����������

André Antunes de Sá

André Antunes de Sá, a PhD candidate in the Ann and H.J. Smead Aerospace Engineering Sciences Department, is co-author of a new paper published in The work, released in August, is titled “Lightning Distance Estimation Using LF Lightning Radio Signals via Analytical and Machine-Learned Models.” We asked about applications of this research in the real world, how he quickly learned about machine learning for the project and what questions remain unanswered. 

Question: How would you describe and summarize the work and results of this paper?
Answer: Our main goal for this study was identifying the location of lightning events by using just a single portable instrument. The lightning flashes we see during a thunderstorm are the result of an electrostatic discharge, where a large flow of charged particles will move between two electrically charged regions in the cloud or the ground. The electric discharges in the clouds emit strong radiation across a large portion of the electromagnetic spectrum. They light up the sky with beautiful visible flashes and also transmit a signature radio pulse. We call these lightning radio pulses "sferics,” from the word "atmospherics,” and we can obtain these signature pulses using a specialized radio receiver.

The radio sferic we observe on a receiver depends on many properties of the originating lightning discharge and of the path between the lightning and the receiver, including the length of that path. 

In this paper, we investigated building a model that takes the sferic observation at the receiver and estimates the distance to the originating lightning discharge. To do that, first we looked into an analytical model that exploits a well-known relationship between the duration of a sferic and distance to the lightning discharge. A major disadvantage of this relationship is its dependence on the state of the ionosphere, a layer in the Earth's upper atmosphere that can change significantly with time and cannot be predicted accurately enough for our needs. 

We then looked into training machine-learned models for the same purpose, which ultimately did not display the same precision limitation from an unknown ionosphere. In the end, our models were found to be more practical and slightly more accurate with a 30-mile uncertainty than other single-instrument lightning distance estimation methods. We expect even more accurate estimates as an appropriately comprehensive sferic dataset is created.

Q: Beyond forecasting improvements, what are some real-world applications of this work? Why do we want to study this?
A: While there are several lightning location services that provide reliable near real-time information on lightning discharges, these services require network connectivity. That reduces their suitability for users at more remote locations faced with expensive or non-existent data channels. Commercial and general aviation, remote scientific operations, the maritime sector, and even local community event planning would benefit from a standalone device capable of detecting and locating lightning events through the models we investigated, either for lightning and severe weather avoidance, or for decision-making on mission operations involving lightning research. Additionally, given the ionospheric effects on the observed sferics, our research can be leveraged in studying the ionosphere further.

Q: Was this a research question or area you were particularly interested in before joining the project? 
A: I actually didn't know anything about lightning research and very little about machine learning when my advisor first introduced this research opportunity to me. I have always been interested in applied math, and using machine learning to study lightning sounded very exciting! In the end, this study was a great opportunity to learn about the two fields and in starting my graduate research career.

Q: That probably gave you a really unique perspective. Can you explain exactly how machine learning was used in this project and why? What advantages did it offer?
A: I would say the hardest part of this work was in understanding many concepts and implications of machine learning well enough to use it in atmospheric sciences research and to write a compelling paper on the topic. The field of machine learning is vast with a large amount of recent publications, and it was hard to find comprehensive material discussing machine learning techniques from an intermediate-level perspective. 

This difficulty was only made apparent as I delved deeper into the applications of machine learning specific to my research problem. I eventually compiled my own list of very useful references that helped me through this effort, and I am happy to share that if anyone is just starting out now.

The advantages of the machine learning approach include the ability to uncover relationships between the inputs and outputs that may not be obvious or readily accessible when relying on an analytical solution and a fast run time of the trained models given that they use simple arithmetic operations, which is important in keeping the lightning instrument low-power and low-cost.

Q: What research questions are still to be answered after this paper? Where will the work go from here?
A: One of the most important conclusions in the paper relates to the limitations of the machine learning approach, particularly in regard to data pre-processing. Future research on improving the machine-learned models should include a more comprehensive sferic dataset and a study of different sampling and data pre-processing techniques, as suggested in the paper. 

In terms of lightning research, our group deployed multiple radio instruments for studying lightning during the field campaign in Argentina. By using multiple receivers we can get a more precise location for a lightning discharge of about a few miles. Currently, I am studying the most energetic lightning present in the RELAMPAGO storms and investigating the environment in which they occur, which might help us in better forecasting severe thunderstorm weather.

was supported in part by the Integrated Remote and In Situ Sensing Program through the University of Colorado Boulder and in part by NSF under Grant AGS 1661726. The work of Andre L. Antunes de Sá was supported by the NASA Earth and Space Science Fellowship under Grant 80NSSC17K0392. Other ���������� Boulder authors include Robert Marshall.