The biggest challenge of aviation today is to drastically abate its growing environmental impact. Aviation is the second biggest source of transport emissions after road transport and at the same time is the fastest-growing source of greenhouse gas emissions. Specifically, if global aviation were a country, it would rank in the top 10 emitters. As a result, sustainability improvements in aviation, are essential for tackling climate change and the resulting ecological crisis.
Artificial Intelligence (AI) is proving to be a real game changer since AI models have achieved promising results toward clean aviation. However, these achievements come at a cost: they are not yet sustainable. In 2019, researchers at the University of Massachusetts Amherst analyzed various AI natural language processing training models in order to estimate the energy cost in kilowatts required to train them. Converting this energy consumption into approximate carbon emissions and electricity costs, the authors estimated that the carbon footprint of training a single big language model is equal to around 300,000 kg of carbon dioxide emissions. This is of the order of 125 round-trip flights between New York and Beijing for one person. Therefore, moving towards clean aviation, sustainable AI models must be developed first.
To that end, the research in this Ph.D. project will be on developing novel Bayesian adaptable physics-informed prognostic models for aeronautical structures and systems so as to transform ‘power-hungry’ models into ‘power-sustainable’ models. Towards a physics-informed model, the user will be able to engage physical meaning and replace parts of the training process that will allow better understanding and interpretation of the results, while maintaining a high level of learning and performance. Furthermore, the related CO2 emissions will be reduced since less learning data will be needed, and as a result, the computational effort of the training process will be drastically decreased. Finally, utilizing adaptation characteristics the physics-informed model will be able to learn from unexpected phenomena during the operation and adapt its estimated parameters so as to obtain more accurate and reliable prognostics, along with decreasing the needed learning data.