[00:00:22] Bridget Scanlon: I'm pleased to welcome Alex Sun to the podcast. Alex is a data scientist currently working for the Department of Energy, National Energy Technology Lab. And previously Alex worked as a senior research scientist at the Bureau of Economic Geology at the University of Texas, and we both work together at the Bureau.
Today we're going to focus the discussion on the current status and future directions of artificial intelligence and machine learning in our sciences, AI ML Alex published an excellent review article on how big data and machine learning benefit in the environment and water resources management in environmental research letters in 2019.
Thank you so much, Alex, for joining me today.
[00:01:15] Alex Sun: Yeah, sure. It is my pleasure.
[00:01:17] Bridget Scanlon: So Alex, you've been working with AI ML for over a decade and now, and recognize the value of it. And we are all beginning to recognize it more and more. And I guess it's just all around us. Some people think maybe we can just turn it off or whatever, but I think it's too late for that.
But even driving around, you see these self-driving electric vehicles and people in their homes, they have Alexa and they're asking questions all the time. Maybe you can describe how some of the concepts for AI ML are used. Say for example, for the self-driving EVs or things like that.
[00:01:56] Alex Sun: Yeah, sure. So,self-driving technology, that basically is a perfect example for demonstrating how AI and ML is affecting our everyday lives, right? So, if you look at self-driving car for example, a Tesla, then AI ML actually everywhere in the car and they work altogether to help the car, drive under this super region mode, right?
Full self-driving cars. And so basically we talk about AI, let me probably first define the terms. So, we often use AI and ML together.
So AI basically is just a generic term referring to the use of computers or machines to imitate human-like behaviors.
And then ML is actually a branch orsubset of AI that trains models to learn or reproduce human behavior or human thinking to solve the real world problems.
So in the case of self-driving car like Waymo or Tesla, then the AI ML basically work in coordination with the hardware, right?
First, the hardware like LIDAR system and GIS system and onboard cameras, they capture the real time information. And then so the online models, for example like convolution neural network models can be used to detect the real objects on the road and for example, pedestrians and road signs and traffic lights.
So then that provides the real time feedback to onboard processors and the processor can then make real time decision on behalf of a human being. And so that, that's a really almost one of the most complicated AI ML systems out there. So, basically kind of reflect most aspects of what we have achieved in the world of AI ML, from the hardware and software. So this is a great example to basically illustrate how AI ML has influenced our everyday work. Of course, we have many other examples like, the smart home Alexa system we just mentioned, and even our smartphones, which can do much more these days, So we're basically, living in a world full of those smart things and it is just impossible to turn them off like you mentioned.
[00:04:40] Bridget Scanlon: It's interesting, you ask anybody in the past, if you go to New City or whatever, you might ask somebody where such and such a place was or whatever. Now everybody just looks at their smartphone and even you go to the grocery store and you say, which aisle is blah, blah, blah and they pull out their phone.
So, it just seems like we, it's making us a bit dumb. We don't remember any phone numbers anymore. I hardly remember anybody's name or anything because I can search in my outlook and figure out what the name is, we're losing some of our memory and outsourcing it to smartphones or to other things.
So, Yeah, thanks for explaining these electric vehicles and how they operate. And another example is a friend of mine was going to a conference and she booked one of these through Uber, but it took her to the wrong conference center. But, I think large language models then are used, I think, to respond to voice commands so, I think we are getting more and more comfortable with these things and maybe somewhat or maybe a bit too comfortable.
And then in the homes, Alexa, people have conversations and asking them all sorts of questions is sort of addictive really, to be able to get information so readily in an easy way.
So Alex you have seen artificial intelligence and machine learning evolve a lot in the last more than the past decade.
And maybe you can describe a little bit to us, the different stages it has gone through and how AI and ML evolved along with the computer infrastructure that also evolved to support it and how it was this marriage between the two that has made some of these things possible today.
[00:06:26] Alex Sun: Yeah, sure. The research of ML actually started a time ago. So then, these days we tend to call those like classic ML. So the modern, deep learning era actually started probably like around 2010.
Then between 2010 and I would say, 2016 we see this like algorithm dominated era. So, many deep learning models are developed during that time, such as the Convolution Neural Net and LSTM. LSTM actually was first developing 1990s.
