About the Author

Robot Swarm Detects Carbon Dioxide to Forecast Volcanic Eruptions

By Jillian Kunze

Volcanoes are major emitters of carbon dioxide, an odorless and invisible gas that contributes to climate change. Volcanic eruptions are preceded by changes in the ratio of carbon dioxide to sulfur dioxide, so monitoring this quantity is important for providing early eruption warning and evacuating threatened areas in a timely manner. However, the current methods for surveying volcanic carbon dioxide are both dangerous and fragmentary. 

Figure 1. The building and testing process for one of the unpiloted air vehicles in the Volcano Co-robots with Adaptive Natural algorithms swarm.
During a minisymposium presentation at the 2021 SIAM Conference on Applications of Dynamical Systems, which is taking place virtually this week, G. Matthew Fricke of the University of New Mexico described a new approach for monitoring volcanic gasses: The Volcano Co-robots with Adaptive Natural algorithms (VolCAN) swarm. “Currently, the state of the art is for geologists to clamber on foot over very hazardous conditions carrying sensors,” Fricke said. But a much more preferable method is for unpiloted air vehicles, or UAVs, to sample the carbon dioxide levels at volcanoes instead.

In 2019, over a dozen teams gathered in Papua New Guinea to test pilot 20 different UAVs at the Tavurvur and Manam volcanoes. All but one of these drones were destroyed by volcanic activity, illustrating that any volcanic robot swarm must anticipate the loss of UAVs. The teams also found it to be very challenging to pilot a UAV around a volcano since it was difficult to maintain a line of sight, leading them to believe it would be preferable to use UAVs that could automatically pilot themselves. Based on the known limitations from these tests, Fricke and his collaborators aimed to develop UAVs that autonomously measure carbon dioxide emission by first finding the source of the gas, then measuring the extent of the gas plume being emitted from the ground. 

There were a number of additional constraints on the development of these UAVs. The system needed to be portable by backpack over rough terrain and must carry carbon dioxide sensors. They also needed to survey a large area—up to several square kilometers—but could not emit carbon dioxide themselves, as that would interfere with the measurements. Fricke and his collaborators developed the VolCAN autonomous UAV swarm to contend with all these constraints. 

Fricke worked with a multidisciplinary team to design, build, and test the UAVs in the VolCAN swarm, including faculty in computer science (Melanie Moses and Jared Saia), electrical engineering (Rafael Fierro), and the earth and planetary sciences (Tobias Fischer and Scott Nowicki), as well as students (John Ericksen, Samantha Wolf, Abir Islam, Karissa Rosenberger, Julie Hayes, and Jarret Jones). The team built four dragonfly drones, each made out of 3D printed parts and incorporating a very accurate sensor that can tolerate pressure and temperature changes (see Figure 1). “We built our own because that’s easiest and cheapest,” Fricke said. Each drone cost about $2,000 to build and is able to fly autonomously for an hour.

Figure 2. The simulated flight patterns of unpiloted air vehicles (UAVs) using the LoCUS algorithm (red line) towards a carbon dioxide gas plume. Despite losing three UAVs along the way (circled in red), the swarm was able to locate the plume’s source.
Next, they needed to design an algorithm that would direct how each drone searched for the source of carbon dioxide. A simple approach is to do a pre-planned exhaustive survey, in which the UAVs follow a set path to thoroughly search an area. The team developed two styles of this approach, which they called the Linear Programming Lawnmower and the 3D Deterministic Distributed Spiral Algorithm (DDSA). Both of these survey algorithms would be sufficient to search a relatively small area. These approaches also provide a ground truth with which the researchers could compare their more sophisticated algorithms. 

The researchers developed two more complex solutions to detect carbon dioxide, direct the swarm towards it, and estimate the area of the gas plume. “When the area is too large for an exhaustive search, these algorithms react to current carbon dioxide levels to intelligently gather data,” Fricke said. The first solution was an adaptive algorithm called Moth Ballistic Swarms (MoBS), in which each drone is independent — there is no coordination or communication between them. MoBS directs each drone to move randomly and try to locate the edge of a gas plume, then uses an algorithm inspired by moth pheromone detection to follow the plume to its source. 

Their second solution was a highly coordinated sub-swarm approach called Loss-tolerant Cohesive UAV Swarm (LoCUS), in which the swarm behaves as a single gradient measurement instrument. Drones follow the shape of an Archimedes’ spiral while searching the boundary of the gas plume, then follow the gradient of carbon dioxide measurements as calculated by the sensor readings to find the source (see Figure 2). LoCUS is also able to handle the loss of drones explicitly. When a drone is knocked out, it is replaced by a designated heir that takes over its role. 

Figure 3. A hybrid field test of three unpiloted air vehicles (UAVs) using the LoCUS algorithm with simulated sensor data of a carbon dioxide plume. The colored lines show the track of each UAV, and the topological lines show the carbon dioxide concentration in the virtual plume. This test demonstrated the ability of the drones to use their sensor readings to locate the plume’s source.
Simulations comparing MoBS and LoCUS show that LoCUS generally performs better than MoBS if there are perturbations in the gas plume, through this advantage decreases as the number of drones in the swarm gets larger. The researchers moved beyond simulations to perform hybrid testing of the LoCUS algorithm, in which they gave simulated sensor data to a real team of three dragonfly UAVs and observed how they responded in the field. The results were promising, as each UAV did indeed move towards the source of their virtual carbon dioxide readings (see Figure 3). 

In January 2021, Fricke and his team tested out their dragonfly UAVs on volcanic carbon dioxide sources in Valles Caldera National Preserve in New Mexico, attempting both of the simpler lawnmower and DDSA approaches with a single drone. “There was a really nice correlation between areas with surface indications of gas emissions and the areas of high carbon dioxide detected by the drone,” said Fricke. Later in April 2021, the team tested a multi-drone DDSA at Roosevelt Hot Springs Geothermal Area in Utah using two of their UAVs, in the first ever autonomous multi-drone survey of carbon dioxide emission. Geologists were also present at the site, but the drones performed the surveys faster — a good validation for the technology. 

The research team plans to deploy their MoBS and LoCUS algorithms in the field for the first time at the Salton Sea in California this upcoming September. Though much work remains to be done, Fricke described being encouraged by their progress in software and simulations that they will be able to create a robust method for volcanic carbon dioxide measurement through UAVs. 

  Jillian Kunze is the associate editor of SIAM News