Raghavasimhan Sankaranarayanan has over 200 album and film soundtrack credits to his name, and he has performed in more than 2,000 concerts across the globe. He has composed music across many genres and received numerous awards for his technical artistry on the violin. 

He is also a student at Georgia Tech, finishing up his Ph.D. in machine learning and robotics. 

One might wonder why a successful professional musician would choose to become a student again.

“I always wanted to integrate technology, music, and robotics because I love computers and machines that can move,” he said. “There’s been little research on Indian music from a technological perspective, and the AI and music industries largely focus on Western music. This bias is something I wanted to address.”

Sankaranarayanan, who began playing the violin at age 4, has focused his academic studies on bridging the musically technical with the deeply technological. Over the past six years at Georgia Tech, he has explored robotic musicianship, creating a robot violinist and an accompanying synthesizer capable of understanding, playing, and improvising the music closest to his heart: classical South Indian music.

The Essence of Carnatic Music

Carnatic music, a classical form of South Indian music, is believed to have originated in the Vedas, or ancient sacred Hindu texts. The genre has remained faithful to its historic form, with performers often using non-amplified sound or only a single mic. A typical performance includes improvisations and musical interaction between musicians in which violinists play a crucial role. 

Carnatic music is characterized by intricate microtonal pitch variations known as gamakas — musical embellishments that modify a single note’s pitch or seamlessly transition between notes. In contrast, Western music typically treats successive notes as distinct entities.

Out of a desire to contribute technological advancements to the genre, Sankaranarayanan set out to innovate. When he joined the Center for Music Technology program under Gil Weinberg, professor and the Center’s director, no one at Georgia Tech had ever attempted to create a string-based robot.

“In our work, we develop physical robots that can understand music, apply logic to it, and then play, improvise, and inspire humans,” said Weinberg. “The goal is to foster meaningful interactions between robots and human musicians that foster creativity and the kind of musical discoveries that may not have happened otherwise.”

The Brain and the Body

Sankaranarayanan conceptualizes the robot as comprising two parts: the brain and the body. The “body” consists of mechanical systems that require algorithms to move accurately, including sliders and actuators that convert electric signals into motion to produce the sound of music. The “brain” consists of algorithms that enable the robot to understand and generate music.

In robotic musicianship, algorithms interpret and perform music, but building these algorithms for non-Western music is challenging; far less data is available for these forms. This lack of representation limits the capabilities of robotic musicianship and diminishes the cultural richness diverse musical forms can offer. 

Classical algorithms would struggle to capture the nuances of Carnatic music. To address this, Sankaranarayanan collected data specifically to model gamakas in Carnatic music. Then, using audio from performances by human musicians, he developed a machine-learning model to learn those gamakas. 

“You may ask, ‘Why not just use a computer?’ A computer can respond with algorithms, but music’s physicality is vital,” Weinberg said. “When musicians collaborate, they rely on the visual cues of movement, which make the interaction feel alive. Moreover, acoustic sound created by a physical instrument is richer and more expressive than computer-generated sound, and a robot musician provides this.”

Sankaranarayanan built the robot incrementally. Initially, he developed a bow mechanism that moved across wheels; now, the robot violin uses a real bow for authentic sound production.

Developing a New Musical Language

Another challenge involves technologies like MIDI (Musical Instrument Digital Interface), a protocol that enables electronic musical instruments and devices to communicate and sync by sending digital information about musical notes and performances. MIDI, however, is based on Western music systems and is limited in its application to music with microtonal pitch variations such as Carnatic music.

So Sankaranarayanan and Weinberg developed their own system. Using audio files of human violin performances, the system extracts musical features that inform the robot on bowing techniques, left-hand movements, and pressure on strings. The software synthesizer then listens to Sankaranarayanan’s playing, responding and improvising in real time and creating a dynamic interplay between human and robot.