But all these models got pretty popular in that time. And then, starting like around 2019. Then we have this attention-based transformer-based models, right from the natural language processing and that, that's from this famous paper, Attention is All you Need, I think, that was published around probably 2018. And then after that then the transformer technology actually has evolved quite much. And essentially is basically a backbone for all the modern generative AI models from language processing to the vision transformer models.
So, then in 2020, actually just up to, a couple years ago then we started to see the foundation models. So foundation models basically are generic term that refer to those large capacity machine learning models, for example transformer based models. But its unique features are first it has a really large capacity because typical foundation model may have billions of parameters versus the traditional models that we often use before the deep learning era. right? So then, the foundation models actually have created a lot of hype or, excitement, virtually in every field, Now everybody's using the ChatGPT or Copilot. So then, starting a couple years ago we see this huge interest from the commercial sector, from all types of investors that want to push this generative AI forward, So we see this almost like a fierce competition among the big technology firms like, Meta or Google and NVIDIA and as well as startups from all over the world to compete for the best large language models. And then we basically see those models like they release a new model almost like every week, or, in a never seen before speed. So then in recent years, we keep pushing the barrier or for those large language models, basically we want the artificial general intelligence or AGI, so that's basically the thing that's really close to human reasoning capability. So basically, in a recent report published by this tech company called Gottner, it mentions that 2026 is a year for, AI as an infrastructure.
Basically that means that AI is basically just the Microsoft operating system, that's part of everything. And then it is basically pretty normal to work side by side with some kind of AI driven system. The development is pretty fast paced.
Nobody, 2 years ago probably nobody had imagined this kind of fast pace. And also, at the meantime, it's actually the hardware, for example the GPUs from NVIDIA and TPU from Google, and many other startups that basically push all these things forward. So to train bigger and larger models we need not only more data, but also faster and more powerful hardware system. So it is these waves of algorithm renovation plus data, hardware renovation, that kind of driven the AI boom we have seen in this past decade.
[00:11:01] Bridget Scanlon: Yeah. And you mentioned, that you didn't think that anybody could envision this 10 years ago or 15 years ago and I think that's correct. So you mentioned a number of different things there, Alex, the first phase you were talking about in 2010 to 2015 16, you talked about deep learning works and so I guess shallow learning is that like random forest or things like that.
And deep learning is you got a number of different layers and it's much more hierarchical structure or, what's the difference between what's new about deep learning relative to what we had before deep learning.
[00:11:40] Alex Sun: Yeah, that, that's a great question. Shallow learning it's mainly about the model capacity. So, shallow learning, in the classic machine learning world, we have this support vector machine or random forest, or even, some kind of linear regression, those are considered, the classic machine learning models, they typically have a relatively small number of model parameters.
Then in those deep learning models, for example, the convolutional neural net, it basically uses this big stack of feature extracting layers. So we are talking about many of those extracting layers with different reception field. They work together to extract information from the image pixels at different scales, right? So that's why the term you just mentioned hierarchical feature system, that kind of automated this feature engineering, because before it is like, there's a lot of manual work that tries to find the best feature or predictors that go into the model.
But with this deep learning models, this feature engineering it's semi or fully automated, so we provide image or big dataset to those models. And then through this hierarchy of different layers, model layers, then we kind of extract a compact representation of the high dimensional object in the raw, in the original space. So this is instrumental actually for the modern AI because the machines are better at understanding this compact representation of data, high dimensional data or high dimensional image, right?
So if we can convert all the data objects or high dimensional data science, especially in the earth science into this compact representations, then we have this potential to learn the underlying, they call the latent representation or embedded that's sort of a shared by many different objects, right? So as a case example, right? Let's say if we look at the river basins in the US, they all look like very different, different climatology, different topology, right? But if we use a deep learning model, then we are able to project all this heterogeneous information in the raw space to a latent space that's relatively small in the dimension. But then, that allows deep learning model to learn all these different things more efficiently. So at the end, we can find this generalizable mode that represents the common sense behind all those different basins, right? So that this is very important because it basically enables a single model to be trained and then transfer to many different contexts.
In other words, this is what the term generalizable or generalized AI means, right? So you train a single model that becomes like a, almost context agnostic, then you can apply to basins or to other places where they're not originally included in your training data set.
So the, automatic feature extraction and hierarchical learning, that's a basis for many of the models we are using today.
