Each year the AAAI recognizes a group of individuals who have made significant, sustained contributions to the field of artificial intelligence by appointing them as Fellows. Over the course of the next few months, we’ll be talking to some of the 2026 AAAI Fellows. In this interview, we met with Yan Liu, University of Southern California, who was elected as a Fellow “for significant contributions to machine learning and development of widely recognized models for time series and spatiotemporal data analysis”. We found out about how time series research has progressed, the vast range of applications, and what the future holds for this field.
Could you start with a quick introduction to your area of research?
My major research area is machine learning for time series and spatial temporal data, interpretable machine learning and physics informed machine learning. I work on different types of models for time series data and making them interpretable. My research is closely connected with real-world applications, in sustainability, healthcare and science.
How has your time series research evolved over the years?
Almost my whole career I’ve been working on machine learning for time series, and we develop different types of models to address some of the important problems in time series analysis, such as forecasting, causal analysis, anomaly detection, and building generative models. In a nutshell, time series refers to the data that is recorded in a sequential order. We then want to analyze the data itself as well as the relationships within the sequence and the order.
I would say that 10-15 years ago, time series was not a major research area in machine learning. It was mostly statisticians and economists working on these problems, developing simple models based on statistical-based approaches.
I started working on the problem more than 20 years ago, looking into how we could utilize machine learning-based approaches to address the problem. I started with state-space models, where we were able to model the inherent latent space and transitions between different timestamps. Then I moved on to Granger causal models where we tried to model the causal relationships between time series. Then later, when neural networks came in as a main powerhouse in machine learning and AI, we started to look into how we could utilize machine learning models. For example, long short-term memory machines (LSTM) and recurrent neural network (RNN) models for time series forecasting. Later, I started to work on graph convolutional neural networks (GCNs) for modeling time series and spatial temporal data.
More recently, we built a general-purpose foundation model for time series so that we’d be able to forecast anomaly detection or other types of time series analysis tasks in the zero-shot case. Zero-shot means that you can ask for a weather forecast, or a future stock price without any training data in that domain. Previously, our time series models utilized a large amount of training data, but now that we are able to do this zero-shot or few-shot. This basically means that we’re able to build a model with very high generalization capability. These models can do fine-tuning to quickly adapt to the specific applications, and they achieve similar (or sometimes better) results than building the models using a large amount of domain-specific training data. I think that is a really major breakthrough that we’ve seen in the past few years.
What are you working on at the moment – what’s the next step in this journey?
Currently I’m working on a physics-informed time series foundation model. This means that we try to incorporate physics knowledge into the foundation model so that it conforms to physics principles or partial differential equations (PDEs). We can use them for scientific fields where the model has to conform to the physical principles, constraints and equations in that particular application. Our aim is to develop this next generation simulation model that will be accurate and be able to quickly adapt to the specific domain that experts are working on. This will be particularly valuable in applications where there is limited data and the current simulation system doesn’t do very well.
Right now, industry is working really hard on deploying time series models into real-world applications. For example, Google, Amazon, and Salesforce all have specific teams developing industry solutions based on the developments in time series foundation models from the past few years. But I think in the academic world, the next frontier is really trying to develop next generation simulation models to address breakthroughs in the scientific world.
Figure from Yan’s recent work on physics-informed foundation models showing the PINFDiT Architecture. (A): PINFDiT framework with diverse multivariate time series from different domains with multi-resolution or missing values; (B): Detailed structure of PINFDiT block; (C): Physics Injection: employing physics knowledge as a plugin mechanism, injecting known physical residuals F to refine predictions without requiring architectural modifications or retraining.
Could you talk about some of the application domains you’re working on?
We’re working actively with Earth scientists on transport flow modeling. That is basically studying underneath the ground, the rocks, the dirt, and then how fluids and air actually flow underneath and between the rocks. The Earth scientists have a very limited amount of data, because we don’t have sensors everywhere underneath the ground to know what’s happening there. This lack of data means that their models don’t perform very well. We hope to be able to use our next generation AI simulator to help.
Besides this transport flow modeling, we work on climate applications. For example, weather forecasting. There are lots of people working on this application so we are able to compare our results to benchmark datasets, and our models perform very well.
We also work with experts in structural biology. We want to see if we can actually use our model, which also needs to satisfy all these molecule physics constraints, to help with the molecule design for drug discovery.
In another application, we work with domain experts in transportation systems, modeling traffic flow. We built this graph convolutional neural network for traffic forecasting, and for demand forecasting for ride sharing. This relies on the condition that we have a lot of data. We were able to get very, very good results and our model was deployed by companies. It has been one of the mainstream approaches for many years for traffic and demand forecasting for ride-sharing companies. And now we think that with these new generation simulation models, we can actually solve even more problems than before. Let’s say we want to model how the traffic flows within the whole city, or build the simulations for traffic light control. The current simulation system in traffic and transportation is still not very good, but that’s another domain where we think this can potentially play a major role.
