Data-local federated learning for fixed-step depth-ahead prediction of LWD gamma-ray logs in heterogeneous oilfields

Accurate real-time interpretation of logging-while-drilling (LWD) gamma-ray (GR) logs is important for formation evaluation and drilling decisions, but behind-bit sensing delays formation information. This study evaluates a data-local federated learning framework for fixed-step depth-ahead GR prediction across three oilfields with different feature sets, data volumes, and native measured-depth sampling. The main benchmark uses 300 historical samples to predict the next 50 GR samples with preprocessing fitted only on the corresponding training well. Under uniform client weighting, FedAvg reduced mean MAE from 7.13 to 4.44 gAPI in Volve, from 2.23 to 1.49 gAPI in BH, and from 5.76 to 2.79 gAPI in XJ relative to local training, while substantially reducing across-seed dispersion; the paired mean differences did not reach p < 0.05. Four strictly causal baselines were weaker than FedAvg on the evaluated holdouts. Additional diagnostics show field-dependent effects of sample-size weighting and gap-aware windowing, while a matched-volume source-composition test does not support attributing the observed gains to cross-field source diversity itself. The XJ 300/50 case contains only five highly overlapping test windows covering 54 unique target GR points. Local inference latency is millisecond-scale, whereas federated model updating is a separate minutes-scale process whose communication tolerance depends on bandwidth and model-refresh interval. Overall, the evidence supports data-local collaborative training for fixed-step depth-ahead GR prediction, with important limitations from irregular sampling, severe client imbalance, domain heterogeneity, and small-sample evaluation.