“Like in many other fields, bias also exists in the area of music AI, with many researchers and companies focusing on Western music and using AI to understand tonal systems,” Weinberg said. “Raghav’s work aims to showcase how AI can also generate and understand non-Western music, which he has achieved beautifully.”

Giving Back to the Community

Carnatic music and its community of musicians shaped Sankaranarayanan's musical sensibility, motivating him to give back. He is developing an app to teach Carnatic music to help make the genre more appealing to younger audiences.

“By merging tradition with technology, we can expand the reach of traditional Carnatic music to younger musicians and listeners who desire more technological engagement,” Sankaranarayanan said. 

Through his innovative work, he is not just preserving Carnatic music but also reshaping its future for a digital age, inviting a new generation to engage with its deep heritage.

To better understand why some cancer patients struggle to fight off infections, Georgia Tech researchers have created tiny lab-grown models of human immune systems.

These miniature models — known as human immune organoids — mimic the real-life environment where immune cells learn to recognize and attack harmful invaders and respond to vaccines. Not only are these organoids powerful new tools for studying and observing immune function in cancer, their use is likely to accelerate vaccine development, better predict disease treatment response for patients, and even speed up clinical trials. 

“Our synthetic hydrogels create a breakthrough environment for human immune organoids, allowing us to model antibody production from scratch, more precisely, and for a longer duration,” said Ankur Singh, Carl Ring Family Professor in the George W. Woodruff School of Mechanical Engineering and professor in the Wallace H. Coulter Department of Biomedical Engineering at Georgia Tech and Emory. 

“For the first time, we can recreate and sustain complex immunological processes in a synthetic gel, using blood, and effectively track B cell responses,” he added. “This is a game changer for understanding and treating immune vulnerabilities in patients with lymphoma who have undergone cancer treatment — and hopefully other disorders too.”

Led by Singh, the team created lab-grown immune systems that mimic human tonsils and lymph node tissue to study immune responses more accurately. Their research findings, published in the journal Nature Materials, mark a shift toward in vitro models that more closely represent human immunology. The team also included investigators from Emory University, Children’s Healthcare of Atlanta, and Vanderbilt University.

Designing a Tiny Immune System Model

The researchers were inspired to address a critical issue in biomedical science: the poor success rate of translating preclinical findings from animal models into effective clinical outcomes, especially in the context of immunity, infection, and vaccine responses. 

“While animal models are valuable for many types of research, they often fail to accurately mirror realistic human immune biology, disease mechanisms, and treatment responses,” said Monica (Zhe) Zhong, a bioengineering Ph.D. student and the paper’s first author. “To address this, we designed a new model that faithfully replicates the unique complexity of human immune biology across molecular, cellular, tissue, and system levels.”

The team used synthetic hydrogels to recreate a microenvironment where B cells from human blood and tonsils can mature and produce antibodies. When immune cells from healthy donors or lymphoma patients are cultured in these gel-like environments, the organoids support longer cell function, allowing processes like antibody formation and adaptation to occur — similar to the human body. Utilizing the organoids for individual patients helps predict how that individual will respond to infection.

The models also enable researchers to control and test immune responses under various conditions. The team discovered that not all tissue sources are the same, and tonsil cells struggled with longevity issues. They used a specialized setup to study how healthy immune cells react to signals that help them fight infections, which failed to trigger the same response in cells from lymphoma survivors who seemingly have recovered from immunotherapy treatment.

Using organoids embedded in a novel immune organ-on-chip technology, the team observed that immune cells from lymphoma survivors treated with certain immunotherapies do not organize themselves into specific “zones,” the way they normally would in a strong immune response. This lack of organization may help explain some immune challenges cancer survivors face, as evidenced by recent clinical findings. 

A Game-Changing Technology

This research is primarily of interest to infectious disease researchers, cancer researchers, immunologists, and healthcare professionals dedicated to improving patient outcomes. By studying these miniature immune systems, they can identify why current treatments may not be effective and explore new strategies to enhance immune defenses. 

