Aurelia Butler

A research team led by Professor Randolf Pohl from the Institute of Physics at Johannes Gutenberg University Mainz (JGU) has achieved a significant breakthrough in determining fundamental properties of atomic nuclei. For the first time, the team conducted laser spectroscopy experiments on muonic helium-3 at the Paul Scherrer Institute in Switzerland. Muonic helium-3 is a special form of helium in which the atom’s two electrons are replaced by a single, much heavier muon. Yesterday, the results have been published in the journal Science.
“Our experiments with muonic helium-3 provide the most accurate value to date for the charge radius of this nucleus,” says Randolf Pohl, who is also a member of the PRISMA+ Cluster of Excellence at JGU. “This is primarily because the heavy muon orbits much closer to the atomic nucleus than electrons in regular atoms, making it far more sensitive to the nucleus’s size and structure.” Helium nuclei always consist of two protons — this is what defines them as helium. Different isotopes are distinguished by the number of neutrons in the nucleus: helium-3 contains one neutron alongside the two protons, while the heavier helium-4 contains two neutrons. Professor Pohl’s team had already successfully measured helium-4 using laser spectroscopy and muons several years ago.
Measurement Confirms Theoretical Models
Traditionally, nuclear radii are determined using particle accelerators, such as MAMI at JGU or, in the future, MESA. However, the new value obtained from muonic helium measurements is fifteen times more precise, coming in at 1.97007 ± 0.00097 femtometers. Laser spectroscopy with electrons has previously been successfully applied to the lightest atomic nuclei, such as hydrogen and deuterium. For helium, highly precise measurements also exist, but the presence of two electrons in the helium atom makes theoretical calculations more complex, preventing accurate determination of the nuclear radius from such measurements alone. Nevertheless, it has already been possible to determine the difference in charge radii between various helium isotopes (nuclei with the same number of protons but different numbers of neutrons). The new results from muonic helium measurements align well with recent experiments on regular helium conducted by a research team in Amsterdam, also published in Science, yesterday.
“In combination with our earlier results on muonic helium-4, which were published in Nature in 2021, we have now been able to precisely determine the difference in charge radii between helium-3 and helium-4 — an important advancement,” says Pohl.
Close Interaction Between Theory and Experiment in the PRISMA+ Cluster of Excellence
The strong agreement between the measurements conducted by the Mainz and Amsterdam teams confirms existing knowledge about the nuclear physics of the lightest atomic nuclei. This knowledge is based in part on key theoretical calculations of nuclear structure, also carried out within the PRISMA+ Cluster of Excellence.

Professor Sonia Bacca’s team calculated the influence of the muon on the structure of the helium nucleus, while Professor Marc Vanderhaeghen and Dr. Franziska Hagelstein explored the roles of protons and neutrons. These theoretical foundations enabled the extraction of reliable information about the atomic nucleus from the precise experimental data.
“Accurate knowledge of nuclear charge radii is essential for determining fundamental physical constants such as the Rydberg constant,” Pohl explains. “It is also crucial for the search for new physics — particles and forces not yet included in the Standard Model. The previous lack of precise data in this area introduced significant uncertainties.”
Precise Measurement of Additional Atomic Nuclei Planned
Looking ahead, the team of experimental and theoretical physicists at Mainz plans to apply their methods to other atomic nuclei — from lithium to neon — with ten times the accuracy compared to particle accelerator-based methods. Instead of lasers, they will use innovative X-ray detectors. This work, like the previous experiments led by Pohl’s group, is supported by the German Research Foundation (DFG) as part of the Collaborative Research Center 1660 at JGU. More

