Computer scientists show the way: AI models need not be SO power hungry

UNIVERSITY OF COPENHAGEN - FACULTY OF SCIENCE

 

IMAGE: 

EACH DOT IN THIS FIGURE IS A CONVOLUTIONAL NEURAL NETWORK MODEL WITH THE ENERGY CONSUMPTION ON HORIZONTAL AXIS AND PERFORMANCE ON VERTICAL AXIS. CONVENTIONALLY, MODELS ARE SELECTED ONLY BASED ON THEIR PERFORMANCE - WITHOUT TAKING THEIR ENERGY CONSUMPTION INTO ACCOUNT - RESULTING IN MODELS IN THE RED ELLIPSE. THIS WORK ENABLES PRACTITIONERS TO SELECT MODELS FROM THE GREEN ELLIPSE WHICH GIVE GOOD TRADE-OFF BETWEEN EFFECTIVENESS AND EFFICIENCY. 

view more 

CREDIT: FIGURE FROM SCIENTIFIC ARTICLE (HTTPS://IEEEXPLORE.IEEE.ORG/DOCUMENT/10448303)

The fact that colossal amounts of energy are needed to Google away, talk to Siri, ask ChatGPT to get something done, or use AI in any sense, has gradually become common knowledge. One study estimates that by 2027, AI servers will consume as much energy as Argentina or Sweden. Indeed, a single ChatGPT prompt is estimated to consume, on average, as much energy as forty mobile phone charges. But the research community and the industry have yet to make the development of AI models that are energy efficient and thus more climate friendly the focus, computer science researchers at the University of Copenhagen point out.

"Today, developers are narrowly focused on building AI models that are effective in terms of the accuracy of their results. It's like saying that a car is effective because it gets you to your destination quickly, without considering the amount of fuel it uses. As a result, AI models are often inefficient in terms of energy consumption," says Assistant Professor Raghavendra Selvan from the Department of Computer Science, whose research looks in to possibilities for reducing AI’s carbon footprint.

But the new study, of which he and computer science student Pedram Bakhtiarifard are two of the authors, demonstrates that it is easy to curb a great deal of CO2e without compromising the precision of an AI model. Doing so demands keeping climate costs in mind from the design and training phases of AI models.

"If you put together a model that is energy efficient from the get-go, you reduce the carbon footprint in each phase of the model's 'life cycle'. This applies both to the model’s training, which is a particularly energy-intensive process that often takes weeks or months, as well as to its application," says Selvan.

Recipe book for the AI industry

In their study, the researchers calculated how much energy it takes to train more than 400,000 convolutional neural network type AI models – this was done without actually training all these models. Among other things, convolutional neural networks are used to analyse medical imagery, for language translation and for object and face recognition – a function you might know from the camera app on your smartphone.

Based on the calculations, the researchers present a benchmark collection of AI models that use less energy to solve a given task, but which perform at approximately the same level. The study shows that by opting for other types of models or by adjusting models, 70-80% energy savings can be achieved during the training and deployment phase, with only a 1% or less decrease in performance. And according to the researchers, this is a conservative estimate.

"Consider our results as a recipe book for the AI professionals. The recipes don’t just describe the performance of different algorithms, but how energy efficient they are. And that by swapping one ingredient with another in the design of a model, one can often achieve the same result. So now, the practitioners can choose the model they want based on both performance and energy consumption, and without needing to train each model first," says Pedram Bakhtiarifard, who continues:

"Oftentimes, many models are trained before finding the one that is suspected of being the most suitable for solving a particular task. This makes the development of AI extremely energy-intensive. Therefore, it would be more climate-friendly to choose the right model from the outset, while choosing one that does not consume too much power during the training phase."

The researchers stress that in some fields, like self-driving cars or certain areas of medicine, model precision can be critical for safety. Here, it is important not to compromise on performance. However, this shouldn’t be a deterrence to striving for high energy efficiency in other domains.

"AI has amazing potential. But if we are to ensure sustainable and responsible AI development, we need a more holistic approach that not only has model performance in mind, but also climate impact. Here, we show that it is possible to find a better trade-off. When AI models are developed for different tasks, energy efficiency ought to be a fixed criterion – just as it is standard in many other industries," concludes Raghavendra Selvan.