"Lymphoma patients treated with CD20-targeted therapies often face increased susceptibility to infections that can persist years after completing therapy. Understanding these long-term impacts on antibody responses could be key to improving both safety and quality of life for lymphoma survivors,” said Dr. Jean Koff, associate professor in the department of Hematology and Oncology at Emory University’s Winship Cancer Institute and a co-author on the paper. 

“This technology provides deeper biological insights and an innovative way to monitor for recovery of immunological defects over time. It could help clinicians better identify patients who would benefit from specific interventions that reduce infection risk,” Koff added.

Another critical and promising aspect of the research is its scalability: An individual researcher can make hundreds of organoids in a single sitting. The model’s capability to target different populations — both healthy and immunosuppressed patients — vastly increases its usability for vaccine and therapeutic testing. 

According to Singh, who directs the  Center for Immunoengineering at Georgia Tech, the team is already pushing the research into new dimensions, including developing cellular therapies and an aged immune system model to address aging-related questions. 

“At the end of the day, this work most immediately affects cancer patients and survivors, who often struggle with weakened immune responses and may not respond well to standard treatments like vaccines,” Singh explained. “This breakthrough could lead to new ways of boosting immune defenses, ultimately helping vulnerable patients stay healthier and recover more fully.” 

The work was initially funded by the Wellcome Leap HOPE program. This support has led to a boost in recent funding, including a recent $7.5 million grant from the National Institute of Allergy and Infectious Diseases.

Citation: Zhong, Z., Quiñones-Pérez, M., Dai, Z. et al. Human immune organoids to decode B cell response in healthy donors and patients with lymphoma. Nat. Mater. (2024).

DOI: https://doi.org/10.1038/s41563-024-02037-1

Funding: Wellcome Leap HOPE Program, National Institutes of Health, National Institute of Allergy and Infectious Diseases, National Cancer Institute, and Georgia Tech Foundation

Ankur Singh, Carl Ring Family Professor in the George W. Woodruff School of Mechanical Engineering and professor in the Wallace H. Coulter Department of Biomedical Engineering at Georgia Tech and Emory, and Monica (Zhe) Zhong, a Bioengineering Ph.D. student and the paper’s first author.

This month, the future of artificial intelligence (AI) was spotlighted as more than 120 academic and industry researchers participated in the Georgia Tech School of Interactive Computing’s inaugural Summit on Responsible Computing, AI, and Society.

With looming questions about AI's growing roles and consequences in nearly every facet of modern life, School of IC organizers felt the time was right to diverge from traditional conferences that focus on past work and published research.

“Presenting papers is about disseminating work that has already been completed. Who gets to be in the room is determined by whose paper gets accepted,” said Mark Riedl, School of IC professor.

“Instead, we wanted the summit talks to speculate on future directions and what challenges we as a community should be thinking about going forward.”

The two-day summit, held at Tech’s Global Learning Center Oct. 28-30, convened to discuss consequential questions like:

• Is society ready to accept more responsibility as greater advancements in technologies like AI are made?
• Should society stop to think about potential consequences before these advancements are implemented on its behalf, and what could those consequences be?
• What policies should be enacted for these technologies to mitigate harms and augment societal benefits?

A highlight of the summit’s opening day was Meredith Ringel Morris's keynote address. As director of human-AI interaction research at Google DeepMind, she presented a possible future in which humans could use AI to create a digital afterlife.

In her remarks, Morris discussed AI clones, which are AI avatars of specific human beings with high autonomy and task-performing capabilities. Someone could leave such an agent behind as a memory for loved ones to enjoy once they are gone, and future generations could access it to learn more about an ancestor.

On the other hand, it could easily lead to loved ones experiencing extended grief because they have trouble moving on from losing a family member.