Self-driving cars which eliminate traffic jams, getting a healthcare diagnosis instantly without leaving your home, or feeling the touch of loved ones based across the continent may sound like the stuff of science fiction.
But new research, led by the University of Bristol and published today in the journal Nature Electronics, could make all this and more a step closer to reality thanks to a radical breakthrough in semiconductor technology.
The futuristic concepts rely on the ability to communicate and transfer vast volumes of data much faster than existing networks. So physicists have developed an innovative way to accelerate this process between scores of users, potentially across the globe.
Co-lead author Martin Kuball, Professor of Physics at the University of Bristol, said: “Within the next decade, previously almost unimaginable technologies to transform a wide range of human experiences could be widely available. The possible benefits are also far-reaching, including advances in healthcare with remote diagnostics and surgery, virtual classrooms and even virtual holiday tourism.
“In addition, there is considerable potential for advanced driver assistance systems to improve road safety and industrial automation for greater efficiency. The list of possible 6G applications is endless, with the limit just being human imagination. So our innovative semiconductor discoveries are hugely exciting and will help drive forward these developments at speed and scale.”
It is widely acknowledged that the shift from 5G to 6G will demand a radical upgrade of semiconductor technology, circuits, systems, and associated algorithms. For instance, the main semiconductor components involved, in other words the radio frequency amplifiers made from a wonder conductor called Gallium Nitride (GaN), need to be much quicker, emit greater power, and be more reliable.
The team of international scientists and engineers has tested a new architecture, catapulting these special GaN amplifiers to unprecedented heights. This was achieved by discovering a latch-effect in GaN, which unlocked a much greater radio frequency device performance. These next generation devices use parallel channels which then require the use of sub-100nm side fins — a type of transistor which controls the flow of current passing through the devices.

Co-lead author Dr Akhil Shaji, Honorary Research Associate at the University of Bristol, explained: “We have piloted a device technology, working with collaborators, called superlattice castellated field effect transistors (SLCFETs), in which more than 1000 fins with sub-100 nm width help drive the current. Although SLCFETs have demonstrated the highest performance in the W-band frequency range, equating to 75 gigahertz -110 GHz, the physics behind it was unknown.
“We recognised it was a latch-effect in GaN, which enables the high radio frequency performance.”
The researchers then needed to pinpoint exactly where this effect occurred, by simultaneously using ultra precision electrical measurements and optical microscopy, so it could be further studied and understood. After analysing more than 1,000 fins findings located this effect to the widest fin.
Prof Kuball, who is also Royal Academy of Engineering Chair in Emerging Technologies, added: “We also developed a 3D model using a simulator to further verify our observations. The next challenge was to study the reliability aspects of latch effect for practical applications. The rigorous testing of the device over a long duration of time showed it has no detrimental effect on device reliability or performance.
“We found a key aspect driving this reliability was a thin layer of dielectric coating around each of the fins. But the main takeaway was clear — the latch effect can be exploited for countless practical applications, which could help transform people’s lives in many different ways in years to come.”
Next steps for the work include further increasing the power density the devices can deliver, so they can offer even higher performance and serve wider audiences. Industry partners will also be bringing such next generation devices to a commercial market.
Researchers at the University of Bristol are at the forefront of improving electrical performance and efficiency in a wide range of different applications and settings.
Professor Kuball leads the Centre for Device Thermography and Reliability (CDTR), which is developing next generation semiconductor electronic devices for net zero, and for communications and radar technology. It also works on improving device thermal management, electrical performance and reliability, using wide and ultra-wide bandgap semiconductors. More