The “recipe book” put together in this work is available as an open-source dataset for other researchers to experiment with. The information about all these 423,000 architectures is published on Github which AI practitioners can access using simple Python scripts. 

 

 

[BOX:] EQUALS 46 YEARS OF A DANE’S ENERGY CONSUMPTION

The UCPH researchers estimated how much energy it takes to train 429,000 of the AI subtype models known as convolutional neural networks in this dataset. Among other things, these are used for object detection, language translation and medical image analysis.

It is estimated that the training alone of the 429,000 neural networks the study looked at would require 263,000 kWh. This equals the amount of energy that an average Danish citizen consumes over 46 years. And it would take one computer about 100 years to do the training. The authors in this work did not actually train these models themselves but estimated these using another AI model, and thus saving 99% of the energy it would have taken.

 

 

[BOX:] WHY IS AI’S CARBON FOOTPRINT SO BIG?

Training AI models consumes a lot of energy, and thereby emits a lot of CO2e. This is due to the intensive computations performed while training a model, typically run on powerful computers. This is especially true for large models, like the language model behind ChatGPT. AI tasks are often processed in data centers, which demand significant amounts of power to keep computers running and cool. The energy source for these centers, which may rely on fossil fuels, influences their carbon footprint.

 

 

[BOX:] ABOUT THE STUDY

• The scientific article about the study will be presented at the International Conference on Acoustics, Speech and Signal Processing (ICASSP-2024).
• The authors of the article are Pedram Bakhtiarifard, Christian Igel and Raghavendra Selvan from the University of Copenhagen’s Department of Computer Science.

---

DOI

10.1109/ICASSP48485.2024.10448303 

ARTICLE TITLE

EC-NAS: Energy Consumption Aware Tabular Benchmarks for Neural Architecture Search

 

A new 'Deep Learning' model predicts with great accuracy water and energy demands in Agriculture

Researchers develop a model based on the ‘Transformer’ architecture to guide irrigation communities' decision-making

UNIVERSITY OF CÓRDOBA

 

IMAGE: 

RESEARCHERS EMILIO CAMACHO, JUAN ANTONIO RODRÍGUEZ AND RAFAEL GONZÁLEZ, FROM THE AGRONOMY DEPARTMENT AT THE UNIVERSITY OF CÓRDOBA.

view more 

CREDIT: UNIVERSITY OF CORDOBA

Water scarcity and the high cost of energy represent the main problems for irrigation communities, which manage water for this end, making it available to agriculture. In a context of drought, with a deregulated and changing electricity market, knowing when and how much water crops are going to be irrigated with would allow those who manage them to overcome uncertainty when making decisions and, therefore, guide them towards objectives like economic savings, environmental sustainability and efficiency. For this, data science and Artificial Intelligence are important resources.

Researchers from the Hydraulics and Irrigation group with the María de Maeztu Unit of Excellence in the Agronomy Department at the University of Córdoba (DAUCO) are working to apply this cutting-edge technology to the field of precision agriculture. An example of this is the HOPE project, focused on the development of a holistic precision irrigation model that also involves the application of AI to guide decision-making. Within the framework of this effort, prediction models have been developed that would furnish irrigation communities with rigorous estimates of the amount of water that growers will need to meet their crops' needs. The latest model developed, and the most accurate to date, makes it possible to predict the actual demand for irrigation water one week ahead and with a margin of error of less than 2%, thus make possible the effective management of resources, all without detracting autonomy from its users.

According to researchers Rafael González, Emilio Camacho and Juan Antonio Rodríguez, this advance represents another step in the line of digitization applied to irrigation developed by the AGR 228 "Hydraulics and Irrigation" research group. Now they have applied the revolutionary architecture of Transformer Deep Learning to the field of precision irrigation. Since its appearance in 2017, this has been implemented in various sectors and is at the root of Artificial Intelligence milestones, such as ChatGPT. The ‘Transformer’ architecture stands out for its ability to establish long-term relationships in sequential data through what are known as 'attention mechanisms.' In the case of irrigation, this data architecture allows a lot of information to be processed simultaneously, delegating the selection and extraction of the information necessary for optimal prediction to its artificial neural network.

