Category: Uncategorized

High Altitude Balloons (HABs) have long been used to perform atmospheric measurements and Earth observations. With the growing availability of commercial off-the-shelf (COTS) hardware and open-source software, HABs have been increasing in popularity to rapidly prototype and test concepts in near-space conditions. In this case, near-space conditions refer to the conditions seen at the lowest altitude required for a satellite to begin orbiting the earth. The Innovation to Flight (I2F) program at the Jet Propulsion Laboratory (JPL) has developed iterations of this technology for years, decreasing the materials cost of HABs while simultaneously improving their capabilities.

This project builds upon the Zephyrus program whose previous work demonstrated the capabilities of cost-effective HABs in carrying science payloads and performing small-scale experiments in the upper atmosphere [1,2]. Low development costs and efficient prototyping make HABs ideal for inexpensive, effective Earth observation and environmental testing in near-space conditions. The primary roadblock is HABs’ intractability (see Figure 1), which causes difficulty in obtaining consistent scientific data in unpredictable weather conditions and adds restraints for receiving flight licenses issued by the Federal Aviation Administration (FAA) [2].

In this paper, we propose a first-generation altitude control system for the Zephyrus program that meets several stringent requirements. We present a scalable, ballast-based system to accommodate variable HAB payload sizes and still fit within FAA regulations. Additionally, we expect the implementation of this system to cost less than $100 in production costs for 95% of HAB payloads.

Author: Benjamin Zeng
California Institute of Technology
Mentors: Adrian Stoica and Thomas Lu
Jet Propulsion Laboratory, California Institute of Technology
Editor: Laura Lewis

Introduction

High altitude balloons (see Figures 2 and 3) are a classification of unmanned aerial vehicles whose lack of powered lift introduces important operational benefits and engineering challenges that must be addressed. By using gas that is less dense than atmospheric gasses, these balloons produce lift through buoyancy rather than through propulsion. This significantly reduces the energy required to lift the entire system when compared to powered unmanned aerial vehicles (UAVs) and allows them to spend considerably more time in flight compared to many alternatives, including the vast majority of fixed-wing or rotor-based vehicles.

However, this buoyancy-based system introduces additional complications. Most notably, this methodology greatly inhibits the aerial vehicle’s control over its own trajectory, instead relying on wind patterns for its position and velocity. Given these difficulties, this project aims to remedy these drawbacks and develop solutions to augment control systems aboard the HAB and improve descent consistency and automation of navigation. This can be divided into two categories: lateral control and descent control. With a greater degree of control, the balloon can better follow predetermined flight paths and sets the groundwork for research and development of interest areas such as the investigation of signals-of-opportunity (SoOp) in HAB networks; increasingly weather-agnostic autonomous launch, control, and landing systems; and path-planning enabling constant, consistent real-time communication between the HAB and ground instruments [1].

The contributions of this project are threefold: firstly, we describe an improved balloon flight trajectory prediction model that accounts for environmental change throughout the flight of the HAB. Secondly, we elevate control theory of high-altitude balloons to consider neutral buoyancy flights, enabling the balloon to remain at a relatively constant altitude throughout its flight, allowing for prolonged voyages across longer distances. And lastly, we present a reliable, scalable, and inexpensive design to enable altitude and lateral control systems on HABs of various sizes falling within FAA flight regulations.

Background

Balloon Physics [4]

First, in order to discuss our control system more in-depth, we must develop the basics behind understanding the buoyancy and lift principles that govern HABs. These can be determined by applying the ideal gas law PV = CMT, where M represents the total mass and C is a density-dependent constant. The conservation of mass dictates that the density of the balloon follows the equation ρ = ρ₀(V₀/V), where ρ₀ is the initial density and V₀ is the initial volume. For estimation purposes, we assume temperature is constant. Thus, the ideal gas law gives ρ = ρ₀(P/P₀)

We additionally note that the buoyancy lift of the balloon is a function of the density difference between the balloon and the surrounding air, multiplied by the volume of gas displaced by the lifting gas. We denote ambient pressure as P^A and lifting gas pressure as P^L. The overpressure is thus considered P⁰ = P^L – P^A. Lift can be expressed as:

where the approximation P₀^L = P₀^A = P₀ is assumed for simplicity. For a typical latex balloon, we assume that there is approximately 150 Pa of overpressure which only becomes significant at ~30,000 meters [4]. We also account for the fact that a change in temperature can affect the lift force. While the atmosphere cools greatly at high altitudes, we must consider that solar radiation can affect balloon temperature by as much as 10% [4]. We can use the ideal gas equation to similarly note that:

or equivalently that:

Operating principles

Using these equations, we can begin to develop a control system that dictates the desired altitude of the balloon. Furthermore, it is well-known that various altitudes of the stratosphere host fluid layers which have a variety of velocity patterns. In fact, these layers often change directions and can be used to control the balloon (see Figure 4 for example data from the World Meteorological Organization).

Our system takes advantage of these wind patterns in order to better control the balloon’s position and velocity, ensuring that the balloon’s behavior is as similar to desired behavior as possible. For example, a recent project of interest was the use of HABs in monitoring small land areas for research purposes, simulating the efforts of natural disaster observation. This requires the balloon to remain at a constant altitude, allowing on-board cameras to monitor the land area while the balloon remains as stationary as possible. The National Oceanic and Atmospheric Administration (NOAA) provides wind prediction data to the public, updated once every hour throughout the day in altitude resolutions of around 430 meters. This data could be easily used to estimate optimal altitudes in order to ensure that the balloon travels closer to a desired pathway.