Plasma — the electrically charged fourth state of matter — is at the heart of many important industrial processes, including those used to make computer chips and coat materials. Simulating those plasmas can be challenging, however, because millions of math operations must be performed for thousands of points in the simulation, many times per second. Even with the world’s fastest supercomputers, scientists have struggled to create a kinetic simulation — which considers individual particles — that is detailed and fast enough to help them improve those manufacturing processes.
Now, a new method offers improved stability and efficiency for kinetic simulations of what’s known as inductively coupled plasmas. The method was implemented in a code developed as part of a private-public partnership between the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) and chip equipment maker Applied Materials Inc., which is already using the tool. Researchers from the University of Alberta, PPPL and Los Alamos National Laboratory contributed to the project.
Detailed simulations of these plasmas are important to gain a better understanding of how plasma forms and evolves for various manufacturing processes. The more realistic the simulation, the more accurate the distribution functions it provides. These measures show, for example, the probability that a particle is at a particular location moving at a particular speed. Ultimately, understanding these details could lead to realizations about how to use the plasma in a more refined way to etch patterns onto silicon for even faster chips or memory with greater storage, for example.
“This is a big step forward in our capabilities,” said Igor Kaganovich, a principal research physicist at PPPL and co-author of a journal article published in Physics of Plasmas that details the simulation findings.
Making the code reliable
The initial version of the code was developed using an old method that proved unreliable. Dmytro Sydorenko, a research associate at the University of Alberta and first author of the paper, said that significant modifications of the method were made to make the code much more stable. “We changed the equations, so the simulation immediately became very reliable and there were no crashes anymore,” he said. “So now we have a usable tool for the simulation of inductively coupled plasmas into two spatial dimensions.”
The code was improved, in part, by changing the way one of the electric fields was calculated. An electric field is like an invisible force field that surrounds electric charges and currents. It exerts forces on particles. In an inductively coupled plasma, a wire coil carrying an electric current generates a changing magnetic field, which, in turn, generates an electric field that heats the plasma. It is this field, known as the solenoidal electric field, that the team focused its efforts on.

The code calculates electromagnetic fields based on procedures developed by Salomon Janhunen from Los Alamos National Laboratory. These procedures were optimized by PPPL’s Jin Chen, who acted as a bridge between physics, mathematics and computer science aspects of the challenge. “For a complicated problem, the improvement is significant,” Chen said.
The simulation is known as a particle-in-cell code because it tracks individual particles (or small groups of particles clumped together as so-called macroparticles) while they move in space from one grid cell to another. This approach works particularly well for the plasmas used in industrial devices where the gas pressure is low. A fluid approach doesn’t work for such plasmas because it uses average values instead of tracking individual particles.
Obeying the law of conservation of energy
“This new simulation allows us to model larger plasmas quickly while accurately conserving energy, helping to ensure the results reflect real physical processes rather than numerical artifacts,” said Kaganovich.
In the real world, energy doesn’t randomly appear or disappear. It follows the law of conservation of energy. But a small mistake in a computer simulation can accumulate with each step. Because each simulation might involve thousands or even millions of steps, a small error throws off the results significantly. Making sure energy is conserved helps keep the simulation faithful to a real plasma.
PPPL’s Stéphane Ethier also worked on the new simulation code. The work was supported by a Cooperative Research and Development Agreement between Applied Materials Inc. and PPPL, under contract number DE-AC02-09CH11466. More

A new study from UBC Okanagan says students appear to be using generative artificial intelligence (GenAI) responsibly, and as a way to speed up tasks, not just boost their grades.
Dr. Meaghan MacNutt, who teaches professional ethics in the UBCO School of Health and Exercise Sciences (HES), recently published a study in Advances in Physiology Education. Published this month, the paper — titled Reflective writing assignments in the era of GenAI: student behaviour and attitudes suggest utility, not futility — contradicts common concerns about student use of AI.
Students in three different courses, almost 400 participants, anonymously completed a survey about their use of AI on at least five reflective writing assignments. All three courses used an identical AI policy and students had the option to use the tool for their writing.
“GenAI tools like ChatGPT allow users to interface with large language models. They offer incredible promise to enhance student learning, however, they are also susceptible to misuse in completion of writing assignments,” says Dr. MacNutt. “This potential has raised concerns about GenAI as a serious threat to academic integrity and to the learning that occurs when students draft and revise their own written work.”
While UBC offers guidance to students and faculty about the risks and benefits of using GenAI, policies regarding its use in courses are at the discretion of individual instructors.
Dr. MacNutt, who completed the study with doctoral student and HES lecturer Tori Stranges, notes that discipline-specific factors contribute to the perception that many courses in HES are particularly challenging and many students strive for excellence, often at the expense of their mental wellbeing.
So, how often were the students using AI and what was motivating their use?