Daily data from the irrigation campaigns from 2015 to 2022 in the Zujar Canal's Community of Irrigators, in Don Benito (Badajoz), were used to validate the results of this model. In total, more than 1,800 water consumption measurements were used to train the model, combined with data on temperature, precipitation, solar radiation, evapotranspiration, wind speed, humidity, and crop types, etc.
This has reduced the margin of error from previous models from 20% to just 2%. Applied to integrated decision-making support systems, this can be very useful for managers of irrigation communities by offering an accurate forecast of the daily demand for irrigation water for the next seven days in contexts of water scarcity and high energy prices, but also in the framework of a commitment to sustainable resource management.

---

JOURNAL

Computers and Electronics in Agriculture

DOI

10.1016/j.compag.2024.108723 

METHOD OF RESEARCH

Computational simulation/modeling

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

Attention is all water need: Multistep time series irrigation water demand forecasting in irrigation disctrics

AI can take over key management roles in scientific research

ESMT BERLIN

 

IMAGE: 

PORTRAIT OF HENRY SAUERMANN AND MAXIMILIAN KOEHLER

view more 

CREDIT: ESMT BERLIN

Researchers Maximilian Koehler, PhD candidate at ESMT, and Henry Sauermann, professor of strategy at ESMT, explore the role of AI, not as a “worker” performing specific research tasks such as data collection and analysis, but as a “manager” of human workers performing such tasks. Algorithmic management (AM) suggests a significant shift in the way research projects are conducted and can enable projects to operate at larger scale and efficiency.

With the complexity and scope of scientific research rapidly increasing, the study illustrates that AI can not only replicate but also potentially surpass human managers by leveraging its instantaneous, comprehensive, and interactive capabilities. Investigating algorithmic management in crowd and citizen science, Koehler and Sauermann discuss examples of how AI effectively performs five important managerial functions: task division and allocation, direction, coordination, motivation, and supporting learning.

The researchers investigated projects through online documents; by interviewing organizers, AI developers, and project participants; and by joining some projects as participants. This allowed the researchers to identify projects that use algorithmic management, to understand how AI performs management functions, and to explore when AM might be more effective.

The growing number of use cases suggests that the adoption of AM could be a critical factor in improving research productivity. “The capabilities of artificial intelligence have reached a point where AI can now significantly enhance the scope and efficiency of scientific research by managing complex, large-scale projects,” states Koehler.

In a quantitative comparison with a broader sample of projects, the study also reveals that AM-enabled projects are often larger than projects that do not use AM and are associated with platforms that provide access to shared AI tools. This suggests that AM may enable projects to scale but also requires technical infrastructures that stand-alone projects may find difficult to develop. These patterns point towards changing sources of competitive advantage in research and may have important implications for research funders, digital research platforms, and larger research organizations such as universities or corporate R&D labs.

Although AI can take over important management functions, this does not mean that principal investigators or human managers will become obsolete. Sauermann notes, “If AI can take over some of the more algorithmic and mundane functions of management, human leaders could shift their attention to more strategic and social tasks such as identifying high-value research targets, raising funding, or building an effective organizational culture.”

For more information on this research, please contact Maximilian Koehler. The study “Algorithmic Management in Scientific Research,” published in the journal Research Policy, can be viewed here.

--

About ESMT Berlin

ESMT Berlin is a leading global business school with its campus in the heart of Berlin. Founded by 25 global companies, ESMT offers master, MBA, and PhD programs, as well as executive education on its campus in Berlin, in locations around the world, online, and in online blended format. Focusing on leadership, innovation, and analytics, its diverse faculty publishes outstanding research in top academic journals. Additionally, the international business school provides an interdisciplinary platform for discourse between politics, business, and academia. ESMT is a non-profit private institution of higher education with the right to grant PhDs and is accredited by AACSB, AMBA, EQUIS, and ZEvA. It is committed to diversity, equity, and inclusion across all its activities and communities. esmt.berlin

---

JOURNAL

Research Policy

DOI

10.1016/j.respol.2024.104985 

Trust your doctor: Study shows human medical professionals are more reliable than artificial intelligence tools

New research in the American Journal of Preventive Medicine puts the accuracy of advice given by large language models to the test

