1 / 20
πŸ› National Institute of Technology, Raipur β€” 492010, India

Intelligent Prediction of
Blast-Induced Ground Vibrations
Using Advanced Machine Learning Algorithms

Ashish Kumar Vishwakarma* Afsha Anjum Ankush Dhobi Argha Choudhary Prakash Dhekne

*Corresponding: akvishwakarma.min@nitrr.ac.in

πŸ“ Samaleswari Opencast Mine, Odisha
πŸ“Š 88 Blast Datasets
πŸ€– RF Β· GBR Β· XGBoost Β· SVR
πŸ† XGBoost RΒ² = 0.862
Mining Engineering Β· Machine Learning Β· Ground Vibration Prediction
02

Problem Statement

Blast-Induced Ground Vibration β€” A Critical Engineering Challenge
πŸ’₯

Blasting in Mining

Drilling & blasting is the most economical method for rock excavation. Only 20–30% of explosive energy contributes to effective rock breakage.

⚠️

Energy Dissipation

Remaining 70–80% dissipates as environmental nuisances: air overpressure, flyrock, and blast-induced ground vibrations (BIGV).

πŸ—οΈ

Structural Risks

BIGV causes backbreak, overbreak, wall failure, slope instability, and damage to surrounding rock mass and nearby structures.

πŸ“‹ DGMS (India) PPV Regulatory Limits

Structure Type>25 Hz8–25 Hz<8 Hz
Sensitive / Historical1052
Domestic (brick/cement)15105
Industrial Buildings252010
Industrial (limited life)502515

Values in mm/s Β· Source: DGMS, 1997

PPV
Peak Particle Velocity β€” Primary Vibration Metric
BIGV
Most Critical Environmental Concern in Mining
03

Research Gap

Limitations of Existing Empirical Models
Empirical Models
βœ“ Distance (D)
βœ“ Max. Charge (Q)
βœ— Spacing (S)
βœ— Burden (B)
βœ— Hole Depth (H)
βœ— Rock Properties
βœ— Geological Structure
RΒ² β‰ˆ 0.09 – 0.24
β†’
Research
Gap
β†’
ML Models (This Study)
βœ“ Distance (D)
βœ“ Max. Charge (Q)
βœ“ Spacing (S)
βœ“ Burden (B)
βœ“ Hole Depth (H)
~ Rock Properties
~ Geological Structure
RΒ² β‰ˆ 0.637 – 0.862
Scaled Distance Concept: Empirical models reduce PPV prediction to only 2 parameters β€” fundamentally insufficient for complex site conditions
Non-linear Interactions: Spacing, burden, and hole depth exhibit complex non-linear interactions with PPV that empirical formulas cannot capture
Site Specificity: Empirical constants (k, b) are site-specific and cannot generalize across different geological conditions
04

Research Objectives

Scope and Goals of the Study
01
🎯

Predict PPV

Develop accurate prediction models for Peak Particle Velocity using five blast design parameters: D, Q, S, B, H

02
βš–οΈ

Comparative Analysis

Compare performance of empirical equations, multivariate regression, and four ML algorithms on identical datasets

03
πŸ”

Parameter Influence

Quantify the influence of hole depth, burden, and spacing β€” parameters neglected in most prior studies

04
πŸ†

Best Model Selection

Identify the most accurate and generalizable model using RΒ², RMSE, MAE, and cross-validation metrics

INPUT PARAMETERS (5)
D β€” Distance (m) Q β€” Max. Charge/delay (kg) S β€” Spacing (m) B β€” Burden (m) H β€” Hole Depth (m)
β†’
OUTPUT (1)
PPV β€” Peak Particle Velocity (mm/s)
05

Study Area

Samaleswari Opencast Project, Jharsuguda, Odisha
Jharsuguda Odisha INDIA N ~500 km Ib Valley Coalfield
Samaleswari OCP Odisha State
Operator
Mahanadi Coalfields Limited (MCL)
Location
Jharsuguda District, Ib Valley Coalfield, Odisha
Commenced
April 28, 1993
Elevation
210–230 m AMSL Β· Depth: 100–120 m
Geology
Barakar Formation, Gondwana Supergroup β€” interbedded sandstone, shale, Lajkura & Belpahar coal seams
Overburden
Medium-to-coarse grained sandstone (high strength & elasticity)
Mining System
Highly mechanized opencast β€” thick multi-seam coal deposits (Lower Gondwana Basin)
06

Data Collection

88 Experimental Blast Datasets β€” Samaleswari OCP
88
Total Datasets
10
Experimental Blasts
5
Input Parameters
1
Output (PPV)