Existing work

This methodology of altitude control has been proven to work in other implementations, including with the ValBal team at the Stanford Space Initiative [4]. In this use case, the team aimed to improve upon longevity of the balloon’s flight in terms of flight distance and flight time, expanding upon what was previously achievable by latex-class balloons in the atmosphere. However, we differ in that the ValBal team controlled strictly altitude with little regard to lateral position. We adapt this methodology to prioritize tractability of the balloon over survivability. We conducted research on how to expand this work to a variety of flight models (short-term to long-term flights) and desired flight paths (long-distance traversal to stationary positioning).

Methodology & Development Process

To prove the viability of our system, we created models via Simulink to test the tractability of the balloon’s altitude through modulation of the balloon’s buoyancy. The test flight information based solely on the balloon’s altitude information is available in Figure 5.

Compared to commonly trusted balloon predictors available online, such as the Cambridge University Space Flight (CUSF) HAB balloon predictor, our results have approximately 3% error in burst altitude and 31.37% error in burst time. We discovered that these commonly used online predictors failed to account for the changing air pressure as a function of altitude and the resulting impact this would have on the size of the balloon and its buoyancy. Removing this factor from our model allowed us to achieve a burst altitude and time that was comparable to that of the online model.

Interestingly, we had found that not only is controlling the altitude of the balloon possible, but also that with precise enough instrumentation, we are capable of producing a balloon that is neutrally buoyant in the atmosphere—an instrumental step in enabling long-distance flights traversing across further distances (see Figures 6 and 7).

As a result of our verification models, we turned our attention to a ballast system. This new system implements a solenoid that leaks helium from the balloon to decrease buoyancy (see Figure 8) and a ballast payload that would drop small weights that would serve to increase buoyancy of the balloon (see Figure 9 and 10). We utilized small, biodegradable plastic ball bearings (BBs) as a part of our ballast. Not only did these BBs come in various densities—which allowed a wide degree of adaptability for various payload sizes and rapidly variable buoyancies of a wide range of HAB sizes—the BBs additionally were cost-effective and readily available, allowing them to be employable across all HAB tests.

Individually, these parts have been tested and verified to work reliably across operations that simulated more than 15 consecutive flight hours. We experienced no jams with our revolver chamber-style BB dispenser system nor did we face issues with electronic control problems of our solenoid. It is important to note that these tests were conducted at standard atmospheric conditions near sea level at constant room temperature. These parts were not able to be stress tested at altitude or near-space conditions including exposure to radiation, thermal stressing and cycles, or environmental conditions. Given restrictions with FAA flight licenses, we were not able to launch and further test these components in flight during our internship.

Future Work

The extent of this project was limited in scope due to its short timeline and FAA license restrictions. There are many areas of interest and improvement that can be pursued in future generations of the program and iterative design improvements.

Firstly, one particular area of interest is to increase real-world testing scenarios and adapt the design based on individualized use cases, allowing investigators to develop a more accurate picture of the system’s capabilities in a wide set of situations. This would include variations in environmental and system factors including wind speed, weather conditions, and balloon payload sizes. Such tests would provide insight into how a balloon may be adapted and modified to fit various use cases for a variety of applications including extending the length of flights, targeting specific locations for surveillance, and accumulating more accurate weather-gathering capabilities.

Secondly, investigating system autonomy and software improvement are warranted. While our algorithm relied on simple balloon physics and buoyancy calculations, we were unable to confirm our findings in real-world testing which includes various environmental anomalies such as updrafts, downdrafts, and other localized wind variations. By improving software-hardware integration and extensively testing the resulting product, we can vary the ballast drop and helium leakage rates to achieve better degrees of control and even an extensive duration of flight.

Overall, these future research directions have the potential to expand our understanding of the proposed design approach, further improve its performance, and identify new areas of application and development. This paves the way for HABs to fulfill not only pre-existing use cases—such as for near-space experimentation and surveillance—but also for new, expanded roles that are yet to be seen. Already, they have been employed for an abundance of novel fields in recent decades including geospatial imaging, urban development & planning, and space mobility logistics. Their low cost, ease of transport & launch, and payload flexibility have long been recognized as a key selling point for their implementation within the field of space exploration. However, with the introduction of a controls system that would be capable of increasing desired control over its flight path, its true limitations of use are yet to be seen.

Acknowledgments

We would like to thank the following for their gracious support of Innovation to Flight:

Adrian Stoica and Thomas Lu who graciously offered to mentor the many students, supporting us in our engineering endeavors and constantly providing feedback and support to improve our designs and ideas. The results obtained throughout this internship would not have been possible without their assistance and guidance.

Attila Komjathy who graciously communicated with JPL on behalf of our team to arrange available lab space and enabled the team to work on lab and in-person, greatly contributing to our collaborative efforts.

Hunter Hall who provided a plethora of resources and mentorship and whose years of experience with HABs gained us an invaluable set of experiences to build upon.

And lastly, we would like to thank NASA, JPL, Caltech and the education offices for their gracious funding in support of our programs and helping make a program like this possible.

References

[1] Garrison, James L. et al “Signals of Opportunity: Enabling New Science Outside of Protected Bands,” IEEE Xplore, 2018.

[2] Hunter, Hall et al, “Project Zephyrus: Developing a Rapidly Reusable High-Altitude Flight Test Platform,” IEEE Xplore, 2018.

[3] Hunter, Hall et al, “Utilizing High Altitude Balloons as Low-Cost CubeSat Test Platform,” IEEE Xplore, 2020.

[4] Sushko, et al, “Low-Cost, High Endurance, Altitude-Controlled Latex Balloon for Near-Space Research (ValBal),” Stanford Student Space Initiative, 2017.