While only about one-third of the students used AI, the majority of users, 81 per cent, reported their GenAI use was inspired by at least one of the following factors: speed and ease in completing the assignment, a desire for high grades and a desire to learn. About 15 per cent of the students said they were motivated by all three factors, with more than 50 per cent using it to save time on the assignment.
Dr. MacNutt notes that most students used AI to initiate the paper or revise sections. Only 0.3 per cent of assignments were mostly written by GenAI.
“There is a lot of speculation when it comes to student use of AI,” she says. “However, students in our study reported that GenAI use was motivated more by learning than by grades, and they are using GenAI tools selectively and in ways they believe are ethical and supportive of their learning. This was somewhat unexpected due to the common perception that undergraduate students have become increasingly focused on grades at the expense of learning.”
The study does raise some cautions, she warns. GenAI can be a useful tool for students learning English or people with reading and writing disabilities. But there is also the potential that if paid versions are better, students who can afford to use a more effective platform might have an advantage over others — creating further classroom inequities.
MacNutt says continued research in this area can only provide a better understanding of student behaviour and attitudes as GenAI technologies continue to advance. She also suggests, while AI continues to be used more frequently, that institutions and educators adopt an approach that embodies “collaboration with” rather than “surveillance of” students.
“Our findings contradict common concerns about widespread student misuse and overuse of GenAI at the expense of academic integrity and learning,” says Dr. MacNutt. “But as we move forward with our policies, or how we’re teaching students how to use it, we have to keep in mind that students are coming from really different places. And they have different ways of benefiting or being harmed by these technologies.” More

As climate-related disasters grow more intense and frequent, an international team of researchers has introduced Aurora — a groundbreaking AI model designed to deliver faster, more accurate, and more affordable forecasts for air quality, ocean waves, and extreme weather events. This model, called Aurora, has been trained on over a million hours of data. According to the researchers, it could revolutionize the way we prepare for natural disasters and respond to climate change.
From deadly floods in Europe to intensifying tropical cyclones around the world, the climate crisis has made timely and precise forecasting more essential than ever. Yet traditional forecasting methods rely on highly complex numerical models developed over decades, requiring powerful supercomputers and large teams of experts. According to its developers, Aurora offers a powerful and efficient alternative using artificial intelligence.
Machine learning at the core
‘Aurora uses state-of-the-art machine learning techniques to deliver superior forecasts for key environmental systems — air quality, weather, ocean waves, and tropical cyclones,’ explains Max Welling, machine learning expert at the University of Amsterdam and one of the researchers behind the model. Unlike conventional methods, Aurora requires far less computational power, making high-quality forecasting more accessible and scalable — especially in regions that lack expensive infrastructure.
Trained on a million hours of earth data
Aurora is built on a 1.3 billion parameter foundation model, trained on more than one million hours of Earth system data. It has been fine-tuned to excel in a range of forecasting tasks: Air quality: Outperforms traditional models in 74% of cases Ocean waves: Exceeds numerical simulations on 86% of targets Tropical cyclones: Beats seven operational forecasting centres in 100% of tests High-resolution weather: Surpasses leading models in 92% of scenarios, especially during extreme eventsForecasting that’s fast, accurate, and inclusive
As climate volatility increases, rapid and reliable forecasts are crucial for disaster preparedness, emergency response, and climate adaptation. The researchers believe Aurora can help by making advanced forecasting more accessible.