Statistical Summary of Input and Output Parameters

Parameter Symbol Unit Min Mean Median Max Std Dev CV (%)
Distance from blast face D m 50 146.74 149 333 70.64 48.14
Max. explosive weight/delay Q kg 15 43.35 49 80 17.03 39.28
Spacing S m 3.0 4.62 5.0 5.0 0.61 13.14
Burden B m 2.9 3.88 4.0 4.0 0.34 8.69
Hole Depth H m 2.2 4.47 4.4 7.5 1.32 29.44
Peak Particle Velocity PPV mm/s 1.32 6.63 5.22 30.7 5.25 79.18

CV = Coefficient of Variation = (StDev/Mean) Γ— 100 Β· High CV for PPV (79.18%) indicates significant variability in ground vibration response

πŸ“‘ Instrument: Digital seismograph with tri-axial geophone sensor for ground vibration monitoring
07

Methodology

Generalized Framework for PPV Prediction
πŸ“Š
Field Data Collection
88 datasets Β· 10 blasts Β· Samaleswari OCP
↓
Input Parameters (5)
D  |  Q  |  S  |  B  |  H
β†’
Output
PPV (mm/s)
↓
Dataset Split (80:20)
Training: 70 datasets (80%) Testing: 18 datasets (20%)
↓
Empirical Models
BIS 6922 Β· AAH
LAK Β· USBM
Multivariate Regression
5-parameter linear
regression model
ML Models
RF Β· GBR
XGBoost Β· SVR
↓
Model Evaluation & Validation
RΒ² RMSE MAE 5-Fold CV SHAP Analysis
↓
πŸ†
Best Model Selection & Engineering Interpretation
08

Empirical Models

Scaled Distance-Based PPV Prediction Equations
BIS 6922
Indian Standard (IS-6922, 1998)
PPV = k Β· (Q2/3 / D)1.25
RΒ² = 0.0913
Cube-root scaling; Indian regulatory standard
USBM
Duvall & Petkof (1959)
PPV = k Β· (D / Q1/2)βˆ’b
RΒ² = 0.1515
Square-root scaling; foundational BIGV model
LAK
Langefors & KihlstrΓΆm (1978)
PPV = k Β· (Q / D2/3)2b
RΒ² = 0.0073
Cube-root scaling; energy distribution emphasis
AAH
Ambraseys & Hendron (1968)
PPV = k Β· (Q1/3 / D)βˆ’b
RΒ² = 0.2431
Best among empirical; cube-root scaling
⚑ Scaled Distance Concept
SD = D / Qn   where n = Β½ (USBM) or β…“ (AAH/LAK)
Fundamental Limitation: All empirical models reduce PPV prediction to only 2 parameters (D, Q), ignoring the complex influence of spacing, burden, hole depth, and geological variability β€” leading to consistently low RΒ² values across all site conditions.
09

Multivariate Regression Model

Combined Influence of All Blast Design Parameters
Derived Regression Equation (Eq. 4)
PPV = 13.767 βˆ’ 0.0468D + 0.195Q + 15.426S βˆ’ 19.82B βˆ’ 0.707H
D = Distance (m) Q = Charge (kg) S = Spacing (m) B = Burden (m) H = Hole Depth (m)

Coefficient Interpretation

Spacing (S)
+15.43
Burden (B)
βˆ’19.82
Charge (Q)
+0.195
Hole Depth (H)
βˆ’0.707
Distance (D)
βˆ’0.047
Fig. 5: Predicted vs. Actual PPV β€” Multivariate Regression (RΒ² = 0.456)
↑ Q & S β†’ Higher PPV (direct relationship)
↓ D, B & H β†’ Lower PPV (inverse relationship)
RΒ² = 0.456 β€” Moderate improvement over empirical models
10

Machine Learning Models

Four Advanced Algorithms for PPV Prediction
RF

Random Forest

T₁
Tβ‚‚
T₃
β†’ avg β†’
PPV
  • Ensemble of n decision trees (bagging)
  • Each tree trained on random data subset
  • Final prediction = average of all trees
  • Robust to overfitting
RΒ² = 0.637
GBR

Gradient Boosting

E(y, f(x)) = Ξ£{log(1 + exp(f(xi)) βˆ’ yiΒ·f(xi))}
  • Sequential ensemble (boosting)
  • Each tree corrects predecessor errors
  • Minimizes gradient of loss function
  • Combines weak learners β†’ strong model
RΒ² = 0.760
XGB β˜…