[00:15:34] Bridget Scanlon: Wow. And the other term that you mentioned that we will get back to later is the LSTM or long short-term memory, and we'll get back to that application in hydrology soon. I mentioned earlier, Alex wrote an excellent review article for Environmental Research Letters in and published it in 2019.
And that was back then, if you were doing that same review today, maybe you can describe a little bit how you would use the AI ML a bit more to do that review and maybe have your agent do half the work.
[00:16:08] Alex Sun: Yeah. So in 2018, we published that review paper but that kind of reflected literature actually five years before that. Since that time, many things have happened. As I just mentioned like several minutes ago. So if I were to write another, like a review now, I think there many, notable things.
First thing that comes into my mind, of course, is the large language model and the foundation models behind it.
So foundation models basically are those models trained with large capacity, but trained with using large volumes of data. So that basically formed basis for emerging behavior. When your model is scaled to a certain level, then you have the emerging behavior that the model becomes suddenly becomes smarter, right? So then it is able to handle many different tasks, including the tasks that you probably didn't even include during the training, right? Again, using the ChatGPT now it is actually trained with text prompts, They search all the texts available on the internet and many other open source texts and then use that to train the model. But after training the large language model behaves like almost like a very smart person. It knows a lot and all based on the context and the natural language and the relations between all these different contexts that it learned during training. So that represents the most exciting development since 2019 review paper.
And then the other thing is the modern AI has actually gone beyond the natural language processing the thing that large language model was originally designed for So this is, we have many VLM or Vision Language Model so that can handle not only text, but videos and images. So we call those the multi-model data. I think that's exciting and very relevant to earth science in general because in earth science, we are facing a lot of this data sparsity problem.
In general we don't have many stream gauges or other type of sensors deployed in the field. On the other hand, we may have remote sensing data, but the resolution sometimes is coarse. But if we can somehow combine the text data numerical data and also the image data, then that basically can significantly augment our capability to answer things and do the tasks. And now we have not only better interpretation of the subject of study.
For example, that during flood time, the trained foundation model may be able to recognize the flood level image and also may search the social media for the text. And then the model can synthesize all these different sources of multi model data to arrive at some kind of real time decision support for the evacuation. So this is very exciting, especially for earth science and hydrology.
And the third one is also related to the new capabilities showing by this new generation of foundation models, right? So this is like a GenTech AI. That means we can train AI agents to do different things. For example, in the typical GIS analysis task, for example, we need to gather different maps from different sources. And then we want to click some sub area for the area of study. And then we manually look at those different layers and then perform some calculation and then get some conclusions. So, about where to, for example, to find new mineral sources or where to drill new wells or, stuff like that. But with this GenTech AI, it is basically AI augmented pipeline system.
It helps us to first note, divide this whole task into many sub tasks, right? And then this is theory called the planner in the Agent AI world. Then the planner delegates the subtasks to different agents. So, for example, the GIS analysis agent that can extract or gather and extract the maps.
And then the analysis agent can then analyze the maps including extracting the object directly from the roster maps to find the objects were interested in. All of these agents, when they work together, they basically can replace many of the systems we are using today.
So this is very exciting on one hand, but it's also pretty scary on the other hand because if you see in the news these days, many people are concerned about the traditional jobs, whether they will still exist if we deploy the Agentic AI everywhere. So those are the, major things I can think of, since we wrote the 2019 review paper.
[00:22:02] Bridget Scanlon: Yeah, that, that's fascinating, Alex. And my impression of AI and machine learning, is that you can throw the kitchen sink at the problem, and AI can handle so many different types of data. And you mentioned, text and images and social media.
And when we had the flooding in July in Texas, and we had some meetings afterwards with David Maidment and some other groups, one of the groups was talking about a system in that they were developing in Virginia. GPS system that you use in your car and that it would help with the flooding.
And so if people saw the flood level or whatever, then they would feedback and then the next people would know like people say, where there's police and, and the feedback you get from crowdsourcing data or other things. So I really, it's, it is extremely powerful.
And and then, this Agentic AI seems like you basically have an army working for you. And a colleague was describing to me recently that they were using Claude Code development, and describing in detail what they wanted it to do. And so then it would just go out and instead of just writing code in numerical code in the traditional way, it would describe it in detail.
And then it would the Claude Code developer then would go through all those steps and and do it all. Of course, you have to verify all that sort of thing. And I think that gets back to the review article that you wrote, Alex, which was Big Data and AI ML, we need a lot of data, but it can handle a lot of data.
