Abstract
A breakthrough methodology has been developed that combines continuous wavelet transform (CWT) signal processing with deep learning to predict microseismic events during hydraulic fracturing operations. This innovative approach transforms treating pressure data into normalized CWT scalograms, creating unique signatures for different fracture propagation modes. The technique was validated using data from the Marcellus Shale Energy and Environment Laboratory (MSEEL) and demonstrates exceptional accuracy in predicting three-dimensional microseismic event clouds from treating pressure alone. The deep learning model achieves prediction errors below 0.025% for horizontal coordinates and below 0.2% for vertical coordinates, while generating results in near real-time (40 seconds for a 2-hour fracturing job).
Introduction
Understanding hydraulic fracture propagation events is crucial for optimizing completion designs, estimating fracture geometry, and maximizing production. Traditional methods like the Nolte-Smith technique and moving reference point (MRP) analysis, while useful, have limitations in real-time application and require complex workflows or prior knowledge of closure pressure.
Microseismic monitoring provides valuable insights into fracture propagation by detecting small-scale earthquakes resulting from rock ruptures during high-pressure fluid injection. However, microseismic monitoring is expensive and not always available. The ability to predict microseismic events from readily available treating pressure data would provide operators with critical fracture characterization capabilities at significantly reduced cost.
This study introduces a novel signal processing approach that treats hydraulic fracturing as an input-output system, where pumping rate and proppant concentration are inputs and treating pressure is the output. By analyzing the output signal using advanced CWT techniques, the method reveals fracture propagation dynamics without requiring simplifying assumptions common in traditional approaches.
Methodology
Normalized CWT Scalogram Technique
The foundation of this methodology lies in applying CWT to treating pressure data during hydraulic fracturing. CWT provides continuous scaling and shifting of wavelet functions, offering high resolution in both frequency and time domains. This makes it particularly suitable for analyzing non-stationary signals like treating pressure during fracturing operations.
Figure 1: The CWT mechanism involves comparing a selected wavelet (Complex Morlet Wavelet) to segments of the treating pressure signal, calculating correlation coefficients at different scales and time positions.
[Figure shows the wavelet transform mechanism: original signal → wavelet comparison at multiple scales → energy scalogram generation]
The process involves three key steps:
- Signal Analysis: A Complex Morlet wavelet is systematically compared to the treating pressure signal across its entire duration
- Scale Variation: The wavelet is stretched and compressed to capture both fine (high-frequency) and coarse (low-frequency) features
- Energy Calculation: Wavelet coefficients are squared to create an energy scalogram showing dominant frequencies at each time interval
The energy scalogram undergoes normalization using minimum-maximum scaling to ensure consistency across different fracturing jobs:
(log₂ E)ₙₒᵣₘ = [(log₂ E) – (log₂ E)ₘᵢₙ] / [(log₂ E)ₘₐₓ – (log₂ E)ₘᵢₙ]
This normalization creates normalized CWT coefficients ranging from 0 to 1, enabling direct comparison between different wells and formations.
Fracture Event Classification
The normalized CWT scalogram reveals distinct signatures for different fracture propagation modes:
- Limited Width Propagation: Normalized coefficients 0.0-0.25 (blue regions)
- Length Growth: Normalized coefficients 0.4-0.6 (green regions)
- Height Growth/High Leak-off: Normalized coefficients 0.8-0.95 (red regions)
Figure 2: Example normalized CWT scalogram showing color-coded fracture propagation events over time. The scalogram acts as a “mathematical microscope” revealing subtle changes in treating pressure that correspond to different fracture behaviors.
[Figure shows heat map with time on x-axis, scale on y-axis, and color intensity representing normalized CWT coefficients, with distinct regions corresponding to different fracture events]
Deep Learning Architecture
The deep learning model employs a compact neural network designed specifically for regression tasks with tabular data. The architecture includes:
- Input Layer: 256 features from normalized CWT scalogram
- Hidden Layers: Sequential layers with batch normalization and ReLU activation
- Linear(256→200) → ReLU → Linear(200→100) → ReLU
- Output Layer: Linear(100→3) for X, Y, Z coordinates of microseismic events
Figure 3: Deep learning framework architecture showing the flow from CWT scalogram input through hidden layers to microseismic event coordinate prediction.