XGBoost

L = Ξ£ l(yα΅’, Ε·α΅’) + Ξ£ Ξ©(fβ‚–)
  • Advanced gradient boosting + regularization
  • Handles non-linear multivariate interactions
  • Optimizes bias-variance trade-off
  • Computationally efficient with high accuracy
RΒ² = 0.862 β˜…
SVR

Support Vector Regression

Ξ΅-insensitive tube
  • Ξ΅-insensitive loss function
  • Structural risk minimization
  • Kernel trick for non-linear mapping
  • High generalization capability
RΒ² = 0.857
11

Training Strategy & Validation

80:20 Split with 5-Fold Cross-Validation

Dataset Partitioning

TRAINING
70 datasets (80%)
TESTING
18 datasets (20%)
Total: 88 datasets

5-Fold Cross-Validation

TEST
TRAIN
TRAIN
TRAIN
TRAIN
Fold 1
TRAIN
TEST
TRAIN
TRAIN
TRAIN
Fold 2
TRAIN
TRAIN
TEST
TRAIN
TRAIN
Fold 3
TRAIN
TRAIN
TRAIN
TEST
TRAIN
Fold 4
TRAIN
TRAIN
TRAIN
TRAIN
TEST
Fold 5
Each fold: 4 parts training + 1 part testing β†’ 5 iterations β†’ averaged performance

Performance Metrics

RΒ² β€” Coefficient of Determination
RΒ² = 1 βˆ’ (SSres / SStot)
Proportion of variance explained by the model. Range: 0–1; higher is better.
RMSE β€” Root Mean Square Error
RMSE = √(Ξ£(yα΅’ βˆ’ Ε·α΅’)Β² / n)
Penalizes large errors. Units: mm/s. Lower is better.
MAE β€” Mean Absolute Error
MAE = Ξ£|yα΅’ βˆ’ Ε·α΅’| / n
Average magnitude of errors. Units: mm/s. Lower is better.
Train-Test Gap
Gap = RΒ²train βˆ’ RΒ²test
Measures overfitting. Values near 0 indicate good generalization.
12

Results Comparison

Comprehensive Performance Assessment Across All Models
Rank Model Type Test RΒ² RMSE (mm/s) MAE (mm/s) CV RΒ² (Mean Β± SD) Train-Test Gap Assessment
πŸ₯‡ 1 XGBoost ML 0.862 2.281 1.667 0.461 Β± 0.257 0.034 Highest test accuracy
πŸ₯ˆ 2 SVR ML 0.857 2.316 1.705 0.533 Β± 0.390 βˆ’0.026 Most generalizable
πŸ₯‰ 3 GBR ML 0.760 3.004 1.973 0.276 Β± 0.908 0.117 Moderate stability
4 RF ML 0.637 3.694 2.307 0.328 Β± 0.512 0.146 Lowest ML accuracy
5 MVR REG 0.456 β€” β€” β€” β€” Moderate interpretability
6 AAH EMP 0.243 β€” β€” β€” β€” Best empirical model
7 USBM EMP 0.152 β€” β€” β€” β€” Moderate empirical
8 BIS 6922 EMP 0.091 β€” β€” β€” β€” Poor performance
9 LAK EMP 0.007 β€” β€” β€” β€” Weakest model
13

Performance Plots

Predicted vs. Actual PPV β€” All Machine Learning Models
Random Forest (RF)
RΒ² = 0.637 Β· RMSE = 3.694 mm/s
Gradient Boosting Regressor (GBR)
RΒ² = 0.760 Β· RMSE = 3.004 mm/s
XGBoost β˜… Best Model
RΒ² = 0.862 Β· RMSE = 2.281 mm/s
Support Vector Regression (SVR)
RΒ² = 0.857 Β· RMSE = 2.316 mm/s
Dashed line = perfect prediction (y = x). Points closer to the line indicate higher accuracy.
14

SHAP Analysis

Shapley Additive Explanations β€” Feature Importance & Physical Interpretation

SHAP Summary Plot (XGBoost)

Distance (D)
1st
Charge (Q)
2nd
Spacing (S)
3rd
Burden (B)
4th
Hole Depth (H)
5th
← Negative SHAP (↓ PPV) Positive SHAP (↑ PPV) β†’
β–  High feature value β–  Low feature value

Physical Interpretation

1st
Distance (D) β€” Most Influential
↓ Negative contribution at high D
Seismic wave energy attenuates with distance due to geometric spreading and inelastic energy dissipation in rock mass
2nd
Max. Charge (Q) β€” Strong Positive
↑ Positive contribution at high Q
Higher explosive charge releases greater energy β†’ increased vibration amplitudes β†’ higher PPV
3rd
Spacing (S) β€” Moderate Positive
↑ Positive contribution at high S
Larger spacing reduces cumulative blasthole interaction β†’ enhanced PPV generation per hole
4th
Burden (B) β€” Weak Negative
↓ Negative contribution at high B
Larger burden increases confinement β†’ more energy absorbed in rock breakage β†’ lower PPV
5th
Hole Depth (H) β€” Weakest
↓ Slight negative contribution
Deeper holes distribute energy over larger rock volume β†’ marginal reduction in surface PPV
15

