The motivation of this study was to increase the scientific return on planetary exploration missions through automating the localization of areas of high scientific interest through mimicking the search strategies of insects and animals used when following odour plumes.

Turbulent odour plumes are representative of the environments which methane would likely be found on Mars. Turbulent plumes can be defined as odor pulses separated by gaps of lower concentration or clean air, where the odor gradient does not necessarily point towards to the source of the odor. Creating a representative model of a turbulent environment was the first step taken towards achieving the goals of this study. To test the generality of future algorithms, a 3D model of the laminar plume caused by Enceladus’ icy jets was also created. To test the accuracy of these models, the methane model was compared with experimental measurements of a turbulent odor plume found in the literature. The icy jet model was compared to previous models which were completed at larger scales than the model created for this project. After completion of the models, three control schemes were analysed for conceptual design: neural control, Kalman filtering, and source likelihood mapping.

Results:

The power of neural control is in the training process. Whereas genetic algorithms seemed to get some good results, direct training the neural network based on training sets did not result in a network that could successfully navigate turbulent odour plumes. In addition, combining the desire to search for odour sources and other tasks such as obstacle avoidance would result in needing to begin the training process from the beginning. More tradition methods like creating source likelihood maps involve more computation and are limited by the search area in the map, but would be able to interface with other system algorithms much easier. Therefore, continual work will be put into creating more effective and less computationally heavy source likelihood mapping algorithms.