And so I guess one of the things that you mentioned is the flooding, so operational flood management and evacuation and all of that. It seemed like we cannot rely on the traditional approaches and the latency might be too much when we are. And we really need AI ML to handle all of the remote sensing data and all of the data that's coming in from social media and all of these things to try to manage hazards like flooding or earthquakes or all of these things.
[00:24:13] Alex Sun: Yeah. Yeah, that, that's right. So flooding is actually very, one, one of the most I would say developed area because of, of course, because first because it's the huge damage is economic damage causes every year, right? And this is, we see interest not just from the academic world, but also from the like big tech firms like, such as Amazon or Google, right?
So I, I think Google now runs this thing called the Flat Hub, where it basically you know aggregates information from many different sources at a global scale, so this is a very important application word. The hazard mitigation and to potentially save many lives.
So this is also very exciting in terms of the, AI's social impact, right? So it's not just for explaining certain natural phenomena, but also to be used in operational manner to also reduce latency related to the traditional operational systems, right?
So I think that's the application of AI that's something that's going to continue to add value to our everyday life.
[00:25:32] Bridget Scanlon: Yeah, and because sometimes when I look at insurance reports and stuff like AON or others, and they'll say we might have more hazards, but the mortality is not as high as it used to be, and things like that to help reduce anything that we can do would be valuable.
And you mentioned that people are concerned about jobs and AI ML taking jobs, but my feeling is that more will be expected of us. AI and machine learning will free up our minds from certain tasks, like maybe data cleansing, wrangling, or whatever, and, preparing data and writing code and things like that.
But then we'll be expected to do more because we can never seem to get ahead. Long ago I would have to go to the library to get my papers to read, and now, they're readily available, but there are more of them, and so they expected to read more or whatever. So I'm not too concerned about job loss, but maybe certain jobs may be more vulnerable than others.
[00:26:31] Alex Sun: Yeah. Yeah, maybe. Yeah. Definitely myself, I have, benefited from the improved productivity because use of AI and all those chatbots to come up with some basic code and stuff like that. So that's a plus, but I also agree with you that we probably have more work to do to safeguard the results of AI.
So as a domain scientist, I think that's a very important job to look after results or whatever outcomes generated by AI and make sense of the, AI generated decisions. And I think by providing this human feedback to the large language model reasoning system, I'm hoping that we can generate more useful and more valuable information than before. In the old style, we look at paper, we do some research, and we think ourselves. But now it's like many virtual agents that work side by side with us and help us to think more comprehensively. And to me, I think that's very important because there's always something that we don't know.
So, I think that's a very important aspect going forward.
[00:27:54] Bridget Scanlon: And the kind of a side issue, Alex, since we are both hydrologists is the resource intensity of AI ML, how much energy, electricity it uses, how much water it uses. And I was just looking the other day and one paper suggested like 10 to 50 medium size large language model queries would use up about a half a liter of water and maybe 50 watt hours of electricity that would be, equivalent to running an LED bulb for about a few hours. People are scrolling 24/7, you see it all the time, and I don't think people are aware. And maybe we can manage it with large data centers and electricity demands and stuff like that. But I think maybe with time, maybe people will become more cognizant of the actual resources that are used for these processes.
[00:28:46] Alex Sun: Yeah. So that is indeed a major concern right now. The energy consumption that's needed for training all those large language models, it takes millions of dollars. That's what I heard to train a large language model. And it's mainly, related to the computing hours. You basically need a lot of computing power to process all the data. So that represent, a significant power consumption. It's just like you were saying there, we ask questions, but without realizing how much power was used to train those models.
That's also a major bottleneck because in terms of investment, I mentioned foundation model, one of the features of emerging behavior, like by scaling up your data, you start to see this new behavior that's not observed when the data base is small. But more recent findings are, like this scaling up, has, probably hit a plateau. After the companies crawl all the web text and use all the open source code repositories or documents they can find, then there's basically no more good label data to use to train those models.
So then this is what they call the scaling up plateau. And it is like the major bottleneck for basically going to the next level of AI. And with this scale of the model we're seeing right now. If you spend more CPU or GPU to train those models, they return is actually marginal.
And I think that's dilemma, not just for the companies, but also for the investors. So, that and combined with this energy consumption. I think those are the major concerns of the AI ML developers right now. How to reduce the power consumption and how to break the data limitation. I think that's very active area of research and the outcome of that research basically will lead to the next generation of even smarter AI models.