Key Findings

Summary of Model Performance and Critical Observations
RΒ² Comparison Across All Model Types
01
ML Outperforms All Other Approaches

Machine learning models achieve RΒ² = 0.637–0.862, significantly surpassing empirical (0.007–0.243) and regression (0.456) models

02
XGBoost is the Best Model

RΒ² = 0.862, RMSE = 2.281 mm/s, MAE = 1.667 mm/s β€” highest predictive accuracy with minimal train-test gap (0.034)

03
SVR Shows Best Generalization

Train-test gap = βˆ’0.026 (negative = test slightly outperforms train), indicating excellent generalization to unseen data

04
Distance is the Dominant Parameter

SHAP analysis confirms D has the highest impact on PPV, followed by Q, S, B, and H β€” validating the physical attenuation mechanism

05
Empirical Models Fundamentally Insufficient

All four empirical models show RΒ² < 0.25, confirming that 2-parameter scaled distance formulations cannot capture site-specific BIGV complexity

16

Engineering Interpretation

Practical Implications for Blast Design and Safety Management
πŸ›‘οΈ

Safety Compliance

XGBoost model enables accurate pre-blast PPV prediction, allowing engineers to verify compliance with DGMS limits (2–50 mm/s) before detonation β€” preventing structural damage to nearby infrastructure

πŸ“

Blast Design Optimization

SHAP-guided parameter sensitivity enables targeted optimization: reducing Q or increasing D are the most effective levers for PPV control; spacing adjustments offer secondary control

πŸ“‘

Monitoring Strategy

Distance is the most influential parameter β€” monitoring stations should be strategically placed at critical distances from blast faces, particularly near sensitive structures

βš™οΈ

Production Balance

The model enables optimization of charge per delay (Q) to maximize rock fragmentation while keeping PPV within regulatory limits β€” balancing productivity and safety

PPV Control Strategy β€” Based on Model Findings
↓ Reduce PPV
Increase D Β· Decrease Q Β· Decrease S Β· Increase B Β· Increase H
|
↑ Increase PPV
Decrease D Β· Increase Q Β· Increase S Β· Decrease B Β· Decrease H
17

Limitations

Constraints and Boundary Conditions of the Study
πŸ“‰
L1

Small Dataset Size

Only 88 datasets from 10 experimental blasts. Limited sample size may affect model generalization and statistical robustness, particularly for cross-validation stability (high SD in CV RΒ²)

Impact: High CV standard deviation (e.g., GBR: Β±0.908)
πŸ”§
L2

Limited Input Variables

Initiation sequence, stemming conditions, powder factor, and local geological discontinuities were maintained relatively consistent and not included as input variables

Impact: Potential unexplained variance in PPV predictions
πŸ—ΊοΈ
L3

Single Mine Site

All data collected from Samaleswari OCP only β€” a single geological formation (Barakar Formation, Gondwana Supergroup). Model may not generalize to different rock types or geological settings

Impact: Limited transferability to other mine sites
πŸͺ¨
L4

No Geotechnical Parameters

Rock mass properties (UCS, Young's modulus, P-wave velocity, RQD, joint orientation) were not included as input parameters despite their known influence on seismic wave propagation

Impact: Physical mechanisms partially unaccounted for
πŸ”„
L5

No Hyperparameter Optimization

Advanced optimization techniques (PSO, GA, Bayesian optimization) were not applied to tune ML model hyperparameters, which could further improve prediction accuracy

Impact: Potential for further accuracy improvement
πŸ“Š
L6

Low CV RΒ² Values

Cross-validation RΒ² values (0.276–0.533) are considerably lower than test RΒ² values (0.637–0.862), suggesting that model performance may vary with different data splits

Impact: Uncertainty in real-world deployment performance
18

Future Work

Directions for Enhanced PPV Prediction Research
Near-Term
πŸ“Š

Larger Datasets

Expand to 300+ datasets across multiple mine sites and geological formations to improve model reliability, generalization, and cross-validation stability

πŸͺ¨

Geotechnical Parameters

Include UCS, Young's modulus, P-wave velocity, RQD, and joint orientation as additional input variables to capture rock mass influence on BIGV