Interviewer: Maggie Sui

Trivia

• Favorite book: Slaughterhouse Five. I found the author Kurt Vonnegut’s works at specific, formative point in my life and it very much resonated with me. It’s this blend of sci-fi and commentary on human nature, the need to challenge orthodoxy, and the absurdity of it all. It’s always reminded to think about how the grand plans laid out by the ‘people in charge’ affect ordinary people, and to have some humility about what the role of good leaders should be.
• Favorite place: New Hampshire. I think some of my best memories are from vacationing in the small towns of New Hampshire like Lincoln or Conway when my kids were young. I’ve never been in the wintertime, but in the summer it’s just so beautiful and peaceful. There are beautiful hikes, fun swimming holes, bouldering, and lots of good fishing.
• Favorite food: Hot pot. Around LA, I’ve gotten really into hot pot places. There are just so many of them, and it’s a great meal where everyone can get what they want.
• Favorite protein molecule: ATP synthase. It’s such a beautiful structure and an incredibly elegant molecular machine. I remember being blown away when I learned about its structure and how that relates to its function. It’s also a protein complex that is essential to life.

Can you provide an overview of your research?

I am a sensory physiologist. My lab studies the synaptic and circuit mechanisms of how the brain makes internal representations of the world around us, and we investigate this in the context of chemical information in insects, mostly in the fruit fly. Chemical communication underlies all of life on Earth. Our experience as humans is dominated by sight and sound. But, for the vast majority of organisms on the planet, their main mode for interacting with the world is through chemical signals, so I view olfaction as the language of life. Olfaction also serves as a particularly compelling system for studying questions concerning how sensory information can be translated into useful behaviors and actions in the animal. In the smell circuit, high level representations of sensory objects are constructed in very few steps of neural processing, so this circuit is a particularly tractable system to study questions like how the brain generates behavior.

What are some possible applications of your research?

Our research provides insight into basic neuroscience inquiries, such as how does the brain translate environmental stimuli into behavior? How does the brain generate internal representations of the world? Such questions have important implications for understanding how to help people with neuropsychiatric or neurodevelopmental disorders who struggle with normal sensory processing. Understanding our sense of smell also has a huge potential for changing the way we live. We all walk around with little machines in our pockets that can capture visual and auditory stimuli, and then also play those back to us, but nothing comparable exists for chemical signals. Enabling machines to sense, encode, and decode chemical signals would be significant for applications as wide ranging as search and rescue in natural disasters, environmental monitoring, and medical diagnostics – there is a chemical signature for many diseases.

What are some exciting recent developments in your field?

Just in the last few years, we’ve finally seen the first protein structures for the odor receptors. The first insect odorant receptor structure was published about four years ago, and, just a month ago, the first human odorant receptor structure was published by a collaborative research group from UCSF, Duke, and City of Hope, just a few miles away. These structures are beautiful because they show the mechanistic basis of how these receptors detect their odorant ligands. One of the most intriguing aspects of this work is that, so far, the solutions for how odor ligands are recognized by insect vs mammalian odorant receptors are quite different. In one case, binding is quite promiscuous and appears to be mediated by many, very weak interactions, whereas in the other case, binding appears more selective and mediated by a lock-and-key-like mechanism. However, we just have the structure for one odorant receptor for each case, and each is somewhat of a special case. Most olfactory systems have dozens to up to more than a thousand odorant receptors, and the structures of more receptors will undoubtedly be soon solved, so in the next decade or so, it should become clear what is the more “typical” mechanism for odor recognition, or if different odorant receptor families from different parts of the evolutionary tree solve the same overall problem with distinct mechanisms. Either would be fascinating!

What aspects of research do you most enjoy?

There are so many things. I think working with students is probably at the top of the list – the back and forth, brainstorming and puzzling, getting stuck and both thinking hard and then suddenly gleaning insight and being able to move ahead – I enjoy the starts and stops. It’s not a steady process.

I also really love doing experiments and being at the bench. I think something that’s the most disappointing about progressing in science is that you get to do fewer and fewer experiments. You work so hard to be able to run a lab, and you’re excited that you can because you finally have the platform to try out all these ideas that you’ve had for years, but then you don’t really have time to do the actual experiments yourself anymore because you have to be writing all the time. So I spend more time in front of the computer than I would like, honestly, but that’s just the nature of things. But watching a student learn and grow, seeing the quality of their data and analyses steadily improve, and guiding them to discovery – that is really what I enjoy the most.

What do you most like about running experiments?

I find beauty in the biological specimens themselves. There’s a structure to everything – from what a neuron looks like to how they’re organized in layers or compartments or bundles. There’s a physical beauty to biology. I also really enjoy pushing the boundaries of what can be done working in a very small brain, so there’s also that technical aspect of pushing the envelope and trying to see something that nobody else has seen. The first time you record something, and you realize it’s the first time anyone has seen this particular phenomenon – it’s pretty amazing.

What influenced you to pursue neuroscience, and more generally, scientific research?

I was very motivated by disease and wanting to understand things like neurodegenerative disease and mental diseases of the brain. Neuroscience in the past 20 years has definitely become more dominated by circuits and systems neuroscience research, but, initially, being very motivated to try to understand the basis of diseases, I started on the molecular side. During my PhD, I worked on the biochemical pathways and the transcriptional events that occur downstream of calcium entry triggered by neural activity and how those things might go wrong in different types of genetically based neurodevelopmental diseases. We would take neurons out of animal brains, dissociate them which destroys all their connections, and regrow them in a petri dish. This allows us to have enough biological material to study the biochemical events that are happening in those neurons, but unfortunately, then we must study them outside their native context in the brain. Eventually, I became more attracted to studying neurons in their endogenous circuit context and that led me more into the systems neuroscience side of things. Ultimately, in many areas of biology, the goal is to bridge scales of explanation. I think we will eventually get back to the molecular level and how it affects circuit behavior, but tracing the etiology of a disease through multiple biological levels is a technically and conceptually challenging thing to do, but I still am very motivated to make headway on this problem and try to contribute to understanding disease states.