‘Development cycles that once took years can now be completed in just weeks by small engineering teams,’ notes AI researcher Ana Lucic, also of the University of Amsterdam. ‘This could be especially valuable for countries in the Global South, smaller weather services, and research groups focused on localised climate risks.’ ‘Importantly, this acceleration builds on decades of foundational research and the vast datasets made available through traditional forecasting methods,’ Welling adds.
Aurora is available freely online for anyone to use. If someone wants to fine-tune it for a specific task, they will need to provide data for that task. ‘But the “initial” training is done, we don’t need these vast datasets anymore, all the information from them is baked into Aurora already’, Lucic explains.
A future-proof forecasting tool
Although current research focuses on the four applications mentioned above, the researchers say Aurora is flexible and can be used for a wide range of future scenarios. These could include forecasting flood risks, wildfire spread, seasonal weather trends, agricultural yields, and renewable energy output. ‘Its ability to process diverse data types makes it a powerful and future-ready tool’, states Welling.
As the world faces more extreme weather — from heatwaves to hurricanes — innovative models like Aurora could shift the global approach from reactive crisis response to proactive climate resilience concludes the study. More

Is artificial intelligence (AI) capable of suggesting appropriate behaviour in emotionally charged situations? A team from the University of Geneva (UNIGE) and the University of Bern (UniBE) put six generative AIs — including ChatGPT — to the test using emotional intelligence (EI) assessments typically designed for humans. The outcome: these AIs outperformed average human performance and were even able to generate new tests in record time. These findings open up new possibilities for AI in education, coaching, and conflict management. The study is published in Communications Psychology.
Large Language Models (LLMs) are artificial intelligence (AI) systems capable of processing, interpreting and generating human language. The ChatGPT generative AI, for example, is based on this type of model. LLMs can answer questions and solve complex problems. But can they also suggest emotionally intelligent behaviour?
These results pave the way for AI to be used in contexts thought to be reserved for humans.
Emotionally charged scenarios
To find out, a team from UniBE, Institute of Psychology, and UNIGE’s Swiss Center for Affective Sciences (CISA) subjected six LLMs (ChatGPT-4, ChatGPT-o1, Gemini 1.5 Flash, Copilot 365, Claude 3.5 Haiku and DeepSeek V3) to emotional intelligence tests. ”We chose five tests commonly used in both research and corporate settings. They involved emotionally charged scenarios designed to assess the ability to understand, regulate, and manage emotions,” says Katja Schlegel, lecturer and principal investigator at the Division of Personality Psychology, Differential Psychology, and Assessment at the Institute of Psychology at UniBE, and lead author of the study.
For example: One of Michael’s colleagues has stolen his idea and is being unfairly congratulated. What would be Michael’s most effective reaction?
a) Argue with the colleague involved
b) Talk to his superior about the situation

c) Silently resent his colleague
d) Steal an idea back
Here, option b) was considered the most appropriate.
In parallel, the same five tests were administered to human participants. “In the end, the LLMs achieved significantly higher scores — 82% correct answers versus 56% for humans. This suggests that these AIs not only understand emotions, but also grasp what it means to behave with emotional intelligence,” explains Marcello Mortillaro, senior scientist at the UNIGE’s Swiss Center for Affective Sciences (CISA), who was involved in the research.
New tests in record time
In a second stage, the scientists asked ChatGPT-4 to create new emotional intelligence tests, with new scenarios. These automatically generated tests were then taken by over 400 participants. ”They proved to be as reliable, clear and realistic as the original tests, which had taken years to develop,” explains Katja Schlegel. ”LLMs are therefore not only capable of finding the best answer among the various available options, but also of generating new scenarios adapted to a desired context. This reinforces the idea that LLMs, such as ChatGPT, have emotional knowledge and can reason about emotions,” adds Marcello Mortillaro.
These results pave the way for AI to be used in contexts thought to be reserved for humans, such as education, coaching or conflict management, provided it is used and supervised by experts. More