Mid-Term
πŸ”€

Hybrid ML Models

Develop hybrid models combining XGBoost with metaheuristic optimization (PSO, GA, WOA) for automated hyperparameter tuning and improved accuracy

🧠

Deep Learning

Explore LSTM, CNN-LSTM, and Transformer architectures for capturing temporal patterns in sequential blast data and wave propagation dynamics

Long-Term
🌐

Multi-Site Generalization

Develop transfer learning approaches to adapt models trained on one mine site to new geological settings with minimal additional data collection

⚑

Real-Time Prediction System

Deploy ML models in an integrated blast monitoring system for real-time PPV prediction and automated safety compliance verification before each blast

19

Conclusion

Summary of Contributions and Recommendations
C1

Conventional empirical equations (BIS 6922, AAH, LAK, USBM) are insufficient for accurate PPV estimation β€” they rely solely on scaled distance and ignore collective blast design parameter influence

C2

Multivariate regression provides improved interpretability (RΒ² = 0.456) and confirms direct relationships of Q and S with PPV, and inverse relationships of D, B, and H

C3

Machine learning models significantly outperform empirical and regression approaches by capturing complex non-linear multivariate interactions among blast parameters

C4

XGBoost is recommended as the best model for PPV prediction with RΒ² = 0.862, RMSE = 2.281 mm/s, and MAE = 1.667 mm/s β€” superior to all other models tested

C5

SHAP analysis confirms Distance > Charge > Spacing > Burden > Hole Depth in terms of influence on PPV β€” providing physically interpretable feature importance

Best Model Performance Summary

XGBoost Regressor
0.862
RΒ²
2.281
RMSE (mm/s)
1.667
MAE (mm/s)
0.034
Train-Test Gap
πŸ† Recommendation

XGBoost with 5-fold cross-validation is recommended for blast-induced ground vibration prediction in opencast mining operations. Future work should incorporate geotechnical parameters and larger multi-site datasets for improved generalization.

20

References

Key Citations β€” APA / Elsevier Format
[1] Ambraseys, N.R., & Hendron, A.J. (1968). Dynamic behaviour of rock masses. Rock Mechanics in Engineering Practices. Wiley, London, pp. 203–207.
[2] Chandrahas, N.S., et al. (2022). XG Boost Algorithm to Simultaneous Prediction of Rock Fragmentation and Induced Ground Vibration. Applied Sciences, 12(10), 5269. https://doi.org/10.3390/app12105269
[3] Duvall, W.I., & Petkof, B. (1959). Spherical propagation of explosion-generated strain pulses in rock. US Bureau of Mines Report No. 5481–5485.
[4] Fissha, Y., et al. (2024). Predicting ground vibration during rock blasting using relevance vector machine improved with dual kernels. Scientific Reports, 14(1). https://doi.org/10.1038/s41598-024-70939-w
[5] Ghosh, A., & Daemen, J.J. (1983). A simple new blast vibration predictor. 24th US Symposium on Rock Mechanics. AIME.
[6] Hosseini, S., et al. (2023). Assessment of ground vibration during blasting using different computational approaches. Scientific Reports, 13(1), 18582. https://doi.org/10.1038/s41598-023-46064-5
[7] IS-6922 (1998). Criteria for safety and design of structures subject to underground blast. Bureau of Indian Standards, New Delhi.
[8] Khandelwal, M., & Singh, T.N. (2009). Prediction of blast-induced ground vibration using artificial neural network. Int. J. Rock Mech. Min. Sci., 46(7), 1214–1222.
[9] Langefors, U., & KihlstrΓΆm, B. (1978). The Modern Technique of Rock Blasting. John Wiley & Sons.
[10] Nguyen, H., et al. (2019). Predicting Blast-Induced Ground Vibration in Open-Pit Mines Using Vibration Sensors and SVR. Sensors, 20(1), 132.
[11] Rana, A., et al. (2025). Sustainable management of near-field blast-induced ground vibration. Environmental Geochemistry and Health, 47(12).
[12] Roy, P.P. (1991). Vibration control in an opencast mine based on improved blast vibration predictors. Mining Science and Technology, 12(2), 157–165.
[13] Vishwakarma, A.K., et al. (2024). Prediction of ground vibration at surface for ring blasting in sublevel stoping. Mining, Metallurgy & Exploration, 41(3), 1567–1584.
[14] Yan, Y., et al. (2020). Random Forest for blast-induced ground vibration prediction. Engineering with Computers. AOA-optimised RF approach.
Acknowledgements: The authors thank Samaleswari Opencast Mine for sponsoring the project and the management of NIT Raipur for their cooperation and support.