What are some of the challenges you’ve recently faced? How did you overcome them?

Well, I mean, there’s always the everyday scientific challenges. We usually cope by pivoting to a different but still interesting aspect of the problem if we hit a technical wall. Often we can make some progress in another area, and, in the meantime, new tools or methods will emerge that can allow us to return to the original approach. One personal challenge that I’m coming up against is realizing my kids are growing up very quickly. My older child is a sophomore in high school. So I’m realizing I just have two years left with him before he’s off and living his adult life. I’ve realized, okay, I need to balance my job with spending as much time as I can with my children. There is this perpetual work-life balance challenge that you have to deal with.

What advice would you give to Caltech students and other young students interested in research?

The most important things are things you’ll learn outside your classes. Classes are important. They give you a foundation. But a lot of times I see students killing themselves to get in that fifth class or over-uniting. However, when I look back on the things that helped me to crystallize what I was interested in, they came from outside the classroom. The key thing to figure out is what is the level of explanation in science that you find satisfying. One person’s mechanism can be the next person’s phenomenology. I consider myself mechanistic, but I don’t study things down to electrons moving in the orbitals of atoms. Each person has a level of explanation that they find compelling and that they’re excited about. For a psychologist that could be behavioral, and for a structural biologist, that’s atomic, so figuring that out is important, and you can’t figure that out without a lot of exposure to different kinds of science, lectures, papers, and people who work across all these different scales – see how they think about things. All that kind of stuff is really hard to get from a classroom, so it’s possible to have students who are very, very well prepared on paper based on their coursework, but they’re not intellectually ready to pick a problem to focus on for their thesis in graduate school. Honestly, the reason to come to a place like Caltech is the people you’ll meet. Bug people and ask them about their work. Ask them how they decided what they were interested in. Go to lectures and ask the lecturer questions. In the grand scheme of things, that is what will have a much bigger impact on the development of your thinking and the range of creativity you’ll have.

Interviewer: Arushi Gupta

Trivia

• Favorite Book: The Foundation series – Isaac Asimov
• Favorite Place: Santa Fe, New Mexico
• Favorite Food: Enchiladas
• Favorite Country: United States

Could you give an overview of the research you and your lab works on and its applications?

My lab conducts research across a variety of subject areas, ranging from climate science, sustainability, election administration, and election outcomes to more general issues about where people get their information and how information works in either good or bad ways to influence people’s attitudes and behaviors. Like most faculty members at Caltech, we take a very quantitative approach to our research. We use big data, machine learning, and deep learning to explore datasets and uncover new insights into human behavior in social, political, and economic areas.

What influenced you to pursue the fields of social/political science, social/decision neuroscience, and research methodologies? What influenced you to incorporate computational methods and machine learning in your research?

There is a lot of serendipity in how academics develop their research strategies. I did my undergraduate work as a geology major at Carleton College but struggled with certain analytical subjects like physics. Eventually, I had to switch majors, and I explored both political science and economics, largely because I had enjoyed classes in those subjects and could still graduate on time. I ended up majoring in political science because I enjoyed it. During my time, I learned a lot from both qualitative and quantitative political scientists on the faculty of Carleton, including future US Senator Paul Wellstone. After completing my undergraduate degree, I attended law school but found out that I absolutely hated it.

When my fiance started a Ph.D. program in biochemistry in North Carolina, I looked at political science and law programs at Duke University. For reasons that are still sort of mysterious to me today, I was admitted to Duke’s political science program. This is where the serendipity comes in—I hit that program during a time when it was transitioning from being a traditional qualitative political science program to a more quantitative one. It was precisely at a moment when someone with my analytical interests could learn a lot from the incredible new faculty they had just recruited. This led to my interest in studying individual human behavior, specifically regarding where people get their information and how campaigns can influence people’s attitudes and actions. This became the focus of my dissertation and much of my early work at Caltech in the 1990s.

Over the years, my interests have continued to evolve, largely because of the environment that I’m in here at Caltech, with collaborations that span computer science, climate, and sustainability science, and the interests that students have.

What advantages do you see in integrating various disciplines?

What I think is unique about being a faculty member at Caltech and working with both undergraduate and graduate students here is that we are a small, non-traditional university. I’m currently sitting in Baxter Hall surrounded by an incredibly diverse group of colleagues. In the offices next door to me, there’s an economist, a psychologist, a decision theorist, and a neuroscientist, among others. I have this very eclectic mix of people around me.

It’s very easy for me to walk over to Annenberg to work with computer scientists or to Gates-Thomas to collaborate with those working in climate science. Caltech makes it easy, and by doing so, it is possible for us to do work that spans various intellectual and academic disciplines. This not only makes our work more exciting but also pushes all these disciplines in new directions. By injecting social science concerns into computer science research and vice versa, we’re producing innovative and cutting-edge research that has a significant impact on both science and society.

What aspects of your research do you most enjoy? What about working with students do you most enjoy?

The answer to the first question is working with students. Working with students is the most enjoyable part of my day, and the good news is that it’s what I spend most of my day doing. I work with Ph.D. students, undergraduates, and postdocs on all these different research projects we have. I bounce from meetings that cover our research project with Activision, where we are using their data to better understand player toxicity in their platform, to climate science, where we are trying to understand why people do not trust climate science and how we can change that, to fundamental work in computer science, where we are building deep learning tools that can detect harassment and toxicity in social media and text conversations. I go from meeting to meeting all day with students where I hear all of their exciting ideas and I try to push those ideas into publication-quality research projects.