[Figure shows neural network diagram with input layer (256 nodes), hidden layers, and output layer (3 nodes for X,Y,Z coordinates)]
The model was trained on 256,555 data points from 48 hydraulic fracture stages in the Marcellus Shale, with a 70/30 train/test split. Stratified K-fold cross-validation with 10 folds ensured robust performance evaluation across different microseismic event distributions.
Results and Validation
Three-Case Validation Approach
The technique underwent rigorous validation through three distinct cases:
Case 1 – Simulated Fracture: Using a commercial planar 3D simulator with realistic geomechanical properties, the normalized CWT scalogram successfully identified three distinct fracture propagation stages, matching simulated fracture geometry evolution.
Case 2 – Field Data Comparison: Real field data from Well X1 was analyzed and compared with the previously validated MRP technique. The normalized CWT scalogram showed excellent correlation with MRP results while offering significantly faster analysis.
Case 3 – Microseismic Validation: Using MSEEL project data from wells MIP-3H and MIP-5H, the technique was directly calibrated against recorded microseismic events, providing physical validation of fracture event detection.
Performance Results
Figure 4: Comparison of predicted vs. actual microseismic events for Stage 20 in well MIP-5H, showing excellent agreement in both spatial distribution and temporal evolution.
[Figure shows 3D plot comparing actual microseismic events (blue dots) with predicted events (red dots), demonstrating spatial accuracy and event cloud similarity]
The deep learning model achieved remarkable accuracy:
- X and Y directions: <0.025% error for training data, <0.02% for testing data
- Z direction: <0.2% error for training data, with slightly higher but acceptable errors in testing
- Processing Speed: 20 seconds to generate CWT scalogram, 40 seconds to predict microseismic cloud
- Real-time Capability: Updates every 40 seconds during operations
Field Applications
The methodology was successfully applied to multiple fracture stages in the Marcellus Shale, demonstrating consistent performance across varying geological conditions. The technique accurately predicted:
- Fracture propagation direction and extent
- Temporal evolution of microseismic events
- Three-dimensional fracture geometry development
- Identification of height growth vs. length propagation periods
Industry Impact and Applications
Operational Benefits
This technology offers significant advantages for hydraulic fracturing operations:
Real-Time Monitoring: Unlike traditional microseismic monitoring, which requires expensive equipment and post-processing, this method provides near-instantaneous feedback using only treating pressure data.
Cost Reduction: Eliminates the need for dedicated microseismic monitoring equipment while providing similar insights into fracture propagation.
Enhanced Decision Making: Operators can make real-time adjustments to pumping parameters based on predicted fracture development, optimizing treatment effectiveness.
Completion Optimization: Understanding fracture propagation patterns enables better stage spacing, cluster positioning, and completion design.
Comparison with Existing Methods
Figure 5: Performance comparison between normalized CWT scalogram, MRP technique, and Nolte-Smith method showing superior event detection capability and simplified workflow.
