The El Niño Southern Oscillation (ENSO) is an irregular climate pattern that brings heavy rain, droughts and temperature changes around the world. ENSO cycles between El Niño and La Nina every 2-7 years, and is not always easy to forecast. Right now the ENSO status is neutral, meaning neither El Niño nor La Niña conditions are present.

Scientists have now used machine learning to predict ENSO 23 months in advance. It is through timely predictions that governments and communities can better prepare for potential flooding and droughts, which can cause disruptions to farming, water supply, buildings and transport.

We spoke to Dr Ioana Colfescu, a research scientist at the National Centre for Atmospheric Science (NCAS) and the University of St Andrews, about her ENSO machine learning research.

“The project has used machine learning to improve ENSO predictions and identify new indicators to improve predictability. The work showed it is possible to predict ENSO sometimes up to 23 months in advance. Since ENSO events significantly influence global weather patterns, including droughts, heavy rainfall, and temperature variations, improved predictions can help communities prepare for these changes, which is crucial for public safety. 

“Improving the accuracy of ENSO predictions using machine learning is relevant to government agencies, businesses, researchers, and the general public. There should be a lot of interest in terms of scientific advancement as researchers and scientists have a vast interest in new methodologies, like machine learning, that have the potential to  improve predictive capabilities and be faster than traditional methods. Sharing successful outcomes can advance the field, foster further research, and encourage researchers to trust such methodologies more, or not to! ML it is not however a silver bullet for all problems and in all cases and as with any new methodology there is a lot we have to understand in terms of why and how it works but there is a lot of progress too”. 

“ENSO predictions are vital for understanding climate impacts. Emerging technologies, especially machine learning, also provide unprecedented opportunities to tackle old challenges with large volumes of data that were previously unavailable. Leveraging technology and data is essential for addressing urgent environmental issues, making our efforts more crucial and timely than ever. We are addressing state of the art research questions, like how does ENSO, as well as other similar large scale modes of variability work, how does it change, and how does climate change affect ENSO. 

“We used Convolutional Neural Networks (CNNs) to take large datasets, including historical sea-surface temperatures, atmospheric pressure, and other climate-related variables to find if we have an ENSO or not for a certain year in that data. By looking at past ENSO patterns, this algorithm will learn to recognise and predict future El Niño and La Nina. A CNN is a machine learning methodology designed to process and analyse visual data, although it can also be applied to various types of data. CNNs will automatically detect patterns and features in the input data. For example, they could detect specific things in images, anything you want down to small details such as edges or small lines.”

“The most exciting result of our study is that we can achieve quite accurate predictions of ENSO up to 23 months in advance by focusing solely on wind data from a small area in the Indian Ocean. This approach allows us to simplify the prediction process, contrasting with traditional methods that rely on global maps and numerous variables. Additionally, our findings indicate that while wind in the Eastern Indian Ocean is a crucial factor, its influence is intricately linked to the sea-surface temperatures in the western Indian Ocean.

“We found that relying on a single weather variable, such as wind, within a specific area of the Indian Ocean can yield predictions comparable to those derived from a comprehensive global model. This approach reduces computational resources, paving the way for more efficient data-driven predictions using machine learning.” 

“In a broader context this research is very interesting as what this hints at is a few things. First, using just one weather variable (e.g. wind) as a predictor for a small region over the Indian Ocean can get, in terms of prediction, quite close to what we get by using a global map. Computationally this is very important as it opens lots of potential for making data based predictions using machine learning and using a fraction of the computations needed for running a dynamical model. Second, it addresses directly improving ENSO predictions which we need for mitigating and adapting to the extreme events linked to or enhanced by ENSO. This is increasingly important in a warming world. Last, I think the work clearly demonstrates the need to use physical based understanding, along with data centric methodologies to better understand and predict our climate system.

“At a general level I would say we hope we have achieved improved predictive accuracy. By enhancing the accuracy of ENSO predictions, especially compared to traditional statistical and dynamical models, we identified some important features that significantly influence ENSO events. We have suggested an approach in which we integrate multiple data types, for example observations and ocean temperature from models, with a machine learning approach for a more comprehensive analysis. We have developed a machine learning based model that can provide predictions or updates, facilitating timely decision-making for climate-related issues if further refined and implemented.”

“Machine learning studies present several challenges. The main one is that they require large datasets for training and at the same time such methods do not catch enough of the physical aspects. Incomplete, biased, or low-quality data can also lead to inaccurate predictions. Historical climate data may also contain gaps, making it difficult to create reliable models. Second, we have to increasingly use such data based methodologies along mathematical models and equations if we want to have a more accurate and physical understanding .e.g. Hybrid models are on the rise. 

“Then it is the model interpretability issue: many machine learning models operate as “black boxes”, making it challenging to understand how they arrive at specific predictions. This lack of interpretability can hinder trust among scientists and people who use research findings, like policymakers and insurance companies. Combining machine learning predictions with traditional climate modeling methods can be a way to address some of this. For this research, I did try to use some ‘traditional methods’ to understand how and why machine learning produced the predictive skill we got but there’s still lots to be done and lots to understand.