It’s fun and I never have a dull day. I’m interacting with the world’s smartest students at Caltech. I’m always learning something new, whether it’s some cool new AI tool that a student read about to the latest campaign tactics that are going on in the Republican primary right now between DeSantis and Trump. It’s really exciting because students bring all of these incredible new ideas and thoughts into these conversations. They are constantly stimulating my thinking about ways we can turn these ideas into impactful research. So, it’s just a lot of fun to work with students.

Can you discuss a specific project that your lab is currently working on that you find exciting?

I would say the most interesting and challenging research agenda that we have is the collaboration my group has with Anima Anandkumar. For the past five years, we have been working to put together large datasets and build a variety of machine-learning tools that can sift through the data in real time to identify trolling, harassment, and toxic conversations. This has been the culmination of a lot of different research projects that I have engaged in throughout most of my career.

The issue is that toxicity has become a rampant problem across various platforms, including social media, political discussions, and gaming platforms. We have been trying to build tools that can detect such behavior and conversations in real-time. And now that we think we can do so using current tools that have been developed (by us and others), we are trying to pivot towards answering questions like: What can we do about it? At what point do we actually produce interventions? What do those interventions look like? And how do we test them?

Maya Srikanth, an undergraduate at Caltech who graduated two years ago, presented a paper on building chatbots that can intervene in toxic conversations on social media platforms by subtly or not so subtly changing their nature. With the recent developments in large language models, we are now in a better position to build and test such interventions, not only in social media but also in other relevant areas.

What are the challenges you have faced in the execution of this project?

On the research side, one challenge is that trolling, harassment, and toxicity are low-incidence behaviors, which make it difficult to detect unless large quantities of data are collected. So, we have had to spend a considerable amount of time and effort to find very large datasets where these kinds of behaviors could be detected. A few years ago, when it was very difficult for academics to easily access large quantities of social media data, we partnered with Twitter to create a dataset on all social media conversations on their platform related to the #MeToo movement. This dataset has been very important to us and we have used it to build a lot of our tools. A while after getting that data, Twitter made changes to its policies which allowed academics to collect much more data. We took advantage of the opportunity and collected a series of very large datasets including one about COVID-19 that we now use. Although we have been able to overcome the challenge of dataset collection, it has required a considerable amount of time and effort from our research group.

The second challenge we have faced is that many of the individuals who are systematically engaging in spreading misinformation, negativity, and toxicity through social media accounts are doing so in strategic ways. They know that they are being followed and tracked on platforms like Facebook and Instagram in particular, and so they try to do things to cover their tracks. These individuals change their behavior, usernames, hashtags, and keywords in subtle ways to avoid detection, making it challenging to use machine learning tools that are trained on past conversations to identify them in real-time or future interactions. As a result, we have worked on developing some really interesting machine learning and deep learning algorithms to track changes in conversations over time to detect how these negative actors are modifying their behavior.

The third challenge we encountered is associated with the size of these datasets. We want to apply natural language processing to reduce the dimensionality of the datasets we work with to find trace information on trolls and harassers. However, natural language processing on such large datasets involving tens or hundreds of millions of social media posts is a computationally intensive task that cannot just be done on a desktop computer. To address this, we have implemented two solutions. Firstly, we have provided resources for our students and postdocs to work in Google Cloud or AWS. We have also partnered with NVIDIA through Anima’s group to use their technologies to speed up our use of natural language processing. Second, we have also developed new natural language processing tools and methods of estimation with the help of undergraduate students at Caltech who have helped make fundamental discoveries. Using tensor algebra, we have been able to speed up the computation of many of the types of natural language tools that people commonly use, reducing their computational time from weeks or months to just minutes. Despite the challenges we faced, we were able to develop solutions and make significant contributions to the research literature.

What are your interests outside of teaching and research?

My wife and I live in Pasadena and have a very small Doxon, who’s about 10 pounds, and a very large yellow lab, who’s about 75 pounds. We spend a lot of time with them because our lab in particular needs a lot of exercise. So, I spend a lot of time walking and running with the dogs. My other main hobby these days is running. Before the pandemic, I had already been a runner for a long time, but my daughter challenged me to run a half marathon, which got me serious about endurance training again. I ran that half marathon and enjoyed it and so I decided to do more of those. So, when I’m not working, I mainly spend my time hanging out with the dogs and running.

What advice do you have for undergraduate students who are interested in pursuing research, especially in your areas of interest?

The advice I always give to incoming freshmen at Caltech, or to students who are considering coming here, is to get to know your professors. The wonderful thing about Caltech, especially for undergraduates, is that it’s very small and faculty work on campus. You can easily get to know your professors by meeting with them in their offices or labs, talking to them after class, inviting them to Red Door, etc. When you get to know your professors, you’ll find that many of them are really interested in you and what you are doing. This can create a lot of opportunities for you, whether it’s joining their lab, working on a research project, or pursuing some other activity on or off campus.

Getting to know your professors is a force multiplier for Caltech students because it can really give you life-changing opportunities that you would not be able to get if you were at a larger school. Many of the undergraduates who end up working in my group and lab remain personal and professional friends and contacts long after. I still communicate with Caltech undergraduates that I got to know back in the ‘90s and 2000s, and some of my former Ph.D. students are still active collaborators of mine. So, my advice is to take advantage of this opportunity to get to know your professors and see where it takes you.