ELSEVIER

 

IMAGE: 

ARTIFICIAL INTELLIGENCE (AI) TOOLS ARE NOT RELIABLE ENOUGH TO BE SUBSTITUTED FOR MEDICAL PROFESSIONALS IN PROVIDING ACCURATE MEDICAL INFORMATION, RESEARCH IN THE AMERICAN JOURNAL OF PREVENTIVE MEDICINE SHOWS. FINAL ASSESSMENTS FOUND 28.6% OF CHATGPT-4'S ANSWERS ACCURATE, 28.6% INACCURATE, AND 42.8% PARTIALLY ACCURATE BUT INCOMPLETE. BARD PERFORMED BETTER, WITH 53.6% OF ANSWERS ACCURATE, 17.8% INACCURATE, AND 28.6% PARTIALLY ACCURATE.

view more 

CREDIT: AMERICAN JOURNAL OF PREVENTIVE MEDICINE

Ann Arbor, April 2, 2024 – When looking for medical information, people can use web search engines or large language models (LLMs) like ChatGPT-4 or Google Bard. However, these artificial intelligence (AI) tools have their limitations and can sometimes generate incorrect advice or instructions. A new study in the American Journal of Preventive Medicine, published by Elsevier, assesses the accuracy and reliability of AI-generated advice against established medical standards and finds that LLMs are not trustworthy enough to replace human medical professionals just yet.

Andrei Brateanu, MD, Department of Internal Medicine, Cleveland Clinic Foundation, says, "Web search engines can provide access to reputable sources of information, offering accurate details on a variety of topics such as preventive measures and general medical questions. Similarly, LLMs can offer medical information that may look very accurate and convincing, when in fact it may be occasionally inaccurate. Therefore, we thought it would be important to compare the answers from LLMs with data obtained from recognized medical organizations. This comparison helps validate the reliability of the medical information by cross-referencing it with trusted healthcare data."

In the study 56 questions were posed to ChatGPT-4 and Bard, and their responses were evaluated by two physicians for accuracy, with a third resolving any disagreements. Final assessments found 28.6% of ChatGPT-4's answers accurate, 28.6% inaccurate, and 42.8% partially accurate but incomplete. Bard performed better, with 53.6% of answers accurate, 17.8% inaccurate, and 28.6% partially accurate.

Dr. Brateanu explains, "All LLMs, including ChatGPT-4 and Bard, operate using complex mathematical algorithms. The fact that both models produced responses with inaccuracies or omitted crucial information highlights the ongoing challenge of developing AI tools that can provide dependable medical advice. This might come as a surprise, considering the advanced technology behind these models and their anticipated role in healthcare environments."

This research underscores the importance of being cautious and critical of medical information obtained from AI sources, reinforcing the need to consult healthcare professionals for accurate medical advice. For healthcare professionals, it points to the potential and limitations of using AI as a supplementary tool in providing patient care and emphasizes the ongoing need for oversight and verification of AI-generated information.

Dr. Brateanu concludes, "AI tools should not be seen as substitutes for medical professionals. Instead, they can be considered as additional resources that, when combined with human expertise, can enhance the overall quality of information provided. As we incorporate AI technology into healthcare, it's crucial to ensure that the essence of healthcare continues to be fundamentally human.”

 

 

 

---

JOURNAL

American Journal of Preventive Medicine

DOI

10.1016/j.amepre.2024.02.006 

METHOD OF RESEARCH

Content analysis

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

Accuracy of Online Artificial Intelligence Models in Primary Care Settings

Study: AI writing, illustration emits hundreds of times less carbon than humans

While energy use is much lower, tech should not replace humans, authors argue

Peer-Reviewed Publication

UNIVERSITY OF KANSAS

LAWRENCE — With the evolution of artificial intelligence comes discussion of the technology's environmental impact. A new study has found that for the tasks of writing and illustrating, AI emits hundreds of times less carbon than humans performing the same tasks. That does not mean, however, that AI can or should replace human writers and illustrators, the study’s authors argue.

Andrew Torrance, Paul E. Wilson Distinguished Professor of Law at KU, is co-author of a study that compared established systems such as ChatGPT, Bloom AI, DALL-E2 and others completing writing and illustrating to that of humans. 

