SPE ML Challenge 2025 – Team Polska

This repository contains the code, report and the results for Team Polska in the SPE GCS ML Challenge 2025. It includes scripts for data pre-processing, model training, and inference to reproduce our solution.

Problem statement

The 2025 SPE GCS ML Challenge aimed to accurately predict the Bottom Hole Circulating Temperature (BHCT) for five wells. The problem was divided into two phases:

• Phase 1: Develop a single, generalized model to predict BHCT.
• Phase 2: Fine-tune the model for improved accuracy on new wells. Participants were restricted from using pre-trained models or external datasets, such as the Geothermal Data Repository. Model performance was evaluated using the R-squared (R²) metric, measuring how well predictions matched actual values. The challenge reflected real-world constraints, including limited training data and variations in well conditions.

Our approach

The detailed report describing out approach and finidings is available at report/SPE-GCS-ML-Challenge-2025-report.pdf.

Overview

After testing multiple architecture, we found that Gated Recurrent Unit (GRU) networks proved most effective for geothermal drilling data.

Best Model: GRU

• Accurate, efficient, adaptable
• Simplified architecture for faster training

Data Preparation

• Last 500 rows reserved for testing
• Synthetic augmentation: noise injection, random initialization

Performance

• Median output from 200+ models improves generalization
• Mean Absolute Error (MAE) for evaluation

Results

Ranked 8th place in the final standings.

Head to the results folder to see how our predictions compared to the ground truth.

---

Reproduction Steps

1. Set Up the Environment

Ensure you have Python 3.10 installed. Then, create the environment and install dependencies:

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1
pip install -r requirements.txt

Next, copy the challenge dataset into the data/raw folder.

---

2. Data Preprocessing

To format the dataset for training, run:

bash ./preprocess.sh

The processed dataset will be saved in data/formatted.

---

3. Training the Model

Configuring Training Parameters

Before training, modify the configuration file config.py. It consists of two sections:

• PARAMETERS WHICH CAN BE MODIFIED → Adjust these as needed.
• DO NOT TOUCH → Leave these unchanged.

Key settings:

• Training wells: Multiple wells from both phases can be used for training.
• Validation/Test well: Exactly one well must be selected for validation/testing.
  • Set P1_TEST_HOLE to 0, 1, or 2, or P2_TEST_HOLE to 0 or 1.
  • They cannot both be None.
  • The selected test well must also be included in training.
• Test size: The last N rows of the test well will be used for evaluation.
  • For final model submission, a small value (e.g., 16) is recommended.

Example configuration:

P1_TRAIN_HOLE: int | list[int] | None = 0  # Options: [0, 1, 2], 0, 1, 2, None
P1_TEST_HOLE: int | None = 0  # Options: 0, 1, 2, None

P2_TRAIN_HOLE: int | list[int] | None = None  # Options: [0, 1], 0, 1, None
P2_TEST_HOLE: int | None = None  # Options: 0, 1, None

TEST_SIZE: int = 16  # Number of last rows used for testing

Running Training

Once the configuration is set, train the model by running:

python train.py

Since our approach relies on an ensemble of multiple models, repeat the training process with different configurations. In our case, we trained approximately 40 models per well to improve generalization.

---

4. Inference

After training sufficient models, generate predictions and validation results for all models:

python generate_all_models_prediction_and_validation.py

Once complete, aggregate all predictions into a final output:

python combine_final_predictions.py

Final predictions will be saved in the predictions folder.