[00:31:11] Bridget Scanlon: Alex, one of the things I think of is, there's so many things happening and as you mentioned, it's very dynamic and things happening so fast and you just feel like saying, oh my gosh you just throw up your hands and you say I just can't keep up but one of the things when we were chatting the other day you mentioned is that some of the fundamental AI concepts and ideas and stuff very powerful and they haven't changed. And maybe they've been modified a little bit. Like you mentioned, long short term memory or random forest or variations of these. Maybe you can describe a little bit about some of those fundamental aspects and how they're still there and that we may be looking at variations of those, but it's not just off the scale, out of the box all the time, changing so much.
[00:31:59] Alex Sun: Yeah long short term memory model is actually pretty interesting, I think the original concept was probably introduced in the mid-nineties, So there's the reason it got popular in hydrology because a series paper published by, some US and German researchers.
So, they basically took this US national gauge dataset, and then they show that they were able to train this single LSTM model to basically give reasonable predictions for all gauges in that dataset. So since that time, we have seen new researchers have tried newer model architectures. Like, transformer but still the LSTM, represents, at least very close to state of the art performance. I think that's because in terms of architecture the design of LSTM resembles actually the hydrological processes. It represents the storage component the memory of the hydro system, and also the delayed response because of the the rainfall runoff and routing process. So it has this different components that, mimics the physical process so that basically enable it to model this rainfall runoff process effectively.
I think that basically also echoes what you just mentioned that the newer is not necessarily better. So, I think especially for hydrology, and the natural sciences in general, we have a lot of this physics informed paradigm, data driven, but we want to incorporate our physics laws and prior knowledge into the modern process. That way we learn not just data itself, but also the underlying process, that sort of makes more physical sense to us.
I think in terms of application all these models can coexist. Depending on the context of the application, we can choose one model from many possible candidates. So that way actually we can also reduce power consumption, right?
Instead of training huge models that require a lot of GPU hours,train a small model efficiently while, providing essentially same accuracy. So I think this capability to choose from different model architectures or even to combine different model architectures represents a very interesting research that's often overlooked because, in academia, it is oftentimes that we have to publish a new algorithm that beats an existing algorithm, right? But on the other hand, we can train many of the existing models and use the ensemble of those models to, give us very reasonable or actually sometimes even more accurate answers.
This is basically the classic concept of ensemble learning. So we use the ensemble models to generate more accurate predictions. And also give us this capability to quantify the uncertainty in the model predictions.
Going back to the model question I think basically we can live with different models and we just have a much wider set of tools to choose from these days.
[00:35:50] Bridget Scanlon: Yeah I think, you mentioned LSTM, long short-term memory and how that mimics the physical system. So you input the rainfall and you estimate the runoff, or you could estimate the soil moisture or whatever. But that the memory in that system then replicates what happens with the rainfall infiltrates into the soil moisture, and then becomes runoff and things like that.
So that's a really powerful approach and that's why it's very valuable, probably continue to be valuable forever. And then, a lot of people say AI ML models, data-driven models, black box not truly interpretable. But then you say, okay, let's bring in some of the physical constraints.
So even if the data suggests that the water's moving uphill we just say the water needs to move downhill. And put in some of those constraints and have, a physics informed machine learning. So, and then even the data-driven models, may give us some ideas and stuff and concepts that we wouldn't have thought of from the traditional physics models.
And then you mentioned, the ensemble, if you want to run a thousand models or whatever, the traditional physics informed models might take forever. And reduced order models, more simple models or are data driven or, we might be able to run a thousand and then we might be able to get much better uncertainty quantification with all of those models looking at that large ensemble.
And also may be able to sample extreme events because seems like what we are dealing more and more in hydrology is extreme droughts, extreme floods, and trying to sample that just from a single set of observations from a region. What you have experienced is just one example of what could have happened.
But with the ensemble then you can generate a zillion of them that are consistent with what you've seen and all of that sort of thing. So another example, Alex, where we really need big data in AI/ML is remote sensing. Maybe you can describe a little bit the value of it.
Oftentimes remote sensing data, satellite data, we have, some data that's high spatial, but low temporal resolution or vice versa. And we are trying to combine these, or, we're trying to look at land use, land cover change, or stream flow, all of these different types of parameters.
Maybe you can describe that a little bit.
