By Qi-Fan Wu (Niels Bohr Institutet, University of Copenhagen)

In 1943, when Warren McCulloch and Walter Pitts showed that neurons could be represented by simple electrical circuits, they laid the first foundation for machines that could learn, adapt, and predict. In 2023, when ChatGPT became widely used, my Introduction to Python professor found that it could answer every question correctly on his course exam. In the history of machine learning, there has been a repeated oscillation between “extremely high expectations” and “deep skepticism.” What is machine learning? What should we expect from machine learning, and when should we be skeptical about it? Should the same principle also be applied to other numerical models?
The goal of machine learning is to make computers “learn” from “data”. From an end user’s perspective, it is about understanding your data, making predictions and decisions. Intellectually, it is a collection of models, methods and algorithms that have evolved over more than a half-century now [e]. Just as the human brain, neural networks, as one of the most popular machine learning methods, are theoretically capable of learning complex relationships from data. Theoretically, Neural Networks can compute any function in the world. No matter what the function is, there is guaranteed to be a neural network so that for every possible input x, the output value f(x) (or some close approximations) is output from the network (Figure 1). This result holds even if the function has many inputs and many outputs [a]. However, universal approximation only describes what neural networks are capable of, while the actual goal of machine learning is to fit an unknown function from a finite set of samples, ideally faster than traditional numerical methods.

In this blog, however, I do not want to focus on large language models that help with writing, coding, and basic background research. Instead, I want to discuss the training and use of special-purpose AI models, such as neural networks, for solving problems in physics, which is also the main topic of my PhD project. Nowadays, an increasing number of scientists are working on AI-related topics, including climate physics. If we think of the physical world as a forward dynamics model, then given the current state and the action to be taken, machine learning aims to predict the next state, while the entire world can be viewed as a huge digital database.
However, after the initial “extremely high expectations,” machine learning has also raised “deep skepticism”. In physics, especially climate physics, the “close approximations” mentioned earlier, together with the lack of standardized workflows, are often the source of trouble. The figure below shows rather discouraging results from reproducing ML-for-PDE-solving studies using stronger baselines.

As the quote attributed to von Neumann goes, “With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.” All models are wrong, including physics-based numerical methods and climate models, but many are useful because different well-performing models can still reveal different aspects of the same physical system [c]. The physicist Paul Dirac reached a similar conclusion long ago: due to the limitations of human cognitive ability, scientific theories cannot be both closed and complete at the same time [d]. This means that we cannot have perfect, exact theories. He saw that theories based on approximations could sometimes have a considerable amount of beauty in them, and he began to infer that perhaps all theories of nature are, ultimately, only approximations [d]. Personally, I think the same rule could apply to machine learning models, and indeed to all models.
My journey on the METEOR made me appreciate the importance of data even more from a modeler’s perspective, and it deepened my belief that the people who create datasets deserve more applause and respect from the entire scientific community. Because of model uncertainty, data, especially observational data, become extremely important for understanding reality. Nature itself is the ultimate database, and its ocean of data is too vast to be compressed into a single dataset.
Machine learning models and climate reanalysis systems require these high-quality data to be reliable in real-world applications. Traditional numerical weather prediction and climate models, including general circulation models, have comprehensive physical foundations but require enormous computational resources, have limited spatial resolution, and struggle to integrate multi-source observations such as station, satellite, and radar data [e]. Although AI-based weather models have developed rapidly in recent years, they still suffer from inconsistent training datasets, time periods, and regions, varying evaluation metrics, and a lack of standardized code and experimental workflows – issues similar to those previously mentioned for AI-based approaches to solving PDEs in fluid mechanics [e]. Under these circumstances, data collected by METEOR, along with all observational data, are necessary for accurately modeling weather and climate, as well as for developing model architectures for them (Figure 3). A good model should embody a trinity of observational results, physical insight, and mathematical formalism. These three aspects should correspond perfectly, with no redundancy.