“There are also the risks of the changing climate to consider. Namely, the climate system is complex and can change over time due to various factors, including human activity but also internal dynamics. Models trained on historical data, like the one used in this study, may not perform well yet. The underlying dynamics of ENSO could change in the future in a changing climate. We have no guarantee that a model which works for the past 30 years will work in the next 30 years. Also, even if a machine learning model performs well in a research setting, changing the data or even scaling it for operational use in predicting ENSO can present logistical and technical challenges. 

“For any computational related project there’s always a computational cost involved and here, while machine learning can reduce some computational demands at the same time, training sophisticated or very  models for example still requires significant computational resources and time. It is a matter of balance and of learning where and how to use a tool and where we don’t actually need it.” 

“It was a truly international and collaborative pursuit as the work was done as part of a grant from the University of Oxford (led by Hannah Christensen in Oxford) and colleagues from America at the National Centre for Atmospheric Research. 

“This was my first ML related work and  now I have started working on expanding this work in a few directions in St Andrews and by resuming the collaboration with Hannah.

“Following on from this research there is the opportunity to further develop models that generalise well across different ENSO phases (El Niño and La Niña) and various climatic conditions and other datasets. There is a lot which can be done in terms of model interpretability, for example to achieve a level of interpretability that sheds light on how the machine learning model makes predictions. 

“There is also the option to investigate how models act, finding out to what extent they are robust in adapting to changes in climate dynamics and capable of accounting for variability in ENSO behaviour over time. I do have plans to follow up on this as part of my work on CANARI – a national science project aimed at understanding extreme weather impacts on the UK.” 

“I was at a crossroad in my career. It was the time around the pandemic ending, a strange time. I needed a change and I felt learning something new, especially about Machine Learning, would help me. 

“I decided either pursuing a Masters or leaving Academia for a job within industry was the way to go. As life doesn’t really go according to the plan I did neither and one day I got forwarded an advert for a post-doctoral research position to do with machine learning for ENSO predictions at the University of Oxford. The opportunity to work on ENSO prediction – an important subject for climate scientists – while learning about Machine Learning and being based in Oxford was appealing, so I decided to seize the chance. Also, it fit well with something else – I was just starting to supervise students who frequently inquired about machine learning and wanted to pursue work using machine learning. This made it clear that I needed to deepen my knowledge in this area. 

“I am grateful NCAS fully supported me taking a secondment from my current NCAS position. I went back to being a post-doctoral researcher and moved to Oxford for one year. At that point it seemed like a step backwards career-wise, because when you change direction you can’t publish much (and it took way longer than one year for me to start understanding a bit more what I was doing given the change in my career), you don’t know what you are doing, and your old job doesn’t really advance.

“I wondered quite a few times if what I did was right, if I’ll really use this in the future but in the end I really enjoyed learning all I did and I decided to trust the process and go with it. In the long term it turned out to be really beneficial and I’m absolutely happy I did it – mostly because it was really interesting work and equally I met wonderful people both in Oxford – in particular the project lead and my supervisor, and I met new colleagues and found new opportunities within NCAS. But I am still very much learning about ML.”   

“On a more personal level, this work represented a huge learning pursuit, finding new collaborators, adapting to a new environment as well as a career direction change for me. Some of the hardest aspects of it are however the fact that sometimes you are not seen as ‘an expert’ and you lose credibility if you start something new and it takes time till gaining that back …  . 

“I have met wonderful people and I hope they will be long-term, career-long collaborators so I would do it all again. Yet, this makes me think about the problems we have in academia with keeping up with the digital technologies,and learning and adapting fast enough to all this novelty. Many of us are experts in our own fields and yet we now have to learn all these new technology concepts, then apply them, then keep up with what feels sometimes like two fields – how do you do these both or even more than two aspects ? The time to learn, the best way, is it worth it, to what extent should we learn about technologies – all these questions are really important to address.

“I am aware that some of the  NCAS colleagues, like me, are going through a similar transition and wonder how do we keep up with all the technology and machine learning revolution, how do we learn fast enough or how do we embrace this change in our career paths. 

“There is not real recipe and is a bit like learning by doing and being resilient. 

Nevertheless, in my opinion, we keep up through embracing change; change such as learning new things, seeking multidisciplinarity, and connecting and finding collaborators in open-mindedly colleagues with different skills to those we have.” 

“I learn the most by doing and interacting and being inspired by others – by their passion, their ideas or by sharing their experience. I found some wonderful collaborators during this project. I have this summer been part and gave an invited keynote talk to an AI audience and while I am not yet ‘an expert’,slowly without realising I got back to enjoying what a scientist should enjoy most – solving problems and finding solutions but it is a very long, many times really tough and meandering process.”