These AI capabilities are in development and will soon be publicly available. As industry and academic researchers continue to develop them, the public needs to learn about their eminent impacts.

“There’s a lot that needs to be done in educating people,” Morris said. “It’s hard for well-intentioned and thoughtful system designers to anticipate all the harm. We must be prepared some people are going to use AI in ways we don’t anticipate, and some of those ways are going to be undesirable. What legal and education structures can we create that will help?”

In addition to Morris’s keynote, the summit’s first day included 20 talks about future and emerging technologies in AI, sustainability, healthcare, and other fields. 

The second day featured eight talks on translating interventions and safeguards into policy.

Day-two speakers included: 

• Orly Lobel, Warren Distinguished Professor of Law and director of the Center for Employment and Labor Policy at the University of California-San Diego. Lobel worked on President Obama’s policy team on innovation and labor market competition, and she advises the Federal Trade Commission (FTC). 
• Sorelle Friedler, Shibulal Family Professor of Computer Science at Haverford College. She worked in the Office of Science and Technology Policy (OSTP) under the Biden-Harris Administration and helped draft the AI Bill of Rights. 
• Jake Metcalf, researcher and program director for AI on the Ground at the think tank Data & Society. The organization produces reports on data science and equity for the US Government. 
• Divyansh Kaushik, Vice President of Beacon Global Strategies, has given testimony to the US Senate on AI research and development.

Kaushik earned a Ph.D. in machine learning from Carnegie Mellon University before beginning his career in policy. He highlighted the importance of policymakers fostering relationships with academic researchers.

“Policymakers think about what could go wrong,” Kaushik said. “Academia can offer evidence-based answers.”

The summit also hosted a doctoral consortium, which allowed advanced Ph.D. students to present their research to experts and receive feedback and mentoring.

“Being an interdisciplinary researcher is challenging,” said Shaowen Bardzell, School of IC chair.

“We wanted the next generation to be in the room listening to the experts share their visions and also to provide our own experiences when possible on how to navigate the challenges and rewards of doing work in the intersection of AI, healthcare, sustainability, and policy.”

As greenhouse gases accumulate in the Earth’s atmosphere, scientists are developing technologies to pull billions of tons of carbon dioxide (CO2) from the air and inject it deep underground. 

The idea isn’t new. In the 1970s, Italian physicist Cesare Marchetti suggested that the carbon dioxide polluting the air and warming the planet could be stored underground. The reality of how to do it cost-effectively and safely has challenged scientists for decades. 

Geologic carbon storage — the subterranean storage of CO2 — comes with significant challenges, most importantly, how to avoid fracturing underground rock layers and letting gas escape into the atmosphere. Carbon, a gas, can behave erratically or leak whenever it’s stored in a compressed space, making areas geologically unstable and potentially causing legal headaches for corporations that invest in it. This uncertainty, coupled with the expense of the carbon capture process and its infrastructure, means the industry needs reliable predictions to justify it. 

Georgia Tech researcher and Georgia Research Alliance Eminent Scholar Felix J. Herrmann has an answer. His lab, Seismic Laboratory for Imaging and Modeling (SLIM), uses advanced artificial intelligence (AI) techniques to create algorithms that monitor and optimize carbon storage. The algorithms work as “digital twins,” or digital replicas of underground systems, facilitating the safe, efficient storage of CO2 underground.

“The trick is you want the carbon to stay put — to avoid the risk of, say, triggering an earthquake or the carbon leaking out,” said Herrmann, professor in the School of Earth and Atmospheric Sciences and the School of Electrical and Computer Engineering. “We’re developing a digital twin that allows us to monitor and control what is happening underground.”

Predicting the Best Place

Waveform Variational Inference via Subsurface Extensions with Refinements (WISER) is an algorithm that uses sound waves to analyze underground structures. WISER runs on AI, enabling it to work more efficiently than most algorithms while remaining computationally feasible. To improve accuracy, WISER makes small adjustments using sound wave physics to show how fast sound travels through different materials and where there’s variation in underground layers. This helps to create detailed, reliable images of underground areas for better predictions of carbon storage. 