But what exactly can we do to make progress under data-limited conditions? And which scientifically important problems can be clearly formulated and addressed within an analytical modeling framework? These questions remain like dark clouds hanging over scientists working in related fields. A model that is mathematically beautiful and physically simple may still be inconsistent with observations. Some models considered correct may be mathematically unattractive, and their physical mechanisms may not be clearly explained either. My personal opinion is that, when dealing with this kind of situation in the age of AI, we may still need to rely on our own intuition (and even guessing), trying to understand reality with the help of many scientists who use observations and models to sail by night and expand the boundaries of human knowledge through tiny steps.
Note: Artificial Intelligence’s (AI) stated goal is to mimic human behavior in an intelligent manner, and to do what humans can do, which includes artificial “creativity” like driving cars, playing games, responding to consumer questions, etc. In that sense, AI seeks to create muscle and mind of humans, and mind requires learning from data, i.e. Machine Learning. However, Machine Learning helps learn from data beyond mimicking humans. Having said that, the boundaries between AI and ML are getting blurry day-by-day.
References:
[a] Charniak, E. An Introduction to Deep Learning. Cambridge, MA: MIT Press, 2019; 192.
[b] Nick McGreivy. I got fooled by AI-for-science hype—here’s what it taught me. 2025. https://www.understandingai.org/p/i-got-fooled-by-ai-for-science-hypeheres
[c] Fisher, A.; Rudin, C.; Dominici, F. All Models Are Wrong, but Many Are Useful: Learning a Variable’s Importance by Studying an Entire Class of Prediction Models Simultaneously. Journal of Machine Learning Research 2019, 20(177), 1–81.
[d] Dirac, P. A. M. The Principles of Quantum Mechanics. Oxford University Press: Oxford, 1930.
[e] Bansal, H.; Grover, A.; Jewik, J.; Nguyen, T.; Sharma, P. ClimateLearn: Benchmarking Machine Learning for Weather and Climate Modeling. In Advances in Neural Information Processing Systems 36; 2023; pp 75009–75025. https://doi.org/10.52202/075280-3279.
[f] Global Observing System (GOS). World Meteorological Organization. https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/global-observing-system-gos
Ocean Acidification
How Far Does a Plastic Bottle Really Travel in a Year?
If you’ve ever joined a beach cleanup, you’ve almost certainly picked up a plastic bottle. During Ocean Conservancy’s 2025 International Coastal Cleanup® (ICC), plastic beverage bottles were the top item collected by volunteers around the world. Plastic bottles have been among the top 10 items collected throughout the ICC’s entire 41-year history.
But where do those bottles really come from? We know rivers carry plastics of all shapes and sizes from inland communities to coasts and the ocean. What we know far less about is what makes plastic bottles move within rivers, what speeds them up or slows them down and how likely they are to reach the coast and become ocean pollution. Research on river plastics is growing worldwide, but studies in North America lag behind other regions.
Ocean Conservancy scientists, alongside a team of collaborators from 10 research institutions around the United States and Canada, set out to help close some of these pressing knowledge gaps.
How do you track a plastic bottle?
… with a little help from GPS! We modeled our research after a 2021 project in the Toronto Harbor conducted by our University of Toronto colleagues. Like the Toronto project, our research team outfitted floating plastic bottles with GPS tags that reported each bottle’s location at regular intervals. That let us calculate how far and how fast every bottle traveled. Over a two-year period, we released about 300 GPS-tagged bottles into 30 stretches of river in Los Angeles, California; California’s San Francisco Bay Area; Chicago, Illinois; Washington, D.C.; Miami, Florida; Boston, Massachusetts; and Toronto, Canada.
We chose each of these places on purpose.
The locations were selected to represent varied watershed types (ranging from temperate to subtropical climates and mountainous to flat terrain) and development intensity (suburban to densely urban). We followed the bottles for about three months each, through changing seasons, storms and rainfall before locating and retrieving them. And while the plan was always to recover the bottles at the end of the experiment, we also conducted cleanups in each watershed beforehand to make sure we were leaving things a little cleaner than we started.
What did we find?
