Physics-guided TCN–transformer with multi-objective loss for real-time rate-of-penetration prediction in deep geothermal drilling

In deep geothermal wells, accurate real-time prediction of rate of penetration (ROP) is crucial for increasing drilling efficiency and minimizing operational costs. Existing machine learning (ML) approaches used for ROP prediction are limited by either not accounting for the basic principles of the physics of drilling or how their architectures are designed, wherein they have a very limited temporal receptive field to capture long-term dependencies of sequential downhole data. Our proposed method combines a TCN (temporal convolutional network) and Transformer (TCT-Phy) model that was trained using a physics-guided multi-objective loss to jointly optimize for empirical accuracy, consistency of MSE (mechanical specific energy), and the constraint regarding physical monotonicity. The proposed model was tested using the Utah FORGE MU-ESW1 dataset, which contains 6,748 quality-filtered records of drilling from a deep geothermal well. The average MAE (mean absolute error) of the proposed method was 4.50 ± 0.43 ft/hr, and the R² coefficient of determination was 0.917. Therefore, the proposed method outperformed a strong background gradient boosting baseline (MAE 8.48), plain TCN (MAE 5.14), and an ablated version without physics guidance (MAE 5.24). Additionally, the presence of the physics loss improved the consistency of MSE by 17.4% compared to the architecture-only ablation, while the monotonicity ratio of the outputs of the proposed model compared to the ground truth field data (0.444) indicates that the soft constraints placed on the proposed method enforce physical plausibility without artificially masking real-world formation effects. The model satisfies real-time deployment constraints on standard GPU hardware, and all code, data, and experimental protocols are fully reproducible within a single notebook environment.

Authors

Publication Details

Journal
Journal of Engineering and Applied Science
Published
2026-11-20
DOI
https://doi.org/10.1186/s44147-026-01299-7
Primary Topic
Drilling and Well Engineering
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Physics-guided TCN–transformer with multi-objective loss for real-time rate-of-penetration prediction in deep geothermal drilling

Abdurakhmanova Surayyo, O. G. Khayitov, Sezim N. Mustapayeva, Akhmedova Nargiza et al.
Journal of Engineering and Applied Science
Drilling and Well Engineering
article

Physics-guided TCN–transformer with multi-objective loss for real-time rate-of-penetration prediction in deep geothermal drilling

Abdurakhmanova Surayyo, O. G. Khayitov, Sezim N. Mustapayeva, Akhmedova Nargiza, Bekzod Abduraxmonov, Tulyаgаnоvа Nаrgizа, Jumagulov Aydos, Khalmatova Nargiza, Batirova Nargiza, Shukurullaeva Saodat
article en

Abstract

In deep geothermal wells, accurate real-time prediction of rate of penetration (ROP) is crucial for increasing drilling efficiency and minimizing operational costs. Existing machine learning (ML) approaches used for ROP prediction are limited by either not accounting for the basic principles of the physics of drilling or how their architectures are designed, wherein they have a very limited temporal receptive field to capture long-term dependencies of sequential downhole data. Our proposed method combines a TCN (temporal convolutional network) and Transformer (TCT-Phy) model that was trained using a physics-guided multi-objective loss to jointly optimize for empirical accuracy, consistency of MSE (mechanical specific energy), and the constraint regarding physical monotonicity. The proposed model was tested using the Utah FORGE MU-ESW1 dataset, which contains 6,748 quality-filtered records of drilling from a deep geothermal well. The average MAE (mean absolute error) of the proposed method was 4.50 ± 0.43 ft/hr, and the R² coefficient of determination was 0.917. Therefore, the proposed method outperformed a strong background gradient boosting baseline (MAE 8.48), plain TCN (MAE 5.14), and an ablated version without physics guidance (MAE 5.24). Additionally, the presence of the physics loss improved the consistency of MSE by 17.4% compared to the architecture-only ablation, while the monotonicity ratio of the outputs of the proposed model compared to the ground truth field data (0.444) indicates that the soft constraints placed on the proposed method enforce physical plausibility without artificially masking real-world formation effects. The model satisfies real-time deployment constraints on standard GPU hardware, and all code, data, and experimental protocols are fully reproducible within a single notebook environment.

Journal of Engineering and Applied ScienceVol. 73(1)
Drilling and Well Engineering
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon