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
Dr Heidarzadeh’s TV interview with SkyNews on the massive M7.7 Panama earthquake on 9 October 2026
Dr Heidarzadeh’s TV interview with SkyNews on the massive M7.7 Panama earthquake on 9 October 2026. Click here for the full story.
Dr Heidarzadeh’s TV interview with SkyNews on the massive M7.7 Panama earthquake on 9 October 2026
Our ocean is an incredibly vast and mysterious place. And there is no place where that is more evident than the deepest ocean area in the world: the legendary Mariana Trench. Located east of the Philippines, the Mariana Trench lies in the Mariana Archipelago in the western portion of the North Pacific. The greater area of the Mariana Trench Marine National Monument is made up of more than 95,000 square miles of ocean. Beneath these waters lie incredibly complex aggregations of submerged islands and volcanoes … and of course, an astoundingly massive trench.
Dive in with us to have your biggest questions answered about the famed Mariana Trench.
How deep is the Mariana Trench?
The deepest part of the Mariana Trench, known as Challenger Deep, is almost 36,000 feet below the ocean’s surface. For context, Mount Everest is about 29,000 feet above sea level, so the Mariana Trench is about 7,000 feet deeper than Mount Everest is tall. For one more comparison, the RMS Titanic lies at a depth of about 12,500 feet below sea level. The Challenger Deep is in waters nearly three times deeper than this ship’s final resting place.
Due to its enormous size, the Mariana Trench has at times been affectionately nicknamed the Grand Canyon of the ocean. This is certainly a picturesque image, but don’t let the mammoth size of this U.S. national terrestrial park fool you into thinking the Mariana Trench matches its size. The trench’s footprint is estimated to be 120 times larger than the Grand Canyon’s, and to reach the bottom of Challenger Deep, you’d need to descend through almost six Grand Canyons stacked on top of each other before you’d reach the seafloor.
Oh … and did we mention it’s five times wider than it is deep? It’s about 43 miles wide!
What is Challenger Deep, and how long would it take to reach it?
Challenger Deep is the deepest point of the Mariana Trench. Fewer than 30 people have ever journeyed to this area. The trek below the surface is long, dark and time-consuming, and it’s not an expedition considered lightly by scientists and explorers.
The first people to descend into the depths of this area in 1960 were oceanographers Jacques Piccard and Don Walsh, and it took them close to five hours to reach Challenger Deep. Since then, more advanced submersibles have been created, but it still takes two to four hours for modern vessels to reach such depths. For example, explorer and filmmaker James Cameron reached the seafloor in a little more than 2.5 hours in the first documented solo submarine dive to the area, and it took oceanographer Dr. Dawn Wright about four hours in 2022. It is crucial that these submersibles descend slowly and steadily beneath the surface, as the immense water pressure compounds with every mile a submersible descends. Speaking of pressure …
Get Ocean Updates in Your Inbox
Sign up with your email and never miss an update.
What is the water pressure at the bottom of the Mariana Trench?
If you thought the depth of the Mariana Trench was its only jaw-dropping feature, think again. Water pressure is one of the most important factors affecting deep-sea life. In the ocean, pressure increases by one atmosphere for every 33 feet (10 meters) of depth. One atmosphere (a unit of measurement for pressure equivalent to average air pressure at sea level) is equal to about 14.7 pounds per square inch (PSI), which is the approximate standard atmospheric pressure at sea level. If you were to convert atmospheric pressure into weight, one atmosphere is about the weight of a bowling ball. Stay with me, friends.
If a vessel were to dive 2,000 feet below the surface, the pressure would equal that of 60 bowling balls stacked atop every square inch of that submersible. But 60 bowling balls is simply nothing compared to the pressure in the Mariana Trench. The atmospheric pressure at the depths of Challenger Deep is equal to about 16,000 PSI. That pressure is about the same as the weight of an entire full-grown elephant on every square inch of a submersible!
Need a more extreme comparison? If a human were to be placed at the bottom of the Mariana Trench, the pressure would be equal to that of 50 jumbo jets piled on top of a person!
Do any animals live in the Mariana Trench?
The depths of the Mariana Trench come with immense pressure and extreme darkness, but somehow, life finds a way to survive. Organisms that live at an extreme depth, like that of Challenger Deep, have to be highly adapted to low temperatures (typically just above freezing), high pressure and virtually no light at all.
Species found in the depths of the Mariana Trench might look vastly different from animals from other ocean depths because their anatomy has been refined to thrive in harsh conditions. At such great depths, these animals typically depend on chemosynthesis (using nutrient-rich waters being pumped out of hydrothermal vents) or descending detritus (dead matter that has fallen from higher parts of the ocean) for food, such as whale falls. Just a handful of these animals include species of octopuses and squids, sea cucumbers, amphipods, sea jellies, shrimp, sea worms and some wild-looking fish!
In 2025, more than 7,000 new species were discovered in the Mariana Trench, including the hadal snailfish (Pseudoliparis swirei). This is the deepest-dwelling fish currently known to science. Now, the fish has been dubbed “the Mariana snailfish.”
How was the Mariana Trench formed?
The process by which the Mariana Trench was created is a geologically fascinating one known as subduction. On the Earth’s surface, there are thin plates that cover the planet’s mantle, which is made of molten rock. If two of these plates collide, one pushes down into the Earth’s mantle while the other is pushed over the bottom plate. When this happens, a trench is formed from the drag of the bottom plate. This process can sometimes create incredibly strong earthquakes. The plates that forged the Mariana Trench’s creation are estimated to be up to approximately 180 million years old.
This trench was unknown to scientists until 1875. It was then that a team of explorers on the HMS Challenger measured the depth of the sea by simply lowering a weighted rope into the water. They were baffled when they measured that first portion of the seafloor to be five miles below the surface. And so was born the origin of some of the most daring deep-sea explorations known to mankind to date.
Is the Mariana Trench protected?
The Mariana Trench is in waters off Guam and the Mariana Islands. The region is protected from some human activities under its status as a marine monument. The creation of the Mariana Trench Marine National Monument signified that international, national and local leaders all recognize the importance of protecting this area.
However, just because the region is protected does not mean it’s not vulnerable to modern threats to ocean health. Plastic bags have been spotted in the Mariana Trench during several expeditions, highlighting the far-reaching impact of the marine debris problem. No corner of our ocean is safe from the ever-present threat of plastic pollution. The reality of a changing climate makes the future well-being of wondrous places like the Mariana Trench uncertain. What’s more: More than 70% of the global ocean has yet to be mapped. And what we don’t yet know about, we cannot wisely work to protect.
It’s up to us to work together to defend our ocean, from the sandy shores of our favorite beaches to the deepest, darkest parts of the sea. Together, we can work to protect our ocean as a whole, which ultimately helps protect treasured places like Challenger Deep.
Find out how you can join Ocean Conservancy’s wave of momentum to secure a healthy ocean and a thriving planet, forever and for everyone!
The post The Mysteries of the Mariana Trench appeared first on Ocean Conservancy.
Intro
At the western tip of Luzon, which is the biggest island of the Philippines, lies the small coastal town of Bolinao. Tourists seldom find their way here as it takes around seven hours to get here by bus from Manila. Still, you can spot foreigners walking around every now and then, but they really stand out among the Filipinos. A reason for their presence is often that they study or work at the Bolinao Marine Laboratory (BML), which belongs to the University of the Philippines (UP). It is one of the seven research institutions participating in the GAME project 2026. Every year, GAME brings together Master’s students from all over the world to experience different cultures and conduct globally replicated experime