IN-CONTEXT LEARNING WITH TABULAR FOUNDATION MODELS FOR METHANE EMISSION CLASSIFICATION

Methane emissions from the natural gas production infrastructure can be roughly classified into ‘vented’, or expected during normal operation; or ‘fugitive’, which are emissions related to leaks. This classification is important because new regulations are being defined to impose fines on excessive fugitive emissions. Existing methods for this classification are manual and labor intensive Optical Gas Imaging (OGI) surveys or some form of on-site sensor measurements. The latter, while being less expensive than OGI methods, often have difficulties distinguishing between vented and fugitive emissions. Since fugitive methane emissions or ‘leaks’ need to be fixed in a time sensitive manner, automated classification using machine learning will be more efficient. In particular, we evaluate whether in-context learning (ICL) with a pretrained tabular foundation model delivers competitive results without task-specific neural parameter updates. Using Equipment, methane emission rate, event duration, and source height, we compare TabFM with Logistic Regression, Decision Tree, Random Forest, and XGBoost under a rolling 30-day history to 7-day future-test protocol. With the full 30-day history, Random Forest gives the strongest default-threshold Macro-F1 (0.921 ± 0.135), whereas TabFM provides the strongest ranking performance (AUROC 0.997 ± 0.007; AUPRC 0.996 ± 0.012). A controlled few-shot experiment further shows that with only 10 labeled examples per class, TabFM nearly matches Random Forest in Balanced Accuracy (0.840 vs. 0.841) and Macro-F1 (0.804 vs. 0.805) while achieving markedly higher AUROC (0.958 vs. 0.915). At 25 examples per class, TabFM reaches AUROC 0.989 ± 0.015 and AUPRC 0.983 ± 0.031. These findings suggest that ICL is particularly attractive when methane-specific labels are scarce and robust probabilistic discrimination is more important than optimizing a single fixed decision threshold.