WISER allows researchers to work with uncertainties, which is vital for understanding the risk of these underground storage projects.

Scaling the Algorithm

While Herrmann’s lab has been working to apply neural networks to seismic imaging for years now, WISER required them to increase the networks’ scale. Making multiple predictions is a much larger problem that requires a bigger, more potent network, but these types of neural networks only run on graphics processing units (GPUs), which are known for speed but are limited in memory.

To optimize the GPU, Rafael Orozco, a computational science and engineering Ph.D. student, created a new type of neural network that can train with very little memory. This open-source package, InvertibleNetworks, enables the network to train on very large inputs and create multiple output images conditioned on the observed seismic data.

WISER’s fundamental innovation is for the lab’s next concept: creating digital twins for carbon storage. These twins can act as monitoring systems to optimize and mitigate risks of carbon storage projects. 

Devising the Digital Twin

Digital twins are dynamic virtual models of objects in the real world, capable of replicating their behavior and performance. They rely on real-time data to evolve and have been used to replicate factories, cities, spacecraft, and bodies, to make informed decisions about healthcare, maintenance, production, supply chains, and — in Herrmann’s case — geologic carbon storage.

Herrmann and his team have developed an “uncertainty-aware” digital twin. That means the tool can manage risks and make decisions in an uncertain, unseen environment — because it’s been designed to recognize, quantify, and incorporate uncertainties in CO2 storage.

Probing the Unseen

Subsurface conditions are diverse and complex, making the management of greenhouse gas storage a delicate process. Without careful monitoring, the injection of CO2 can increase pressure in rock formations, potentially fracturing the cap rock that is supposed to keep the gas underground.

“The digital twin addresses this through simulations in tandem with observations,” said Herrmann, whose team linked two different scientific fields — geophysics and reservoir engineering — for a more comprehensive understanding of the subsurface environment. Specifically, they combined geophysical well observations with seismic imaging.

Geophysical well observation involves drilling a hole in the subsurface in a geological area of interest and collecting data by lowering a probe into the borehole to take measurements. Seismic imaging, on the other hand,  uses acoustic waves to create images based on the analysis of wave vibrations.

“Bridging the gap between different fields of research and combining various data sources allows our digital twin to provide a more accurate and detailed picture of what’s happening underground,” Herrmann said. 

To integrate and leverage these diverse datasets built from observations and simulations, the team used advanced AI techniques like simulation-based inference and sequential Bayesian inference, a method of updating information as more data becomes available. The ongoing learning allows researchers to quantify uncertainties in the subsurface environment and predict how that system will respond to CO2 injection. The digital twin updates its understanding as new data becomes available.

Making Informed Decisions

Herrmann’s team tested the digital twin, simulating different states of an underground reservoir, including permeability, which is the measure of how easily fluids flow through rock. The goal was to find the maximum injection rate of CO2 without causing fractures in the cap rock.

“The work highlights how dynamic digital twins can play a key role in mitigating the risks associated with geologic carbon storage,” said Herrmann, whose research group is supported in part by large oil and gas companies, including Chevron and  ExxonMobil. “Companies are now in the process of starting large offshore projects for which the digital twin is being developed.”

But there is still plenty of work to be done, he added. For instance, the digital twin can monitor the subsurface and provide critical information about that uncertain environment. It can inform. But it still needs adjustments by humans for each new CO2 injection site, and Herrmann and his team are working on further developing the technology — giving the digital twins the ability to quickly replicate themselves so they can be deployed massively and quickly to meet the demands of mitigating climate change.

“Our aim is to make them smarter,” Herrmann said. “To make them more adaptable, so they can control CO2 injections, become more responsive to risks, and adapt to a wide range of complex situations in real time.”