Carbonation of cement is a process that involves multiple chemical reactions between the main components of set cement and carbon dioxide such that the carbon dioxide is permanently sequestered. 0.25 gigatons of carbon are estimated to be sequestered in cement-based material annually; however, this estimate is based on a simplified model of the carbonation process. By comparing this simplified model with existing data, we show that the model works best under highly controlled conditions in which the rate of carbonation is limited by diffusion of carbon dioxide into the concrete. To improve the modeling of carbonation, we developed a finite difference analysis that included the unsteady diffusion of carbon dioxide and the rate of reaction between carbon dioxide and the various cement components. This improved model allowed us to examine a range of conditions and provided a more complete estimate of the reactions, which will ultimately enable more accurate global carbon estimates.

Author: Miles Jones
California Institute of Technology
Mentors: Professor Melany Hunt
California Institute of Technology
Editor: Wesley Huang

Introduction

Concrete has been a vital part of modern society, as it is a versatile, cheap, and sturdy building material. However, the production of cement, the main component and “glue” found in concrete, results in the release of large amounts of carbon dioxide gas. Due to the immense amount of cement produced, it is estimated that 6 − 10% of all anthropogenic CO₂ released is from cement [1].Once cement is created, however, it absorbs CO₂ as the gas diffuses and reacts with the calcium hydroxide and a calcium silicate hydrate (CSH) gel. The reaction with calcium hydroxide is as follows:
Ca(OH)₂(s) + CO₂(g) → CaCO₃(s) + H₂O (1) 
The reactions between CSH and CO₂ are: 
(3CaO·2SiO₂·3H₂O)+3CO₂ → (3CaCO₃·2SiO₂·3H₂O) (2)
(3CaO · SiO₂) + 3CO₂ + H₂O → SiO₂ · H₂O + 3CaCO₃ (3)
(2CaO · SiO₂) + 2CO₂ + H₂O → SiO₂ · H₂O + 2CaCO₃ (4)
[2]. These processes can be modeled with a simple equation as done by Xi et al.:
d = k × √ t (5) 
Here, t refers to time, k is the carbonation rate coefficient, and d is how far into the cement the reaction has proceeded [3]. 
The coefficient can be expanded, such that: 
d = (2[CO₂]₀D_e,CO2t/[Ca(OH)₂(s)]⁰+3[CSH]⁰)^0.5 (6) 
[Ca(OH)₂(s)]⁰ and [CSH]⁰ are the initial concentrations of calcium hydroxide and CSH gel respectively, [CO₂]⁰ is the concentration of carbon dioxide at the edge of the material, and D_e,CO2 is a diffusion coefficient based on porosity and water content [2]. 

This allows for the estimation of not only how fast the reaction proceeds, but also the total amount of CO₂ absorbed by cement. Xi et al. estimate that around 0.24 gigatons of carbon are absorbed by this process each year by using equation (5) to find the depth of carbonation and thus the amount of carbon sequestered. They also estimate that the total amount of carbon sequestered by cement from 1930 to 2013 is 4.5 gigatons [3].

The method by which the aforementioned estimates were found are based on large assumptions, especially for equation (5). The reaction is assumed to take place instantly and the rate limiting factor is simply the diffusion of CO₂. Additionally, it assumes that the cement is fully hydrated, which is not necessarily true. These assumptions, as well as other issues that arise from using this equation, coupled with the accuracy of some of the coefficients used led us to question if there was a method of modeling this process that would result in higher accuracy of these estimates.

More complex models not only give better estimates to carbonation depth and amount of carbon stored, but also a more complete picture of the process. The models we developed in MATLAB, based on mass transfer and finite-difference principles give the concentration of CO₂, Ca(OH)₂, and CSH gel at each position and time. Upon comparisons with experimental data from Chang and Chen [4] we find that, given the same initial conditions, our models hold up well.This process has a large impact on global carbon models, as the CO₂ fossil fuel emission estimates decrease from 9.9 gigatons of carbon per year to 9.7 gigatons if included. While this difference (0.2 GTC) may seem small, it is of the same magnitude of the net carbon budget imbalance (0.3 GTC) [5]. Thus, we want the estimate of carbon that is sequestered through these reactions to be as accurate as possible.

Methods

Comparison of Simple Models with Data
For the comparison of “simpler” models which use equations (5) and (6), we utilize data from numerous sources. Multiple papers that took measurements of the carbonation depth at differing times were used for data points. Then, the groups of data with similar strengths and storage (indoors vs. outdoors) conditions were grouped together and compared to the results from (5) and (6), using similar initial conditions. These comparisons were done in Microsoft Excel by inputting the data and calculating the predictions.

Development of Models
To develop the more advanced models, MATLAB was utilized. These models are based on unsteady one-dimensional diffusion in a semi-infinite medium. Both forwards and backwards finite differencing methods were used to create two separate models, each being derived from Fick’s Law of Diffusion and an analysis of the mass diffusing in and out of each position, as well as that used in the reaction at each point. The forwards method estimates the concentration of the species at a certain position based on concentrations from the previous section. The backwards method is more complex, and estimates concentrations from those in the following section. The initial equation is follows:
∂C/∂t = ∂/∂x(D C/∂x) − k_CHCC_CH − 3k_CSHCC_CSH (7) 
This is the unsteady mass conservation equation for CO₂, with partial derivatives of the concentration of CO₂, as well as the following mass conservation equations for first Ca(OH)₂ then CSH: 
∂C_CH/∂t = −k_CHCC_CH (8) 
∂C_CSH/∂t = −k_CSHCC_CSH (9) 

In equations (7)-(9) C, C_CH, and C_CSH are the molar concentrations of CO₂, Ca(OH)₂, and CSH respectively, and k_CH and k_CSH are the rate constants for Ca(OH)₂ and CSH. 