[00:38:12] Alex Sun: Yeah, remote sensing basically represents a big source of big data. That's something, we actually focus on a lot in the 2019 review paper, right? So, like with so many satellites. And the long-term archives from many of those satellites like Landsat
So, they represent very important source of information for earth but traditionally the remote sensing faces several challenges, The data retrieval problem and the missing data problem, and also most importantly the data resolution problem.
So, many data they come in different resolution and representing different techniques. So the main question is how to homogenize the remote sensing data. So, that's something that the recent development of foundation models can do very well.
And for example, Google recently published this AlphaEarth foundation model. Basically, it looks at all the new different satellite data collections that it already has through Google Earth. And then, it combines all this data at a very high resolution and using the representation or embedding technique that I mentioned at the beginning.
So that's something all the deep learning models use. So they basically project all the remote sensing data they have, from Landsat, from other sources, and then they divide the whole planet into a very high resolution. Two meters by two meter pixels.
And then for each of this cell they are able to learn the low dimensional representation. Which is basically a vector that reflects all the remote sensing imagery that they have. So basically it's significant down scaling process.
So by using all the different sources of data they homogenize it and it comes up with this very effective representation. So then that means for the local or downstream applications well we can take the latent representation map, and then we can use that as our predictors and then we can, train our own model to achieve certain tasks. So that way it basically establishes a bridge between the huge amount of remote sensing data and also the local applications, right? So this is a very good example of how big data and large data help the local scale applications. So right now, I think there's a lot of interest in these applications.
For example, Google itself also published the Google Earth's AI model. So the model is built on the Alpha Earth, the foundation model, but it also combines their own Gemini large language model. Then in that sense, we can ask new questions, just like what we would ask ChatGPT or other chatbots. And then the underlying system Gemini can do some spatial reasoning and automate a lot of tasks. All through the natural language query right from us.
So, that basically opens the new window for downstream applications like, basically significantly reduce the time it requires to retrieve all the satellite data, and also it provides a very intelligent way to do some applications using multi-source data.
I think in next couple years that will dominate many of the applications, especially related to the hazard mitigation. So by having this high-resolution representation of the Earth, including, climatology and vegetation and land surface, we can significantly extract more information about upcoming events. For example, during flooding. So yeah, that's very good direction of new development in terms of AI.
[00:42:50] Bridget Scanlon: And I think that kind of reflects if you are working on any problem and you focus on a single set of data, whatever, people would always say 'Well, have you considered these data and these data?' And so by providing this, then fusing all these different data sources Google has provided a platform, then and then we can use that to look at, as you said flooding or wildfire risk or other things.
And if you don't use data like that, then people would say you're missing the boat. And with a lot of different types of data sources. And I think it may be another example would be NASA’s Harmonized Landsat Sentinel (HLS) data, where they combine Landsat with a high spatial 30 meter and low temporal 16 day resolution with Sentinel, that has low, lower spatial and higher temporal resolution. And providing these data for free is just incredible, and allows us to advance the science. We forget that we didn't always have these things. And so NASA and European Space Agency and all of these different groups that are providing these data is just like a Christmas every day.
Is there anything else that you would like to.. AI ML is such a huge topic, Alex, and I really appreciate your tackling it and it's great to see all the work that you have done over the years in this area. And I think what we've tried to do in this podcast is introduce some of the concepts and describe a little bit the evolution and maybe some ideas about where it's going.
And we also provide highlights on the website with links to papers and data sources and definitions and stuff for people who are interested in those details Is there anything else you would like to add at this point, Alex?
[00:44:39] Alex Sun: No I think we have covered a lot and like you mentioned the AI ML is such a big field. I'm always hesitant to give any type of talk because it is just too much. But anyway, I'm hoping the information and the topics we cover today can at least benefit our peers or colleagues.
So in that sense, I'm very glad that you gave me this opportunity to talk about something.
[00:45:08] Bridget Scanlon: It sort of reminds me, Alex, I must, it just makes me laugh. I wrote a review article many years ago and one of the reviewers of the article said I didn't know much about this topic to begin with, but I knew even less after I read this, your review article. So I hope we haven't created that situation.
But our guest today is Alex Sun. Alex is a data scientist working at the Department of Energy, National Energy Technology Lab, and previously worked with us at the Bureau for many years. Thank you so much, Alex, for your time.
[00:45:42] Alex Sun: Yeah, thank you for having me.