A first-of-its-kind algorithm developed at Georgia Tech is helping scientists study interactions between electrons. This innovation in modeling technology can lead to discoveries in physics, chemistry, materials science, and other fields.

The new algorithm is faster than existing methods while remaining highly accurate. The solver surpasses the limits of current models by demonstrating scalability across chemical system sizes ranging from large to small. 

Computer scientists and engineers benefit from the algorithm’s ability to balance processor loads. This work allows researchers to tackle larger, more complex problems without the prohibitive costs associated with previous methods.

Its ability to solve block linear systems drives the algorithm’s ingenuity. According to the researchers, their approach is the first known use of a block linear system solver to calculate electronic correlation energy.

The Georgia Tech team won’t need to travel far to share their findings with the broader high-performance computing community. They will present their work in Atlanta at the 2024 International Conference for High Performance Computing, Networking, Storage and Analysis (SC24).

“The combination of solving large problems with high accuracy can enable density functional theory simulation to tackle new problems in science and engineering,” said Edmond Chow, professor and associate chair of Georgia Tech’s School of Computational Science and Engineering (CSE).

Density functional theory (DFT) is a modeling method for studying electronic structure in many-body systems, such as atoms and molecules. 

An important concept DFT models is electronic correlation, the interaction between electrons in a quantum system. Electron correlation energy is the measure of how much the movement of one electron is influenced by presence of all other electrons.

Random phase approximation (RPA) is used to calculate electron correlation energy. While RPA is very accurate, it becomes computationally more expensive as the size of the system being calculated increases.

Georgia Tech’s algorithm enhances electronic correlation energy computations within the RPA framework. The approach circumvents inefficiencies and achieves faster solution times, even for small-scale chemical systems.

The group integrated the algorithm into existing work on SPARC, a real-space electronic structure software package for accurate, efficient, and scalable solutions of DFT equations. School of Civil and Environmental Engineering Professor Phanish Suryanarayana is SPARC’s lead researcher.

The group tested the algorithm on small chemical systems of silicon crystals numbering as few as eight atoms. The method achieved faster calculation times and scaled to larger system sizes than direct approaches.

“This algorithm will enable SPARC to perform electronic structure calculations for realistic systems with a level of accuracy that is the gold standard in chemical and materials science research,” said Suryanarayana.

RPA is expensive because it relies on quartic scaling. When the size of a chemical system is doubled, the computational cost increases by a factor of 16. 

Instead, Georgia Tech’s algorithm scales cubically by solving block linear systems. This capability makes it feasible to solve larger problems at less expense. 

Solving block linear systems presents a challenging trade-off in solving different block sizes. While larger blocks help reduce the number of steps of the solver, using them demands higher computational cost per step on computer processors. 

Tech’s solution is a dynamic block size selection solver. The solver allows each processor to independently select block sizes to calculate. This solution further assists in scaling, and improves processor load balancing and parallel efficiency.

“The new algorithm has many forms of parallelism, making it suitable for immense numbers of processors,” Chow said. “The algorithm works in a real-space, finite-difference DFT code. Such a code can scale efficiently on the largest supercomputers.”

Georgia Tech alumni Shikhar Shah (Ph.D. CSE 2024), Hua Huang (Ph.D. CSE 2024), and Ph.D. student Boqin Zhang led the algorithm’s development. The project was the culmination of work for Shah and Huang, who completed their degrees this summer. John E. Pask, a physicist at Lawrence Livermore National Laboratory, joined the Tech researchers on the work.

Shah, Huang, Zhang, Suryanarayana, and Chow are among more than 50 students, faculty, research scientists, and alumni affiliated with Georgia Tech who are scheduled to give more than 30 presentations at SC24. The experts will present their research through papers, posters, panels, and workshops. 

SC24 takes place Nov. 17-22 at the Georgia World Congress Center in Atlanta. 