Like cryptocurrency, AI has been subject to debate about the amount of energy it uses and its contributions to climate change. Human emissions and environmental impact have long been studied, but comparisons between the two have been scant. The authors conducted a comparison and found that AI systems emit between 130 and 1,500 times less CO2e (carbon dioxide equivalent) per page of text generated than human writers and illustration systems between 310 and 2,900 times less CO2e per image than humans.

“I like to think of myself as driven by data, not just what I feel is true. We’ve had discussions about something that appears to be true in terms of AI emissions, but we wanted to look at the data and see if it truly is more efficient,” Torrance said. “When we did it, the results were kind of astonishing. Even by conservative estimates, AI is extremely less wasteful.”

The study, co-written with Bill Tomlinson, Rebecca Black and Donald Patterson of the University of California-Irvine, was published in the journal Nature. 

To calculate the carbon footprint of a person writing, the researchers consulted the Energy Budget, a measure that considers the amount of energy used in certain tasks for a set period of time. For example, it is well established how much energy a computer with word processing software uses per hour. When multiplied by the average time it takes a person to write a page of text, on average, 250 words, an estimate can be arrived at. Using the same amount of energy used by the CPUs that operate AI such as ChatGPT, which can produce text much faster, produces an estimate for AI.

Researchers also considered per capita emissions of individuals in the United States and India. Residents of the former have approximate annual emissions of 15 metric tons CO2e per year, while the latter is an average of 1.9 metric tons. The two nations were chosen as they have the highest and lowest respective per capita environmental impact of countries with population higher than 300 million, and to provide a look at different levels of emissions in different parts of the world in comparison to AI.

Results showed that Bloom is 1,400 times less impactful than a U.S. resident writing a page of text and 180 times less impactful than a resident of India.

In terms of illustration, results showed that DALL-E2 emits approximately 2,500 times less CO2e than a human artist and 310 times less than an India-based artist. Figures for Midjourney were 2,900 times less for the former and 370 times less for the latter.

As technologies improve and societies evolve, those figures are almost certain to change as well, Torrance said.

The authors wrote that carbon emissions are only one factor to consider when comparing AI production to human output. As the technologies exist now, they are often not capable of producing the quality of writing or art that a human can. As they improve, they hold the potential to both eliminate existing jobs and create new ones. Loss of employment has potential for substantial economic, societal and other forms of destabilization. For those and other reasons, the authors wrote, the best path forward is likely a collaboration between AI and human efforts, or a system in which people can use AI to be more efficient in their work and retain control of final products.

Legal issues such as the use of copyrighted material in training sets for AI must be considered, the authors wrote, as does the potential for an increase in artificially produced material to result in an increase in the energy it uses and resulting emissions. Collaboration between the two is the most beneficial use of both AI and human labor, the authors wrote.

“We don’t say AI is inherently good or that it is empirically better, just that when we looked at it in these instances, it was less energy consumptive,” Torrance said.

The research was conducted to improve understanding of AI and its environmental impact and to address the United Nation Sustainable Development Goals of ensuring sustainable consumption and production patterns and taking urgent action to combat climate change and its impacts, the researchers wrote.

For their part, the authors have begun to use AI as an aid in producing drafts for some of their writing,  but they also agree on the necessity of carefully editing, and adding to, such drafts manually.

“This is not a curse, it’s a boon,” Torrance said of AI. “I think this will help make good writers great, mediocre writers good and democratize writing. It can make people more productive and can be an empowerment of human potential. I’m hugely optimistic that technology is getting better in most respects and lightening the effects we have on the Earth. We hope this is just the beginning and that people continue to dig into this issue further.”

---

JOURNAL

Nature

DOI

10.1038/s41598-024-54271-x 

METHOD OF RESEARCH

Data/statistical analysis

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

The carbon emissions of writing and illustrating are lower for AI than for humans

From data to decisions: AI and IoT for earthquake prediction

KEAI COMMUNICATIONS CO., LTD.

