Monotone-Constrained Diffusion Models for Long-Horizon Production Forecasting

Summary

The paper introduces Physics-SIMS-TS, a conditional diffusion forecaster that observes roughly the first 20% of an oil and gas well's production history—interpolated onto a 100-point lifetime-fraction grid and min-max normalized per curve—and predicts the remaining 80%. The method combines SIMS-style negative guidance from an auxiliary model trained on synthetic samples, Arps decline-curve penalties applied during DDIM sampling, and a terminal isotonic (PAV) projection guaranteeing monotone decline, supported by IDW spatial augmentation, EMA weights with min-SNR loss, and a 5-model x 20-sample ensemble. Across more than 35,000 wells in British Columbia, Alberta, and Pennsylvania, under a shared evaluation space and matched training budgets, it beats CSDI, Diffusion-TS, TSDiff, and TimeGrad in all twelve baseline-jurisdiction comparisons (win rates 60-82%, p<0.001) but trails ensembled PatchTST by 5.5% and 5.4% MSE on BC and Alberta, with Pennsylvania unresolved. Monotonicity costs at most 0.5% MSE, and applying the same projection to baselines costs about 0.1%. One dispersion factor per jurisdiction restores near-nominal interval coverage. A RevIN variant is the leading diffusion baseline on six standard benchmarks (0.471 vs CSDI's 0.519), behind transformers. The authors themselves show the guidance components are accuracy-neutral, attribute accuracy to pipeline choices, and quantify four protocol decisions that flip rankings. Main limitations: it is the most expensive method (12-21x transformers), monotonicity contradicts actual post-forecast behavior on essentially all wells, and the calibration factor is fitted on half the test set.