Lithology Classification Based on Well Log Data - A Benchmark for Machine Learning Models

The paper proposes a benchmark for lithology classification using well log data, mainly driven by the lack of a consistent experimental protocol to compare different models in the field, since works often use their own datasets, data splitting strategies, and metrics. For that, the authors define a standard protocol for data preparation and model evaluation, using the FORCE and Geolink datasets as experimental sets. In total, 280 experiments were executed using ten methods from the literature, including traditional machine learning models and deep learning architectures. The results indicate that the task is far from solved and reinforce the importance of a public, reproducible, and impartial benchmark to guide future work.