The main objectives of the machine learning work package are:

• To develop machine learning (ML) tools, designed for and tested on planetary science cases submitted by the community, and to provide sustainable, open access to the resulting products, together with support documentation
• To foster wider use of ML technologies in data driven space research, demonstrate ML capabilities and generate a wider discussion on further possible applications of ML

The goal is to build a multipurpose toolset for ML-based data analysis that will be applicable to a range of scientific research questions in planetary science with minor or easily-achievable customization efforts. We provide both, trained models as well as ML code, all of which are free to use and to be adapted by the user. The term “ML tool” refers in our case to ML codes, notebooks, or trained models, with which data can be analyzed. Applying our tools to data will result in (high-level and derived) data products, which will be made available via VESPA, together with the original data.

The tools are shared and made accessible to the wider planetary science community through this webpage, the ML Portal, and our GitHub organization. The ML work package also provides documentation, tutorials, and workhops, to support the beneficiaries and the community users of the ML tools.

In order to reach our goal of building versatile ML based tools for data analysis in planetary science, we pursue a bottom up approach, i.e., building tools that solve a diverse set of specific problems in planetary science with the aim of generalizing the solutions as much as possible. To achieve this, the community was asked to provide scientific use cases during the proposal phase of the project. Out of these proposed cases, a representative set of science cases was selected and the respective institutes were included as beneficiaries.

The idea is that the technical partners (KNOW, Uni-Passau and ACRI-ST) will closely collaborate with the scientific partners of our WP to build tools that solve “real life” scientific problems and at the same time provide the flexibility necessary to adapt the resulting toolset to other related problems in planetary science with passable effort.

A careful assessment of these science cases provided a solid basis to determine the technologies that would be best fitted to build the tools. It turned out that ML tools to classify, filter and search time series data as well as Deep Learning (DL) based classifiers of 2D (e.g., surface images or spectra) and video data, such as Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNNs) in general as well as other techniques (as e.g., Self Organizing Maps) will be applied. Supervised as well as unsupervised approaches, with the latter also being able to leverage unlabelled data, are the technologies expected to benefit the community the most.

Furthermore, we investigate automated pre-processing methods based on automated machine learning in order to optimize the pre-processing step per science case. The respective tools are built in an iterative process, continuously assessed and validated by the scientific partners as well as the planetary science community in general. Hands on sessions at conferences and workshops support the process of gathering feedback to validate and further optimize the developed solutions.

We plan to have applicable and working ML tools for all of our science cases at the end of the workpackage, which will be in July 2023. Of course, as soon as a working prototype is available, we will put the corresponding codes on our GitHub account.

The integration of the data sets and the data products of the first science cases (the IWF science case, the LMSU science case, as well as the GMAP science cases) into VESPA will start in winter 2021. We will then continuously include the data sets from the other science cases.