Figure 1: Geology with parallel horizontal layers
Figure 2: Actual 1D geology with faults and dips
This project addresses the challenge of predicting geological layer-depth sequences for horizontal well drilling operations. Subsurface geology exhibits inherent uncertainty due to complex deposition patterns and tectonic deformations. Our data-driven model generates 10 geological realizations (300 positions each) to account for prediction uncertainty, supporting geosteering decisions during drilling operations.
Predict 1D geological layer sequences (delta_z values) along lateral wells using:
- Known geological data from columns -299 to 0 (current drilling position)
- Generate 10 equiprobable realizations for positions 1-300
- Handling partial observations with guaranteed data only from -49 to 0
- Modeling geological deformations (faults and dips)
- Generating diverse predictions while maintaining accuracy
| Column Range | Description | Normalization |
|---|---|---|
| geology_id | Unique hash-based identifier | - |
| -299..0 | Observed Z-coordinates (partial data) | Z(0) = 0 |
| 1..300 | True geology (realization 0) | Competition target |
| r_{1-9}pos{1-300} | Additional realizations (1-9) | Equiprobable |
- Training data contains complete true geology for all realizations
- Test data contains only partial observations (-299 to 0)
- Position values represent feet relative to drill bit
The evaluation uses a weighted NLL loss accounting for position-dependent errors:
Where:
e_i(x) = b_T(x) - b_i(x)(prediction error)D_T^{-1}(x)= Position-dependent covariance matrix:
log_slopes = [1.0406, 0.0, 7.8353]
log_offsets = [-6.4307, -2.1617, -45.2488][Coming soon]
[Coming soon]
[Coming soon]
# Set random seeds
SEED = 42
torch.manual_seed(SEED)
np.random.seed(SEED)
random.seed(SEED)- Alyaev & Elsheikh (2022) - Geological boundary modeling
- Muhammad et al. (2024) - Particle filter applications in geosteering
- ROGII Inc. - Competition sponsor and data provider