[Figure shows side-by-side comparison of three methods analyzing the same treating pressure data, highlighting the clarity and detail provided by the CWT approach]
The normalized CWT technique offers several advantages over traditional methods:
- No closure pressure required (unlike Nolte-Smith)
- Simplified workflow (compared to MRP complexity)
- Real-time capability with continuous updates
- Higher resolution fracture event detection
- Quantitative predictions rather than qualitative interpretations
Technical Innovations
Signal Processing Advances
The application of CWT to fracture analysis represents a significant advancement in signal processing for petroleum engineering. The technique:
- Captures both transient and continuous fracture events
- Provides time-frequency resolution impossible with traditional Fourier methods
- Adapts to non-stationary nature of fracture propagation signals
- Maintains mathematical rigor while offering practical applicability
Machine Learning Integration
The deep learning component successfully bridges the gap between signal processing and physical phenomena:
- Transforms mathematical features into geological interpretations
- Learns complex relationships between pressure signatures and fracture geometry
- Generalizes across different formations and completion designs
- Provides probabilistic predictions with quantified uncertainty
Implementation Guidelines
Data Requirements
For successful implementation, operators need:
- High-resolution treating pressure data (1 Hz sampling recommended)
- Pumping rate and proppant concentration records
- Initial formation characterization for model calibration
- Historical microseismic data for initial training (if available)
Workflow Integration
Step 1: Real-Time Processing
- Continuous CWT analysis of treating pressure
- Automated scalogram generation and normalization
- Integration with existing data acquisition systems
Step 2: Event Prediction
- Deep learning model inference on CWT features
- Microseismic event coordinate prediction
- Fracture geometry estimation and visualization
Step 3: Decision Support
- Comparison with planned fracture geometry
- Identification of unexpected propagation behavior
- Recommendations for pumping parameter adjustments
Quality Control and Validation
Regular calibration against available microseismic data, pressure matching with fracture simulation models, and continuous refinement based on production outcomes ensure maintained accuracy and reliability.
Future Developments
Technology Extensions
Research is ongoing to expand the methodology’s applicability:
- Multi-Formation Training: Developing models for different geological settings
- Enhanced Physics Integration: Incorporating geomechanical constraints
- Uncertainty Quantification: Probabilistic predictions with confidence intervals
- Multi-Well Analysis: Simultaneous monitoring of pad drilling operations
Digital Integration
Future developments will focus on:
- Integration with digital oilfield platforms
- Automated workflow optimization
- Machine learning model continuous improvement
- Integration with production forecasting models
Economic Impact
Cost-Benefit Analysis
The economic advantages of this technology are substantial:
- Microseismic Monitoring Savings: $100,000-500,000 per pad
- Completion Optimization: 10-20% improvement in production through better fracture design
- Reduced Treatment Failures: Early detection of screen-outs and poor propagation
- Operational Efficiency: Real-time adjustments reduce non-productive time
Conclusions
This research introduces a paradigm shift in hydraulic fracture monitoring and analysis. The combination of advanced signal processing through normalized CWT scalograms with deep learning creates a powerful tool for real-time fracture characterization. Key achievements include:
- Breakthrough Accuracy: Prediction errors below 0.025% for horizontal microseismic event locations demonstrate unprecedented precision in fracture event prediction from treating pressure alone.
- Real-Time Capability: Processing times of 40 seconds for complete fracture analysis enable true real-time monitoring and decision-making during hydraulic fracturing operations.
- Comprehensive Validation: Three-case validation approach using simulated, field, and microseismic data provides robust confirmation of technique effectiveness across different scenarios.
- Operational Simplicity: The method requires only treating pressure data, eliminating the complexity and cost of traditional microseismic monitoring while providing similar insights.
- Universal Applicability: Successful application across multiple formations and fracture stages demonstrates the technique’s broad applicability in the petroleum industry.
The technology represents a significant advancement in completion engineering, offering operators the ability to optimize hydraulic fracturing operations in real-time while reducing costs and improving outcomes. As the industry continues to focus on efficiency and cost reduction, this methodology provides a practical solution that bridges advanced signal processing with operational decision-making.
The successful validation using publicly available MSEEL data demonstrates the technique’s reliability and sets the stage for widespread industry adoption. Future developments will focus on extending the methodology to additional formations and integrating with broader digital oilfield initiatives.
This work exemplifies how advanced mathematical techniques can be successfully applied to solve practical petroleum engineering challenges, providing both academic rigor and immediate industrial value.
You can read the details in the following paper Advanced Deep Learning for microseismic events prediction for hydraulic fracture treatment via Continuous Wavelet Transform
At ATCE 2025, we will evaluate the scalability of the model by validating its results against Rate Transient Analysis (RTA) for a Marcellus Shale well located 3–5 km away from the training dataset. (SPE-228059).