Physics-Guided Diffusion Models for Production Forecasting with Limited Well Data

Temesgen Abraha, Yves Lucet
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:332-341, 2026.

Abstract

Accurate forecasting of oil and gas well production from limited early-time data is critical for reservoir management and timely land cleanup. We present Physics-SIMS-TS, a novel framework that adapts Self-Improving Diffusion Models with Synthetic Data (SIMS) for time series forecasting and integrates physics-based decline-curve constraints. Given only the first 20% of a well’s production history, our model predicts the remaining 80%. We first adapt the SIMS methodology, originally developed for image generation, to time series by training a conditional diffusion model with negative guidance that steers generation away from synthetic artifacts. Experiments show that domain-aware data augmentation (Inverse Distance Weighting) outperforms generic generative approaches (TimeGAN, TimeVAE) by 1.7x, demonstrating that incorporating domain-specific knowledge improves forecasting performance. Building on this insight, we introduce Physics-SIMS-TS, which integrates Arps decline-curve dynamics through gradient guidance during sampling and monotonicity projection via isotonic regression. Experiments on 16,216 gas wells from British Columbia, Canada, spanning multiple geological formations, demonstrate that Physics-SIMS-TS achieves 1.9-6.2x lower prediction error than traditional machine learning baselines across all dataset sizes, with the largest improvements on small datasets where physics constraints most effectively regularize the learning problem.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-abraha26a, title = {Physics-Guided Diffusion Models for Production Forecasting with Limited Well Data}, author = {Abraha, Temesgen and Lucet, Yves}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {332--341}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/abraha26a/abraha26a.pdf}, url = {https://proceedings.mlr.press/v318/abraha26a.html}, abstract = {Accurate forecasting of oil and gas well production from limited early-time data is critical for reservoir management and timely land cleanup. We present Physics-SIMS-TS, a novel framework that adapts Self-Improving Diffusion Models with Synthetic Data (SIMS) for time series forecasting and integrates physics-based decline-curve constraints. Given only the first 20% of a well’s production history, our model predicts the remaining 80%. We first adapt the SIMS methodology, originally developed for image generation, to time series by training a conditional diffusion model with negative guidance that steers generation away from synthetic artifacts. Experiments show that domain-aware data augmentation (Inverse Distance Weighting) outperforms generic generative approaches (TimeGAN, TimeVAE) by 1.7x, demonstrating that incorporating domain-specific knowledge improves forecasting performance. Building on this insight, we introduce Physics-SIMS-TS, which integrates Arps decline-curve dynamics through gradient guidance during sampling and monotonicity projection via isotonic regression. Experiments on 16,216 gas wells from British Columbia, Canada, spanning multiple geological formations, demonstrate that Physics-SIMS-TS achieves 1.9-6.2x lower prediction error than traditional machine learning baselines across all dataset sizes, with the largest improvements on small datasets where physics constraints most effectively regularize the learning problem.} }
Endnote
%0 Conference Paper %T Physics-Guided Diffusion Models for Production Forecasting with Limited Well Data %A Temesgen Abraha %A Yves Lucet %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-abraha26a %I PMLR %P 332--341 %U https://proceedings.mlr.press/v318/abraha26a.html %V 318 %X Accurate forecasting of oil and gas well production from limited early-time data is critical for reservoir management and timely land cleanup. We present Physics-SIMS-TS, a novel framework that adapts Self-Improving Diffusion Models with Synthetic Data (SIMS) for time series forecasting and integrates physics-based decline-curve constraints. Given only the first 20% of a well’s production history, our model predicts the remaining 80%. We first adapt the SIMS methodology, originally developed for image generation, to time series by training a conditional diffusion model with negative guidance that steers generation away from synthetic artifacts. Experiments show that domain-aware data augmentation (Inverse Distance Weighting) outperforms generic generative approaches (TimeGAN, TimeVAE) by 1.7x, demonstrating that incorporating domain-specific knowledge improves forecasting performance. Building on this insight, we introduce Physics-SIMS-TS, which integrates Arps decline-curve dynamics through gradient guidance during sampling and monotonicity projection via isotonic regression. Experiments on 16,216 gas wells from British Columbia, Canada, spanning multiple geological formations, demonstrate that Physics-SIMS-TS achieves 1.9-6.2x lower prediction error than traditional machine learning baselines across all dataset sizes, with the largest improvements on small datasets where physics constraints most effectively regularize the learning problem.
APA
Abraha, T. & Lucet, Y.. (2026). Physics-Guided Diffusion Models for Production Forecasting with Limited Well Data. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:332-341 Available from https://proceedings.mlr.press/v318/abraha26a.html.

Related Material