A new artificial intelligence system pinpoints the origin of 3D printed parts down to the specific machine that made them. The technology could allow manufacturers to monitor their suppliers and manage their supply chains, detecting early problems and verifying that suppliers are following agreed upon processes.
A team of researchers led by Bill King, a professor of mechanical science and engineering at the University of Illinois Urbana-Champaign, has discovered that parts made by additive manufacturing, also known as 3D printing, carry a unique signature from the specific machine that fabricated them. This inspired the development of an AI system which detects the signature, or “fingerprint,” from a photograph of the part and identifies its origin.
“We are still amazed that this works: we can print the same part design on two identical machines -same model, same process settings, same material — and each machine leaves a unique fingerprint that the AI model can trace back to the machine,” King said. “It’s possible to determine exactly where and how something was made. You don’t have to take your supplier’s word on anything.”
The results of this study were recently published in the Nature partner journal Advanced Manufacturing.
The technology has major implications for supplier management and quality control, according to King. When a manufacturer contracts a supplier to produce parts for a product, the supplier typically agrees to adhere to a specific set of machines, processes, and factory procedures and not to make any changes without permission. However, this provision is difficult to enforce. Suppliers often make changes without notice, from the fabrication process to the materials used. They are normally benign, but they can also cause major issues in the final product.
“Modern supply chains are based on trust,” King said. “There’s due diligence in the form of audits and site tours at the start of the relationship. But, for most companies, it’s not feasible to continuously monitor their suppliers. Changes to the manufacturing process can go unnoticed for a long time, and you don’t find out until a bad batch of products is made. Everyone who works in manufacturing has a story about a supplier that changed something without permission and caused a serious problem.”
While studying the repeatability of 3D printers, King’s research group noticed that the tolerances of part dimensions were correlated with individual machines. This inspired the researchers to examine photographs of the parts. It turned out that it is possible to determine the specific machine made the part, the fabrication process, and the materials used — the production “fingerprint.”
“These manufacturing fingerprints have been hiding in plain sight,” King said. “There are thousands of 3D printers in the world, and tens of millions of 3D printed parts used in airplanes, automobiles, medical devices, consumer products, and a host of other applications. Each one of these parts has a unique signature that can be detected using AI.”

King’s research group developed an AI model to identify production fingerprints from photographs taken with smartphone cameras. The AI model was developed on a large data set, comprising photographs of 9,192 parts made on 21 machines from six companies and with four different fabrication processes. When calibrating their model, the researchers found that a fingerprint could be obtained with 98% accuracy from just 1 square millimeter of the part’s surface.
“Our results suggest that the AI model can make accurate predictions when trained with as few as 10 parts,” King said. “Using just a few samples from a supplier, it’s possible to verify everything that they deliver after.”
King believes that this technology has the potential to overhaul supply chain management. With it, manufacturers can detect problems at early stages of production, and they save the time and resources needed to pinpoint the origins of errors. The technology could also be used to track the origins of illicit goods.
Miles Bimrose, Davis McGregor, Charlie Wood and Sameh Tawfick also contributed to this work. More

Increasingly powerful AI models can make short-term weather forecasts with surprising accuracy. But neural networks only predict based on patterns from the past — what happens when the weather does something that’s unprecedented in recorded history? A new study led by scientists from the University of Chicago, in collaboration with New York University and the University of California Santa Cruz, is testing the limits of AI-powered weather prediction. In research published May 21 in Proceedings of the National Academy of Sciences, they found that neural networks cannot forecast weather events beyond the scope of existing training data — which might leave out events like 200-year floods, unprecedented heat waves or massive hurricanes.
This limitation is particularly important as researchers incorporate neural networks into operational weather forecasting, early warning systems, and long-term risk assesments, the authors said. But they also said there are ways to address the problem by integrating more math and physics into the AI tools.
“AI weather models are one of the biggest achievements in AI in science. What we found is that they are remarkable, but not magical,” said Pedram Hassanzadeh, an associate professor of geophysical sciences at UChicago and a corresponding author on the study. “We’ve only had these models for a few years, so there’s a lot of room for innovation.”
Gray swan events
Weather forecasting AIs work in a similar way to other neural networks that many people now interact with, such as ChatGPT.
Essentially, the model is “trained” by feeding it a bunch of text or images into a model and asking it to look for patterns. Then, when a user presents the model with a question, it looks back at what it’s previously seen and uses that to predict an answer.
In the case of weather forecasts, scientists train neural networks by feeding them decades’ worth of weather data. Then a user can input data about the current weather conditions and ask the model to predict the weather for the next several days.