Our bottles went on quite a journey! On average, each bottle traveled 3.8 miles, but distance traveled actually spanned a huge range depending on the bottle and location. For example, one of our bottles barely budged, moving just 240 feet over the deployment period. Another made it further than the length of a marathon, clocking in at 33 miles traveled.
We found river flow was a major driver of how far bottles went, but it wasn’t the only factor—the landscape around each river mattered too. Patterns in bottle movement revealed:
- Forest and vegetation in rivers slowed bottles down. The more forest cover upstream, the less likely a bottle was to reach the coast, because it got snagged in brush, branches and other natural obstructions.
- More urbanized stretches sped them up. Bottles moved farther and faster in more urbanized stretches of river, where there was typically less vegetation to catch them.
Using thousands of GPS points, as well as data about the river systems and watersheds we studied, we built a model to estimate where bottles reaching the coast or the Great Lakes may be coming from. It showed that within a one-year period, about half originate quite close to shore (within seven miles). The other half come from much farther inland, up to 120 miles away!

Why does this matter for our ocean?
Plastic bottles aren’t just unsightly. Ocean Conservancy’s recent, award-winning research found they can be deadly to marine life—either as whole items or after they break up into sharp fragments. There have been documented instances of seabirds like albatross dying after swallowing a plastic bottle whole.
Plastic bottles are among the most prolific plastic pollutants we see out there, and now we understand that many of those we find on beaches or in the ocean may be coming from far upstream.
That has big implications for how we tackle the problem. If half of the bottles reaching the ocean around North America or Great Lakes can start their trip more than 100 miles inland, prevention must happen upstream. Pollution-prevention models, like laws that incentivize the collection and recycling of plastic bottles before they end up in the environment, should be implemented more broadly to prevent plastic from becoming pollution in the first place. Cleanups and other efforts in coastal communities also remain critical. If we want to capture floating plastics like bottles before they reach the coast and eventually our ocean, communities within 120 miles from the ocean or Great Lakes should plan to do those cleanups on an annual basis, or even more frequently.
How can I help keep plastic bottles out of rivers and the ocean?
No matter where you live, the choices you make can help keep plastic out of the ocean:
- Support policies, like those Ocean Conservancy advocates, to stop plastic pollution at the source.
- Join a cleanup along your local river, creek or lakeshore. Every bottle picked up inland is one less that can wash out to sea.
- Choose reusables instead of single-use plastic bottles whenever you can.
We want to extend a huge thank-you to our amazing collaborators on this project at the University of New Hampshire, University of Toronto, Loyola University of Chicago, Smithsonian Environmental Research Center, Florida International University, Moore Institute for Plastic Pollution Research, San Francisco Estuary Institute, University of Waterloo, University of Southern California and University of New Mexico.
Our study, “GPS-Tracked Bottles in North American Rivers Show Marine Plastic Pollution Can Originate from Far Upstream,” was published in the November 13 issue of the scientific journal Environmental Science & Technology-Water. Read the full paper here.
This blog was co-written with Hannah De Frond. Hannah works with Ocean Conservancy and the University of Toronto Trash Team to manage the International Trash Trap Network (ITTN), a global network of local groups using trash traps to increase cleanup efforts, engage communities and inform upstream solutions.
Get Ocean Updates in Your Inbox
Sign up with your email and never miss an update.
The post How Far Does a Plastic Bottle Really Travel in a Year? appeared first on Ocean Conservancy.
Ocean Acidification
How Far Does a Water Bottle Really Travel in a Year?
If you’ve ever joined a beach cleanup, you’ve almost certainly picked up a plastic bottle. During Ocean Conservancy’s 2025 International Coastal Cleanup® (ICC), plastic beverage bottles were the top item collected by volunteers around the world. Plastic bottles have been among the top 10 items collected throughout the ICC’s entire 41-year history.
But where do those bottles really come from? We know rivers carry plastics of all shapes and sizes from inland communities to coasts and the ocean. What we know far less about is what makes plastic bottles move within rivers, what speeds them up or slows them down and how likely they are to reach the coast and become ocean pollution. Research on river plastics is growing worldwide, but studies in North America lag behind other regions.
Ocean Conservancy scientists, alongside a team of collaborators from 10 research institutions around the United States and Canada, set out to help close some of these pressing knowledge gaps.
How do you track a plastic bottle?
… with a little help from GPS! We modeled our research after a 2021 project in the Toronto Harbor conducted by our University of Toronto colleagues. Like the Toronto project, our research team outfitted floating plastic bottles with GPS tags that reported each bottle’s location at regular intervals. That let us calculate how far and how fast every bottle traveled. Over a two-year period, we released about 300 GPS-tagged bottles into 30 stretches of river in Los Angeles, California; California’s San Francisco Bay Area; Chicago, Illinois; Washington, D.C.; Miami, Florida; Boston, Massachusetts; and Toronto, Canada.
We chose each of these places on purpose.
The locations were selected to represent varied watershed types (ranging from temperate to subtropical climates and mountainous to flat terrain) and development intensity (suburban to densely urban). We followed the bottles for about three months each, through changing seasons, storms and rainfall before locating and retrieving them. And while the plan was always to recover the bottles at the end of the experiment, we also conducted cleanups in each watershed beforehand to make sure we were leaving things a little cleaner than we started.
What did we find?
Our bottles went on quite a journey! On average, each bottle traveled 3.8 miles, but distance traveled actually spanned a huge range depending on the bottle and location. For example, one of our bottles barely budged, moving just 240 feet over the deployment period. Another made it further than the length of a marathon, clocking in at 33 miles traveled.
We found river flow was a major driver of how far bottles went, but it wasn’t the only factor—the landscape around each river mattered too. Patterns in bottle movement revealed:
- Forest and vegetation in rivers slowed bottles down. The more forest cover upstream, the less likely a bottle was to reach the coast, because it got snagged in brush, branches and other natural obstructions.
- More urbanized stretches sped them up. Bottles moved farther and faster in more urbanized stretches of river, where there was typically less vegetation to catch them.
Using thousands of GPS points, as well as data about the river systems and watersheds we studied, we built a model to estimate where bottles reaching the coast or the Great Lakes may be coming from. It showed that within a one-year period, about half originate quite close to shore (within seven miles). The other half come from much farther inland, up to 120 miles away!

Why does this matter for our ocean?
Plastic bottles aren’t just unsightly. Ocean Conservancy’s recent, award-winning research found they can be deadly to marine life—either as whole items or after they break up into sharp fragments. There have been documented instances of seabirds like albatross dying after swallowing a plastic bottle whole.
Plastic bottles are among the most prolific plastic pollutants we see out there, and now we understand that many of those we find on beaches or in the ocean may be coming from far upstream.
That has big implications for how we tackle the problem. If half of the bottles reaching the ocean around North America or Great Lakes can start their trip more than 100 miles inland, prevention must happen upstream. Pollution-prevention models, like laws that incentivize the collection and recycling of plastic bottles before they end up in the environment, should be implemented more broadly to prevent plastic from becoming pollution in the first place. Cleanups and other efforts in coastal communities also remain critical. If we want to capture floating plastics like bottles before they reach the coast and eventually our ocean, communities within 120 miles from the ocean or Great Lakes should plan to do those cleanups on an annual basis, or even more frequently.
How can I help keep plastic bottles out of rivers and the ocean?
No matter where you live, the choices you make can help keep plastic out of the ocean:
- Support policies, like those Ocean Conservancy advocates, to stop plastic pollution at the source.
- Join a cleanup along your local river, creek or lakeshore. Every bottle picked up inland is one less that can wash out to sea.
- Choose reusables instead of single-use plastic bottles whenever you can.
We want to extend a huge thank-you to our amazing collaborators on this project at the University of New Hampshire, University of Toronto, Loyola University of Chicago, Smithsonian Environmental Research Center, Florida International University, Moore Institute for Plastic Pollution Research, San Francisco Estuary Institute, University of Waterloo, University of Southern California and University of New Mexico.
Our study, “GPS-Tracked Bottles in North American Rivers Show Marine Plastic Pollution Can Originate from Far Upstream,” was published