“The project’s success came from combining expertise from people with diverse backgrounds ranging from numerical methods to chemistry and materials science to high-performance computing,” Chow said.

“We could not have achieved this as individual teams working alone.”

Just as a chameleon changes colors to mask itself from predators, new AI-powered technology is protecting people’s photos from online privacy threats.

The innovative model, developed at Georgia Tech, creates invisible digital masks for personal photos to thwart unwanted online facial recognition while preserving the image quality.

Anyone who posts photos of themselves risks having their privacy violated by unauthorized facial image collection. Online criminals and other bad actors collect facial images by web scraping to create databases.

These illicit databases enable criminals to commit identity fraud, stalking, and other crimes. The practice also opens victims to unwanted targeted ads and attacks.

The new model is called Chameleon. Unlike current models, which produce different masks for each user’s photos, Chameleon creates a single, personalized privacy protection (P-3) mask for all of a user’s facial photos.

A bespoke P-3 mask is created based on a few user-submitted facial photos. After applying the mask, protected photos won’t be detectable by someone scanning for the user’s face. Instead, the unwanted scan will identify the protected photos as being someone else.

The Chameleon model was developed by Professor Ling Liu of the School of Computer Science (SCS), Ph.D. students Sihao Hu and Tiansheng Huang, and Ka-Ho Chow, an assistant professor at the University of Hong Kong and Liu’s former Ph.D. student.

During development, the team accomplished its two main goals: protecting the person's identity in the photo and ensuring a minimal visual difference between the original and masked photos.

The researchers said a notable visual difference often exists between the original and photos using current masking models. However, Chameleon preserves much of the original photo’s quality among various facial images.

In several research tests, Chameleon outperformed three top facial recognition protection models in visual and protective metrics. The tests also showed that Chameleon offers more substantial privacy protection while being faster and more resource-efficient.

In the future, Huang said they would like to apply Chameleon’s methods to other uses.

“We would like to use these techniques to protect images from being used to train artificial intelligence generative models. We could protect the image information from being used without consent,” he said.

The research team aims to release Chameleon code publicly on GitHub to allow others to improve their work.

“Privacy-preserving data sharing and analytics like Chameleon will help to advance governance and responsible adoption of AI technology and stimulate responsible science and innovation,” said Liu.

The paper on Chameleon, Personalized Privacy Protection Mask Against Unauthorized Facial Recognition, was presented earlier this month at ECCV 2024.

Oxygen is essential for living organisms, particularly multicellular life, to metabolize organic matter and energize all life activities. About half of the oxygen we breathe comes from terrestrial plant life, such as forests and grasslands, while the other half is produced through photosynthesis by marine algae in the ocean's surface waters.

Oxygen concentrations are declining in many parts of the world’s oceans. Experts believe this drop is linked to the ocean’s surface warming and its impacts on the physics and chemistry of seawater, though the problem is not fully understood. Temperature plays a crucial role in determining how oxygen dissolves in seawater; as water warms, it loses its ability to hold gas.

“Calculating the amount of oxygen lost from the oceans is challenging due to limited historical measurements and inconsistent timing,” said Taka Ito, oceanographer and professor in the School of Earth and Atmospheric Sciences at Georgia Tech. “To understand global oxygen levels and their changes, we need to fill in many data gaps.”

A group of student researchers sought to address this issue. Led by Ito, the team developed a new machine learning-based approach to more accurately understand and represent the decline in global ocean oxygen levels. Using datasets, the team further generated a monthly map of oxygen content visualizing the ocean’s oxygen decline over several decades. Their research was published in the Journal of Geophysical Research: Machine Learning and Computation.

“Marine scientists need to understand the distribution of oxygen in the ocean, how much it's changing, where the changes are occurring, and why,” said Ahron Cervania, a Ph.D. student in Ito’s lab. “Statistical methods have long been used for these estimates, but machine learning techniques can improve the accuracy and resolution of our oxygen assessments.”