The AI models are very good at this. Generally, they can achieve the same accuracy as a top-of-the-line, supercomputer-based weather model that uses 10,000 to 100,000 times more time and energy, Hassanzadeh said.
“These models do really, really well for day-to-day weather,” he said. “But what if next week there’s a freak weather event?”
The concern is that the neural network is only working off the weather data we currently have, which goes back about 40 years. But that’s not the full range of possible weather.
“The floods caused by Hurricane Harvey in 2017 were considered a once-in-a-2,000-year event, for example,” Hassanzadeh said. “They can happen.”
Scientists sometimes refer to these events as “gray swan” events. They’re not quite all the way to a black swan event — something like the asteroid that killed the dinosaurs — but they are locally devastating.
The team decided to test the limits of the AI models using hurricanes as an example. They trained a neural network using decades of weather data, but removed all the hurricanes stronger than a Category 2. Then they fed it an atmospheric condition that leads to a Category 5 hurricane in a few days. Could the model extrapolate to predict the strength of the hurricane?

The answer was no.
“It always underestimated the event. The model knows something is coming, but it always predicts it’ll only be a Category 2 hurricane,” said Yongqiang Sun, research scientist at UChicago and the other corresponding author on the study.
This kind of error, known as a false negative, is a big deal in weather forecasting. If a forecast tells you a storm will be a Category 5 hurricane and it only turns out to be a Category 2, that means people evacuated who may not have needed to, which is not ideal. But if a forecast underestimates a hurricane that turns out to be a Category 5, the consequences would be far worse.
Hurricane warnings and why physics matters
The big difference between neural networks and traditional weather models is that traditional models “understand” physics. Scientists design them to incorporate our understanding of the math and physics that govern atmospheric dynamics, jet streams and other phenomena.
The neural networks aren’t doing any of that. Like ChatGPT, which is essentially a predictive text machine, they simply look at weather patterns and suggest what comes next, based on what has happened in the past.
No major service is currently using only AI models for forecasting. But as their use expands, this tendency will need to be factored in, Hassanzadeh said.
Researchers, from meteorologists to economists, are beginning to use AI for long-term risk assessments. For example, they might ask an AI to generate many examples of weather patterns, so that we can see the most extreme events that might happen in each region in the future. But if an AI cannot predict anything stronger than what it’s seen before, its usefulness would be limited for this critical task. However, they found the model could predict stronger hurricanes if there was any precedent, even elsewhere in the world, in its training data. For example, if the researchers deleted all the evidence of Atlantic hurricanes but left in Pacific hurricanes, the model could extrapolate to predict Atlantic hurricanes.
“This was a surprising and encouraging finding: it means that the models can forecast an event that was unpresented in one region but occurred once in a while in another region,” Hassanzadeh said.
Merging approaches
The solution, the researchers suggested, is to begin incorporating mathematical tools and the principles of atmospheric physics into AI-based models.
“The hope is that if AI models can really learn atmospheric dynamics, they will be able to figure out how to forecast gray swans,” Hassanzadeh said.
How to do this is a hot area of research. One promising approach the team is pursuing is called active learning — where AI helps guide traditional physics-based weather models to create more examples of extreme events, which can then be used to improve the AI’s training.
“Longer simulated or observed datasets aren’t going to work. We need to think about smarter ways to generate data,” said Jonathan Weare, professor at the Courant Institute of Mathematical Sciences at New York University and study co-author. “In this case, that means answering the question ‘where should I place my training data to achieve better performance on extremes?’ Fortunately, we think AI weather models themselves, when paired with the right mathematical tools, can help answer this question.”
University of Chicago Prof. Dorian Abbot and computational scientist Mohsen Zand were also co-authors on the study, as well as Ashesh Chattopadhyay of the University of California Santa Cruz.
The study used resources maintained by the University of Chicago Research Computing Center. More