 

IMAGE: 

PROPOSED INTEGRATED SYSTEM ARCHITECTURE WITH MULTIPLE DATA SOURCES USED FOR AI AND ML EARTHQUAKE MODEL PREDICTION.

view more 

CREDIT: PWAVODI JOSHUA, ET AL

The study of earthquake remains a main interest worldwide as it is  one of the least predictable natural disasters. In a new review published in the KeAi journal Artificial Intelligence in Geosciences, a tea, of researchers from France and Turkey explored the role of conventional tools like seismometers and GPS in understanding earthquakes and their aftermath.

“These tools have provided invaluable insights into various seismic parameters, such as ground deformation and displacement waves. However, they face several limitations, including the inability to predict earthquakes in real-time, challenges with temporal data resolution, and uneven spatial coverage,” explains Joshua Pwavodi, lead author of the review. “Despite their historical significance, these tools struggle to distinguish seismic signals from environmental noise.”

Nevertheless, the authors note that recent advancements in AI and IoT have significantly addressed some of these limitations. AI methodologies have proven instrumental in identifying intricate patterns and complex relationships within historical seismic data. By leveraging AI, unique insights into seismic patterns across diverse geological locations have been gained.

“Both classical and advanced machine learning techniques have contributed to the development of robust early warning systems and decentralized prediction models. IoT devices have also played a crucial role by enabling seamless data transmission for real-time monitoring,” adds Pwavodi.

The versatility of IoT devices enhances data accessibility and storage, creating a dynamic network for earthquake prediction. However, challenges such as computational complexity, data quality, and interpretability persist. A major limitation is the integration of primary hydrogeological measurements into AI model training. Monitoring hydrogeological data, including pore-fluid pressures and fluid flow, is often costly. Tools like the Circulation Obviation Retrofit Kits (CORKs) provide in-situ measurements of these parameters, but data transmission is not always in real-time, unlike IoT systems.

“To address these challenges, we proposed a comprehensive approach that integrates diverse datasets, including seismic, GPS, meteorological, and IoT sensor data,” says Pwavodi. “By combining these datasets, researchers can develop more robust earthquake prediction models that account for various contributing factors.”

Specifically, the authors suggest integrating IoT devices with tools like Circulation Obviation Retrofit Kits (CORKs) to enable real-time transmission of hydrogeological measurements influencing earthquakes. This real-time data, combined with other datasets, can be used to construct predictive AI models capable of providing real-time earthquake predictions.

###

Contact the author: Pwavodi Joshua, Hwodye Technology, France, Email: pwavodi@hwodye.com

The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 100 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).

---

JOURNAL

Artificial Intelligence in Geosciences

DOI

10.1016/j.aiig.2024.100075 

METHOD OF RESEARCH

Systematic review

SUBJECT OF RESEARCH

Not applicable

ARTICLE TITLE

The Role of Artificial Intelligence and IoT in Prediction of Earthquakes: Review

Researchers propose new step in tectonic squeeze that turns seafloor into mountains

UNIVERSITY OF TEXAS AT AUSTIN

 

IMAGE: 

THE OPENING AND CLOSING OF THE ROCAS VERDES BASIN, A BACK-ARC BASIN IN PATAGONIA, AS DESCRIBED BY RESEARCHERS AT THE UNIVERSITY OF TEXAS AT AUSTIN IN A STUDY PUBLISHED IN GEOLOGY. PANELS B AND C ILLUSTRATE THE BASIN-CLOSURE PROCESS, IN WHICH AN UNDERLYING PORTION OF THE OCEANIC CRUST IS THRUST INTO A MAGMA CHAMBER (B) AND BREAKS OFF AHEAD OF CONTINENTAL CRUST (C). THE LAST TWO PANELS (C AND D) SHOW THE OCEANIC PLATE AND CONTINENTAL PLATE COLLIDING TOGETHER – SQUEEZING THE BASIN TOGETHER INTO THE ANDES MOUNTAINS OF PATAGONIA THAT WE SEE TODAY. 

view more 

CREDIT: FERNANDO REY ET AL

Scientists use tiny minerals called zircons as geologic timekeepers. Often no bigger than a grain of sand, these crystals record chemical signatures of the geological environment where they formed. In a new study led by scientists at The University of Texas at Austin, researchers used them to describe what could be an overlooked step in a fundamental tectonic process that raises seafloors into mountains.  

