Through this PhD, I discovered that the important thing is not that I try to learn everything there is to know all the scientific disciplines I work with, but that I learn enough about each of them to understand how they can fit together.
To this end, I was lucky to get to closely collaborate with experts from the Institute of Environmental Sciences (CML) at Leiden University and the European Space Agency (ESA), as well as with researchers from my own field (AI). Once I embraced that my job is making connections between concepts from these different fields, rather than trying to be an expert myself in everything at once, this topic became incredibly enjoyable.
See also the project description at AutoAI4EO.
The following publications were part of this PhD project:
- A super secret project that is still under review 😉
- Ill-posedness of physical model inversion
- Training-free cloud removal for satellite data
- Spatial interpolation (VPint)
Physics-aware machine learning (PA-ML) is a broad category of methods where machine learning is combined with the rich domain knowledge on physical systems that is already available to us.
When working with Earth observation (EO) data, we are often interested in variables that represent key parameters in complex physical systems, such as climate, ocean currents and ecosystems, while we do not necessarily have access to reliable ground truth data. This is a strong motivation to incorporate physical domain knowledge into machine learning systems.
Ironically, while this was my PhD topic, I do not have a strong physics background myself, having opted instead for subjects such as history, philosophy and languages back in secondary school. However, though the learning curve was steep, and I would consider myself far from an expert in physics, I found the process of learning to unify the fields extremely rewarding. While I may not be a physicist, it is very satisfying to unify the highly valuable work from those who are experts in it with the AI methods that I am more familiar with (mainly machine learning and optimisation).
PA-ML in a nutshell
One of the most popular examples of PA-ML is the use of physics-informed neural networks (PINNs). In these networks, existing physical models are usually parameterised using a neural network. It is also possible to incorporate physical constraints (such as the conservation of mass constraint) into the training procedure of a neural network, to encourage physically consistent outputs. Although I remain interested in this type of approach (hopefully I can get some students interested in it as well!), I did not end up using them much for my own research projects.
The type of PA-ML I chiefly focused on is the combination of physical simulation models with machine learning techniques. There are various ways in which simulation models can be incorporated into machine learning pipelines. For example, we could use the simulation model to generate data. We could then train a machine learning model to “emulate” this simulator, at a fraction of the computational cost. We could also create “hybrid models” to improve extrapolation performance (a key weakness of deep learning models) by adding additional loss terms for non-observed data points put through the simulator, thereby preventing a model from fitting the observed data well only to completely break down outside of the training distribution. Finally, because physical models are generally founded on core principles of cause and effect, we can use machine learning to learn a mapping function from observed effects to inferred caused (“model inversion”).