
column shows the non-
tidal (NT) HYCOM results
(Exp 19.2); the second
column shows the NT
results translated into
the tidal domain using
the GAN model; the
third column shows the
tidal (T) HYCOM results
(Exp 19.0); and the fourth
column shows the T
results translated into
the NT domain using a
GAN model. From top
to bottom, rows in (b–d)
show water tempera
ture, salinity, eastward
velocity, and northward
velocity, respectively.