In a study published in the journal Geology, the researchers describe zircons from the Andes mountains of Patagonia. Although the zircons formed when tectonic plates were colliding, they have a chemical signature associated with when the plates were moving apart.  

The researchers think that the unexpected signature could be explained by the mechanics of underlying tectonic plates that hasn’t yet been described in other models. This missing step involves a sort of geologic juicing in a magma chamber where zircons form before they reach the surface, with oceanic crust entering the chamber ahead of continental crust.  

“If you put some oceanic basin below this magma, you have a change in the composition of this magma as it’s incorporated,” said the study’s lead author Fernando Rey, a doctoral student at the UT Jackson School of Geosciences. “This is something that was not documented before this study.”

This theory of oceanic magma mixing is important because it could represent a transitional step in the formation of back arc basins – an important geological structure that shapes landscapes, geologic records and helps regulate the planet’s climate.  

These basins form between oceanic and continental tectonic plates, opening up as the plates move apart and closing as they come back together. While the opening of the basin creates oceanic crust, its closing squeezes the crust into mountains – bringing a geologic record of Earth history to the surface where humans can more easily access it, said coauthor Matt Malkowski, an assistant professor at the Jackson School’s Department of Earth and Planetary Sciences. What’s more, the weathering of the ocean crust is a major driver of natural carbon dioxide storage.
“This is the Earth’s way of sequestering carbon. Very effective on its own, but it may take hundreds of thousands if not millions of years,” said Malkowski.

Malkowski collected the zircons examined in the study from rock and sediment samples at a field site in Patagonia. The samples captured the entire record of the back arc basin, called the Rocas Verdes Basin, from opening to closing.

When Rey started analyzing the chemical signatures of the zircons, nothing looked out of place at first. The zircons associated with an opening basin had the expected signature. However, when he started examining zircons associated with the closing of the basin, the signature didn’t undergo the expected chemical shift – known to scientists as a “pull down” because of the way data plotting the isotope ratios goes from steadily rising to falling down.  

When that pull down signature didn’t show up until 200 million years later, appearing in zircons that formed 30 million years ago when the basin was already well into its closure phase, Rey and his collaborators hypothesized a scenario that could help explain the data.

In their paper, they propose a model where the same tectonic forces that squeeze the oceanic crust into mountains could be underthrusting parts of that crust and pushing it toward the magmatic chamber where the zircons are formed – influencing the chemical signatures recorded in the crystals during the early to middle stages of closure. As the continents continue to squeeze together, the oceanic crust is eventually replaced by continental crust, the source of the pulldown signal.

The researchers think this transitional phase where zircons are juiced by oceanic crust could be part of back arc basins around the world. But there’s a good reason why it hasn’t been observed before, said Rey. Most back arc basins close faster than Patagonia did – in a few million years rather than tens of millions of years – meaning a shorter window of time in which these zircons can form.   

Now that scientists have discovered this zircon signal in Patagonia, they can start looking for signs of it in zircons from other places. Rey is currently analyzing zircons from the Sea of Japan – a modern back arc basin that’s in the early stages of closure – to see if there’s signs of oceanic crust influencing the zircon signature.

This research adds to a record of discovery about back arc basins at UT Austin, said Malkowski. Professor Ian Dalziel authored a well-known Nature paper in 1974 that first recognized the Andes of Patagonia as forming due to back arc basin closure. 

“Here we are 50 years later, and we’re still learning new things about these rocks,” Malkowski said.

The research was funded by the National Science Foundation and UT Austin.

Fernando Rey, a doctoral candidate at The University of Texas at Austin Jackson School of Geosciences, posing with the Lower Zapata formation in the Torres del Paine National Park, Chile. 

CREDIT

Jacqueline Epperson.

Jacqueline Epperson, a doctoral student at The University of Texas at Austin, posing close to a fault propagation fold in the outcrops of the Lower Zapata formation in the Torres del Paine National Park, Chile.

CREDIT

Fernando Rey

---

JOURNAL

Geology

DOI

10.1130/G51918.1 

ARTICLE TITLE

Detrital isotopic record of a retreating accretionary orogen: An example from the Patagonian Andes

ARTICLE PUBLICATI