For both methods, these derivations lead to a set of equations that could be used to iterate through a matrix of initial concentrations of CO₂, Ca(OH)₂, and CSH and give their final concentrations at each position after a time determined by the timestep and number of iterations. The forwards finite differencing method is easier to derive and implement, but is limited in that if the timestep is too large, it becomes unsteady and is unusable. The backwards method has no such limitation, but requires more complex operations such as matrix inversions. 

These initial models were then improved with the addition of equations that calculate new diffusion coefficients at each position depending on the initial porosity as well as reaction progression. There were two methods of the backwards finite differencing model that include this: one with a central difference calculation and the other with an average of the diffusion coefficients on each side of each position.

Comparison of Complex Models with Data
Many comparisons were done between the complex models and the simpler models with experimental data from recent literature. All of these comparisons took place in MATLAB, and they can be seen below in the Figures section. The data was then plotted with the estimate(s) from the models.

Results

In Figure 1, we combined numerous sets of data and models to show that, in certain cases, the simple estimates can prove to be decent estimates. All of the experimental data and models here are from high strength concrete stored indoors. The light blue line is the estimated depth from (5) with a carbonation rate coefficient of k = 1.5 mm/yr^0.5, and the dark blue line is the estimated depth from (6). We can clearly see that the data is close to the estimates from the two models in this case.

In comparison to Figure 1, Figure 2 demonstrates how these simple equations do not always work. All of the features are for similar strength classes, but they are not from controlled environments as in the first figure. These estimates and data are all from outside locations and coefficients. As the conditions were different for each, one can see that the slopes vary greatly, and the data does not fit close to them. The figure has estimates from (5) with a carbonation rate coefficient of k = 10.8 mm/yr^0.5 as the yellow line and estimates from (6) as the dark blue line. The estimated value from Xi et al. [3]. is uses their k value for outside high strength mortar. Additionally, the grey circles are from Papadakis et al. [6] and the orange squares are from Nagataki et al. [8].

Figure 3 is a comparison of the two different modeling methods. The red line demonstrates the moving front of the reaction as predicted by (5) and the blue line is the estimate from the backwards differencing model with the more accurate transition from fully reacted to not yet reacted. The initial conditions for both are the same. The figure specifically focuses on the concentration of Ca(OH)₂, one of the main reactants in the reactions. One can see that the method provided by (5) models as an instantaneous change, whereas the more accurate model has a more gradual curve, which is more correct.

This next figure is the first comparison between different backwards finite differencing methods with diffusion coefficient calculations. The red is the central difference method and the blue is the averaging method. The central difference method estimates the first and second derivative at a certain point with the values before and after this point, and then uses them to estimate the diffusion coefficient at that point. The method of averaging is much simpler, and finds the average between each point and the one two ahead of it, and uses that value for the point in between. Additionally, the points are data from Chang and Chen [4]. The Y-axis is the concentration of Ca(OH)₂, one of the main reactants involved in the process. All of these last 4 graphs are just showing the differences in the accuracies of the two methods for certain cases, but note that both methods are very close to reality.

Here is a very similar figure to Figure 4, but instead of taking place after 8 weeks, the data and estimates are after 16 weeks. Note how the data is much closer to the average estimate instead of the central difference estimate, opposite to Figure 4. Once again, the red is the central differencing estimate and the blue is the estimate from the averaging method of calculating diffusion coefficients, with the points being experimental data from Chang and Chen [4].

Similar to Figures 5 and 6, Figure 7 compares the central difference and averaging method for diffusion coefficients, with data from Chang and Chen [4]. However, the Y-axis is the concentration of CaCO₃, the main product of the carbonation process. We have the red line as the estimate from central difference method, the blue line as the estimate from the average method, and the points are the experimental data.

As mentioned above, this figure is similar to Figure 6, but is after 16 weeks, not 8. The lines and points are from the same sources as above. Note again the change in which method fits the data best. Once again, the data is from Chang and Chen [4].

Discussion/Conclusions

The simple models, based on (5) and (6), are adequate at predicting the carbonation depth in controlled environments where the assumptions that those equations make are true. However, they fail in uncontrolled situations where the environment is not kept the same. The more advanced models we develop solve many of these shortcomings by giving accurate estimates for a wider range of environments. Moving forward, the models themselves can be improved to contain more initial conditions to include more of the calculations involved in finding total carbon sequestered.

Acknowledgements

I would first like to thank my mentor, Professor Melany Hunt, for the instrumental insight and assistance that I was provided. I would also like to thank graduate student Ricardo Hernandez for his assistance with the development of the code, specifically the functions to calculate the porosity and diffusion coefficients. Additionally, I am extremely grateful for Dr. and Mrs. Harris for providing funds so that I was able to partake in this research. Lastly, I would like to thank the SFP department for the opportunity to engage in this research.

References

[1] K. L. Scrivener, V. M. John, and E. M. Gartner, “Eco-efficient Cements: Potential Economically Viable Solutions for a Low-CO2 Cement-based Materials Industry,” Cement and Concrete Research, vol. 114, pp. 2– 26, 2018.

[2] V. G. Papadakis, C. G. Vayenas, and M. N. Fardis, “A Reaction Engineering Approach to the Problem of Concrete Carbonation,” AIChE Journal, vol. 35, no. 10, pp. 1639–1650, 1989. 

[3] F. Xi et al., “Substantial Global Carbon Uptake by Cement Carbonation,” Nature Geoscience, vol. 9, no. 12, pp. 880–883, 2016. 

[4] C.-F. Chang and J.-W. Chen, “The Experimental Investigation of Concrete Carbonation Depth,” Cement and Concrete Research, vol. 36, no. 9, pp. 1760–1767, 2006. 