The project began three years ago with support from the National Science Foundation, and the team initially focused solely on Atlantic Ocean data to test the new method. They used a computational model to generate hypothetical observations, which allowed them to assess how well they could reconstruct missing oxygen level information using only a fraction of the data combined with machine learning. After developing this method, the team expanded to global ocean observations, involving undergraduate students and delegating tasks across different ocean basins.

Under Ito’s guidance, Cervania and other student researchers developed algorithms to analyze the relationships between oxygen content and variables like temperature, salinity, and pressure. They used a dataset of historical, ship-based oxygen observations since the 1960s and recent data from Argo floats — autonomous drifting devices that collect and measure temperature and salinity. Although oxygen data existed before the 1960s, earlier records have accuracy issues, so the team focused on data from the 1960s onward. They then created a global monthly map of ocean oxygen content from 1965 to the present.

“Using a machine learning approach, we were able to assess the rate of oxygen loss more precisely across different periods and locations,” Cervania said. “Our findings indicate that incorporating float data significantly enhances the estimate of oxygen loss while also reducing uncertainty.”

The team found that the world’s oceans have lost oxygen at a rate of about 0.7% per decade from 1970 to 2010. This estimate suggests a relatively rapid ocean response to recent climate change, with potential long-term impacts on marine ecosystems’ health and sustainability. Their estimate also falls within the range of decline suggested by other studies, indicating the accuracy and efficacy of their approach.

“We calculated trends in global oxygen levels and the ocean’s inventory, essentially looking at the rate of change over the last five decades,” Cervania said. “It’s encouraging to see that our rate aligns with previous estimates from other methods, which gives us confidence. We are building a robust estimate from both our study and other studies.”

According to Ito, the team’s new approach addresses an ongoing challenge in the oceanographic community: how to effectively combine different data sources with varying accuracies and uncertainties to better understand ocean changes.  

 “The integration of advanced technologies like machine learning will be essential in filling data gaps and providing a clearer picture of how our oceans are responding to climate change.”

Citation: Ito, T., Cervania, A., Cross, K., Ainchwar, S., & Delawalla, S. (2024). Mapping dissolved oxygen concentrations by combining shipboard and Argo observations using machine learning algorithms. Journal of Geophysical Research: Machine Learning and Computation, 1, e2024JH000272. 

DOI: https://doi.org/10.1029/2024JH000272

Funding: National Science Foundation

Basin-scale oxygen inventory trend with global ocean divided to 10 basins. Blue lines and shading show ensemble mean and ensemble range for ship-only reconstructions, and red lines and shading are for ship and Argo float reconstructions. The research team divided up the task by working on different basins. (Credit: Georgia Institute of Technology)

Taka Ito, oceanographer and professor in the School of Earth and Atmospheric Sciences.

Ahron Cervania, a Ph.D. student in Earth and Atmospheric Sciences.

A collaborative team from Georgia Tech, Purdue University, Case Western Reserve University (CWRU), Michigan State University (MSU), and Brown University have been awarded a combined $7.3 million from the Office of Naval Research (ONR) as part of the Multidisciplinary University Research Initiative (MURI) program. MURI seeks to fund research teams with creative and diverse solutions to complex problems and is a major part of the Department of Defense’s (DoD) research portfolio.

Satish Kumar, Frank H. Neely Professor in the George W. Woodruff School of Mechanical Engineering, along with his collaborators, received the five-year award for their project, Machine learning Enabled Two-pHase flow metrologies, models, and Optimized DesignS (METHODS). Georgia Tech, CWRU, and Brown University will lead efforts of machine learning (ML)-assisted discovery of mechanistic two-phase models and generalized two phase flow model development. Purdue University (lead on the project) and MSU will lead efforts in performing two-phase flow experiments and generating high resolution data for the project.

Read the full story on the George W. Woodruff School of Mechanical Engineering website.