Towards Subsurface ML Metrics

Use of metrics are key in application of machine learning in any domain. Good metrics allow us to assess performance of algorithms, gain insight into the behaviour of models and understand the impact of model and parameter choices as well as data and feature selections. Shared metrics allow research and engineering communities share knowledge and communicate effectively at a high level, helping progress and reproducibility.

In applying ML in the subsurface, the first port of call is to use standard ML performance metrics such as accuracy, f1_score and r2 score. These metrics are well know but generic. In some cases they provide effective performance indicators, more so in classification tasks. However they generally don't provide much insight into why model is achieving a particular level of performance, or measure performance in terms of expected or acceptable subsurface behaviour.

In this workshop session, we aim to further the discussion on why development of a common set of meaningful subsurface metrics is important for the our community. We highlight some of the gotchas and shortcomings with typical metrics used in machine learning classification and regression tasks and we propose some potentially routes forward.