I’m interested to hear a bit about your career path to this point. Could you say something about that?
I got my PhD from Carnegie Mellon University (CMU). And so I joined at an interesting time, in 2001. Two major things happened then: one was the burst of the internet bubble, and the second was September 11th when the economy had a major breakdown. Despite the IT bubble bursting, there was actually something fundamental and exciting under development at CMU, and that was machine learning and AI. When I went to the campus for the orientation, the buzzword I heard was “machine learning”. I had no idea what it was, but everybody was talking about it, and I felt like I should do my PhD on this topic. And that’s where I got started. We published papers in the main machine learning conference at the time, the International Conference on Machine Learning (ICML). I remember that there were around a couple of 100 papers, but a third of them came from CMU. It was really the place to study machine learning at that time.
My PhD thesis was on machine learning models for protein structure prediction. DeepMind has famously worked on the protein structure prediction problem more recently. However, at the time of my PhD, there were only around 2,000 protein structures in the protein data bank. And only less than 100 new structures were added to each CASP competition. There were a very limited number of researchers working on that problem at the time. But things changed quickly and by the time DeepMind were working on it there were hundreds of thousands of protein structures out there and researchers could utilize the deep neural network architectures and many other advances in machine learning.
After my PhD, I joined IBM Research where I worked on a variety of different types of problems. And that’s where I encountered time series data. I started to look into this more deeply in terms of the right solutions to work on to solve this particular problem, which has so many applications. I stayed at IBM for three and a half years, before deciding that I wanted to go back to academia and establish a research group fully devoted to the specific area of time series. It was 2010 when I moved from IBM research to University of Southern California (USC) to start my research group and try to build the workforce in this particular area.
I’ve also done a couple of sabbaticals. The first one was with DiDi (the Chinese Uber) where I developed graph convolutional neural network solutions for traffic flow forecasting and demand forecasting. Here I saw my research work translate into real-world application deployment. And then later, I did my second sabbatical with Google Cloud AI, where together we developed this prototype time series foundation model. I think that each exposure to different environments has given me new thoughts about what should be the right directions and what are the right approaches to addressing these fundamental problems.
I’m now working with Amazon as an Amazon scholar and we’re working on solutions for supply chain management using the time series models. So I would say that it’s really an exciting journey and I took a lot of detours, but I really enjoyed every turn, which opened up new directions and thoughts for addressing fundamental problems.
What are some of the open problems in the field?
Many recent achievements in the field of time series have relied on the development of sequence models and large language models. We really have been enjoying this and have made good progress. However, there are also all these fundamental problems that we have not solved, specifically in the time series domain. For example, some of the different types of time series we work on have a lot of missing values while we seldom see missing values in a sequence language model.
There are also multi-resolution observations where we get observations per minute, per second, per hour, per day, per month. The challenge is being able to utilize all this information together and build a more accurate model. That’s what we typically call a multi-resolution time series challenge, which was not appropriately handled by a language model.
Furthermore, causal analysis is very important in our domain, but foundation models for language seldom consider causal models, especially accurately inferring and modeling the causal relationships. This is really a fundamental question in time series analysis.
What are you excited about working on for the next five to ten years, and how do you see the future of the field?
I think in the future, we will open the door to the physical world. So one observation I talked about with my colleagues and students is that I think languages are really an abstraction of the physical world that we’re in. It’s a kind of a vehicle for humans to understand and communicate with each other, but does not fully represent the physical world. Language by itself is an abstraction, it’s fuzzy, but when we look into the physical world, it is really accurate. I call what we work on precise AI, in the sense that we have all these real data values from sensors, and this will help capture a more accurate characterization of the physical world that we’re in. That’s the foundation for us understanding the physical world, uncovering scientific breakthroughs, new materials discovery, robotics, and so on. This is all in the physical world, it’s not in the virtual world.
From that perspective, time series is a very exciting problem that is really understudied. This is a fully interdisciplinary problem and needs experts from different fields to address it. So I think that’s really the future: physics-informed AI models – you can call this a “world model” if you like to use the buzzword. There will of course be ups and downs in this particular search, but it shouldn’t stop us from pursuing this direction, and the future is bright. It’s just a super exciting time, and hopefully you will see more breakthroughs from us.
About Yan
Yan Liu is a Professor in Thomas Lord Department of Computer Science, and Ming Hsieh Department of Electrical and Computer Engineering in Viterbi School of Engineering. She is the Director of the Machine Learning Center and co-director of USC Institute in Ethics and Trust in Computing. She received her Ph.D. degree from Carnegie Mellon University. Her research interest is machine learning for time series and its applications to geo-science, health care, and sustainability. She is a fellow of IEEE and AAAI, and has received several awards, including NSF CAREER Award, Okawa Foundation Research Award, New Voices of Academies of Science, Engineering, and Medicine, Best Paper Award in SIAM Data Mining Conference.