[5] P. Friedlingstein et al., “Global Carbon Budget 2020,” Earth System Science Data, vol. 12, no. 4, pp. 3269–3340, 2020. 

[6] V. G. Papadakis, C. G. Vayenas, and M. N. Fardis, “Fundamental Modeling and Experimental Investigation of Concrete Carbonation,” ACI Materials Journal, vol. 88, no. 4, 1991. 

[7] K. Tuutti, “Corrosion of Steel in Concrete,” Diss. – Stockholm – Tekniska hogskolan, 1982. 

[8] S. Nagataki, M. A. Mansur, and H. Ohga, “Carbonation of Mortar in Relation to Ferrocement Construction ,” ACI materials journal, vol. 85, no. 1, pp. 17–25, 1988.

Convention says that both sides of the brain should be symmetrical. However, contradicting evidence of asymmetric structure can be found even in the common fruit fly (Drosophila melanogaster). This study investigates possible causes of this asymmetrical body (AB) and explores the mechanisms behind communication within the neural circuit. In particular, we examine the asymmetrical morphology of major AB input neurons. We also develop an image analysis system accurately calculate the change of volume in the AB. Our work aims to enhance understanding of the brain using Drosophila as a model organism for studying neuronal networks, specifically with regards to how the coordinated activity of connected neurons gives rise to brain function in an asymmetrical manner.

Author: Arsalan Hashmi
University of California, Santa Barbara
Mentors: Professor Carlos Lois, Aubrie De La Cruz, and Ting-Hao Huang
California Institute of Technology
Editor: Hannah Chen

Introduction

Previously, it was believed that both sides of the brain in Drosophila were completely identical, but recent studies have shown that one side of the brain has an asymmetrical body (AB). This AB is a neuropil — a network of dendrites, axons, and synapses connecting neuronal cell bodies — on the left side of the brain which is only a fourth the size of the corresponding right neuropil (Figure 1). We aim to investigate how different brain functions such as memory and movement may be attributed to these different structures in the left and right hemispheres.

The asymmetrical circuit in question involves two neurons, SA1 and SA2, which innervate into the right AB. Our goal is to investigate its mechanism and function from both a developmental and a behavioral basis. We first use an inward rectified potassium channel that prevents neuronal firing and then analyze the resulting changes in development. Then, I study Drosophila courtship behavior, which could be a gateway to understanding their short- and long-term memory. Previous results have shown that while the asymmetric body does not play a role in short-term memory, it does in long-term memory, which we intend to test [1].

Development

To explore the morphology of the AB on a molecular level, we performed genetic crosses. In immature SA1/SA2, we introduced Kir2.1, an inward rectifying potassium channel that prevents the neurons from firing action potentials. The flies carrying Kir2.1-eGFP were collected, dissected, and immunostained. We captured images under a confocal microscope, which were then merged into a 3D stacked image (Figure 2).

Using the image processing package Fiji, I implemented an image analysis system to accurately calculate the volume and intensity of the RFP/GFP signal within the brains (Figure 3). RFP/GFP allows us to visualize neural activity and confirm that there is a change occurring. The procedure involved mapping out specific structures in the stacked confocal image and calculating the area of each slice.

Initial analysis of the results (Figure 4) showed conformational changes in the samples containing Kir2.1 in comparison to control. As shown below, the sample containing Kir2.1 has more visible branches, indicating added neurite projections. The asymmetric body also became more symmetrical. It was concluded that Kir2.1 does in fact impact the neurons in the region and has some role to play in the development of Drosophila brain. In future work, we aim to excite the neurons to test whether calcium influx is a factor in altering the SA1/SA2 morphology.

Behavior

We examined courtship rejection of males from pre-mated females to understand how their retention and long-term memory work when separated and placed back together. Our control experiment was designed around protocols used in a previous study (Figure 5) [2].

We first set up an assay that allows observation of mating patterns of wildtype Drosophila in real-time, dividing the virgin males and pre-mated females into separate vials. During the training period, individual males and females were placed into the same wells together, left for a nine-hour period, and then separated for 24 hours. We tested whether the males have memory of the courtship rejection from the training period by comparing their actions (Figure 6) with typical mating behaviors (Figure 7).

Results did indicate that females reject the courtship behavior as expected, but the sample size was too small to make any firm conclusions. Additionally, we ran into difficulty gathering the right tools to repeat the experiment. We did not have access to an aspirator for fly transfer, so we used anesthesia instead. We also had to use flies with “dirty” drivers, meaning they had expressions of non-interest neurons.

This first trial served as a valuable learning experience on how to conduct the experiment efficiently. In the future, we plan to repeat the experiment with consideration of the issues we faced. We also aim to perform the courtship behavioral assay on flies with silenced SA1/SA2 from Kir2.1 to compare to the wildtypes.

Acknowledgments

Special thanks to Sophia, Javier, the Read de Alaniz Group and the McNair staff and cohort for their support back at my home institution, UCSB. Thank you to my Caltech mentors Carlos, Aubrey, Ting-Hao and the rest of the Lois lab. Thank you to the coordinators and fellow WAVE/SURF/Amgen Fellows. Lastly, thank you to the Chen Institute for funding and providing me with a platform to research.

References

[1] Pascual, A., Huang, K.-L., Neveu, J., & Préat, T. (2004). Brain asymmetry and long-term memory. Nature, 427(6975), 605–606. https://doi.org/10.1038/427605a

[2] Koemans, T. S., Oppitz, C., Donders, R. A. T., van Bokhoven, H., Schenck, A., Keleman, K., & Kramer, J. M. (2017). Drosophila courtship conditioning as a measure of learning and memory. Journal of Visualized Experiments : JoVE, 124, 55808. https://doi.org/10.3791/55808