🌊 A team of researchers, including #URIGSO's @nickpizzooceans.bsky.social, has developed a new AI-powered way to study ocean currents.
#GOFLOW uses thermal images from weather satellites already in orbit to map ocean surface currents in unprecedented detail, with no new hardware required. 🧪
There is a ton more to do on this to make it an operational product (clouds!), but it still is a step towards better constraining obs of these surface currents. Plus, geostationary satellites *presently* sample many parts of the globe!
Joint PDFs of strain vs vorticity showing that thermal fronts generate a lot of the extremes in the distribution
You can also look at joint PDF of strain, vorticity and divergence. This again is a way to characterize certain features of the flow. At a practical level you can see how well we resolve the larger values here compared to AVISO (green line in first plot).
Currents, temperature gradients, current gradients.
This allows you to start computing things over ranges of scales only previous accessible in models. We are also interested in the gradients of the currents, as they have dynamical significance. Eg the horizontal divergence hints at the vertical motions, important for air-sea exchanges.
There is no geostrophic assumption here, which limits altimeters (and they also are more sensitive to hard-to-constrain internal wave contributions) and we are also not limited by sparse temporal sampling (eg SWOT has a repeat cycle that is ~21 days). This last point is the most significant.
U-Net architecture.
Using a "relatively simple" (his words) U-Net trained on LLC4320 we sought to answer the following question: given temperature information at time t-1hr, t and t+1hr (so this is a conditional statement), can you predict the velocity fields at time t?
At this point we made the best decision of the project and reached out to my grad school friends, Roy Barkan and Kaushik Srinivasan. Kaushik has been talking to me about ML for at least a decade, and had been looking for a good data set to test some of his ideas on.
Great! We had both done lab work together and thought we could applying those techniques -- essentially correlating current features from one frame to another. It turns out doing this when you have a lot of interleaving structures (made clearer by plotting log grad T) is difficult.
Luc noticed if you downloaded the data and changed the color bar it'd look like this
We typically look at visible or IR goes products when path planning, eg this time we were doing experiments in the North Atlantic and were staring at windy. These satellites are amazing -- not only are they high res in space, but they sample every ~5 min. This high freq allows what follows.
I was serving on a panel a few years ago and I saw I had a few missed calls from Luc. When I called him back he told me to check my email -- he had sent me a series of videos in which he'd changed the colorbar from the IR band of GOES and could vividly see thermal structure from the currents!
[Context: Larger scale ocean currents have weak vertical velocities. Smaller scale jets, fronts and eddies at the submesoscale (O(1) Ro # -- essentially .1-10km scale currents) supports vertical motions that are important for air-sea exchanges. But they are hard to measure (more on this later).]
Our recent paper on using thermal imagery from geostationary satellites to infer surface currents was published! We call the method GOFLOW 🧪🌊 www.nature.com/articles/s41...
There is a ton more to do on this to make it an operational product (clouds!), but it still is a step towards better constraining obs of these surface currents. Plus, geostationary satellites *presently* sample many parts of the globe!
You can also look at joint PDF of strain, vorticity and divergence. This again is a way to characterize certain features of the flow. At a practical level you can see how well we resolve the larger values here compared to AVISO (green line in first plot).
This allows you to start computing things over ranges of scales only previous accessible in models. We are also interested in the gradients of the currents, as they have dynamical significance. Eg the horizontal divergence hints at the vertical motions, important for air-sea exchanges.
There is no geostrophic assumption here, which limits altimeters (and they also are more sensitive to hard-to-constrain internal wave contributions) and we are also not limited by sparse temporal sampling (eg SWOT has a repeat cycle that is ~21 days). This last point is the most significant.
Using a "relatively simple" (his words) U-Net trained on LLC4320 we sought to answer the following question: given temperature information at time t-1hr, t and t+1hr (so this is a conditional statement), can you predict the velocity fields at time t?
At this point we made the best decision of the project and reached out to grad school buddies of mine, Roy Barkan and Kaushik Srinivasan. Kaushik has been talking to me about ML for at least a decade, and had been looking for a good data set to test some of his ideas on.
Great, we thought! We had both done lab work together and thought we could applying those techniques -- essentially correlating current features from one frame to another. It turns out doing this when you have a lot of interleaving structures (made clearer by plotting log grad T) is difficult.
Luc noticed if you downloaded the data and changed the color bar it'd look like this
We typically look at visible or IR goes products when path planning, eg this time we were doing experiments in the North Atlantic and were staring at windy. These satellites are amazing -- not only are they high res in space, but they sample every ~5 min. This high freq allows what follows.
I was serving on a panel a few years ago and I saw I had a few missed calls from Luc. When I called him back he told me to check my email -- he had sent me a series of videos in which he'd changed the colorbar from the IR band of GOES and could vividly see thermal structure from the currents!
[Context: Larger scale ocean currents have weak vertical velocities. Smaller scale jets, fronts and eddies at the submesoscale (O(1) Ro # -- essentially .1-10km scale currents) supports vertical motions that are important for air-sea exchanges. But they are a pain to measure (more on this later). ]
Take a timeline cleanse to watch clam mating in action and the resulting clam babies. External fertilization is a hoot!
Bubble rising through stratified flow, viewed w/ Schlieren
With Artemis in the news and space on the big screen: in case you haven't seen them, JPL has posted a series of incredible videos from some of their missions: e.g., youtu.be/1rVH0YlcKyY?... and www.youtube.com/watch?v=Ykl5... 🔭🧪 No hollywood dramatization needed
Work with me!
Earth, Environmental and Planetary Sciences at Brown University invites applications for a Doherty Postdoctoral Fellowship in Ocean Sciences. This competitive postdoctoral fellowship will be awarded for a one-year period, + 2nd year possible. Please see apply.interfolio.com/183855!
Chart of the nonet of vector mesons: a hexagon with 3 mesons at center, for a total of 9.
I'm only now learning about "vector meson dominance", a big idea put forth by Sakurai and others around 1960.
Just as the pion comes in +, - and neutral forms and has spin 0, the rho meson comes in +, - and neutral forms and has spin 1 - so it's described by a "vector", like the photon.
(1/n)
Cover of the Journal of Fluid Mechanics, Volume 1030, dated 10 March 2026. The image shows multiple droplets creating ripples on a liquid surface. The publication is by Cambridge University Press.
New Volume of Journal of Fluid Mechanics available on Cambridge Core https://cup.org/4dfo2TW
New Volume = New Cover
📸 (A16) Bouncing to coalescence transition for droplet impact onto moving liquid pools
Harris, D.M. et al. doi.org/10.1017/jfm.2026.11232
