VsPrediction-Stacking

Hybrid I-CNN Stacking Ensemble for Shear-Wave Velocity Prediction from Conventional Well Logs

A reproducible MATLAB pipeline for shear-wave velocity (V_s) prediction from conventional well logs (GR, RHOB, NPHI, PHIE, V_p). The framework stacks four heterogeneous base learners (PNN, MLFFNN, DFFNN, 1D-CNN) under three meta-learner variants (Ridge stacker, I-CNN stacker, and a Hybrid multi-scale feature-fusion CNN), with Monte-Carlo-Dropout uncertainty quantification and an out-of-distribution (OOD) safety layer for auditable cross-formation deployment. Deployment-model selection is assessed across five random seeds to quantify reproducibility.

This repository accompanies:

Wibowo, R.C., Handoyo, Kumalasari, I.N., Winardhy, I.S., Amijaya, H., Normansyah, Aristoteles, Sarkowi, M. (2026). Hybrid I-CNN Stacking Ensemble for Shear-Wave Velocity Prediction from Conventional Well Logs: A Multi-Seed Robustness Study. Artificial Intelligence in Geosciences. DOI: (pending — update on acceptance).

⚠️ Verify before publishing: confirm the author list, title, and DOI against the final accepted manuscript; update the badges and the How to cite block accordingly.

---

📋 Table of contents

• Overview
• Key results
• Repository structure
• Requirements
• Quick start
• Configuration
• Reproducing paper results
• Data availability
• Output structure
• How to cite
• License
• Contact

---

🌍 Overview

Reliable shear-wave velocity supports reservoir characterization, AVO inversion, and geomechanical caprock assessment — workflows central to CCS/CCUS evaluation in mature basins. Dipole-sonic logging is expensive and absent from most legacy onshore wells. This pipeline closes that gap with machine learning while embedding uncertainty quantification and OOD safeguards that make cross-formation deployment auditable.

The pipeline implements:

1. Data ingestion & preprocessing — Excel reader, IQR + z-score outlier detection, kNN imputation of predictors only (the V_s target is never imputed), Savitzky-Golay denoising, z-score normalization, uniform depth resampling.
2. Target-clean supervised set — only rows with measured V_s are used for training; a hard assertion guarantees the target is never synthetic.
3. Feature selection — mRMR (top-k) and LASSO (10-fold CV), SHAP attribution, and a four-scenario ablation (S1 intersection, S2 all-features, S3 mRMR-only, S4 LASSO-only).
4. Base learners — PNN, MLFFNN, DFFNN, CNN1D.
5. Meta-learners — Ridge stacker, I-CNN stacker, and the Hybrid multi-scale feature-fusion CNN (kernels 3/5/7 → fusion → dense → regression).
6. Multi-seed robustness — stages 3–9 repeated over seeds {7, 99, 123, 2026, 42}; the canonical seed (42) supplies the deployed model and the figures.
7. Safety layer — MC-Dropout (T = 200), OOD detection (|z| > 3), physical clipping.
8. Deployment & indirect validation — Vp–Vs crossplot vs Castagna and Greenberg-Castagna, plus geomechanical post-processing (ν, G, K).

---

🎯 Key results

Applied to a North East Java Basin well pair — training well NEJ-1 (Kujung carbonate, 1415–1993 m, measured V_s) and blind well NEJ-2 (Plio-Pleistocene clastic overburden, no V_s). Held-out test R² reported as mean ± std across five seeds in the physical km/s domain:

Model	R² (mean ± std)	Note
Hybrid I-CNN	0.821 ± 0.031	⭐ Deployment model (wins 3/5 seeds)
Ridge stacker	0.812 ± 0.029	Operational alternative (tie-break, 2/5)
PNN	0.811 ± 0.029	Best base learner
I-CNN stacker	0.808 ± 0.022	Most stable (lowest seed variance)
DFFNN	0.790 ± 0.031
MLFFNN	0.777 ± 0.031
CNN1D	0.730 ± 0.041	Highest seed sensitivity
GC limestone	0.549 ± 0.040	Empirical (lithology-matched baseline)
GC shale	0.513 ± 0.044	Empirical baseline
Castagna mudrock	0.200 ± 0.052	Empirical baseline

Three principal findings:

1. The Hybrid multi-scale feature-fusion CNN is the top deployment model (mean R² = 0.821, winning outright in 3 of 5 seeds). The Ridge stacker and PNN are practically tied within seed-to-seed variability and serve as defensible operational alternatives; a ΔR² < 0.005 tie-break promotes the simpler Ridge stacker in the two remaining seeds.
2. Retaining all five logs (no feature selection, S2) gives the strongest stacking. LASSO assigns a zero coefficient only to RHOB, and because the mRMR top-3 set {GR, NPHI, V_p} is a subset of the LASSO-retained set, the intersection scenario S1 coincides with mRMR-only S3 by construction.
3. Cross-formation deployment is auditable but demanding. 13.5% of NEJ-2 samples are OOD-flagged (up to ~30% in the shallowest bin); within non-OOD samples the predicted V_s tracks Greenberg-Castagna shale within ~1% and lies 7.9% below Castagna mudrock — consistent with the clastic, shale-dominated lithology. A previously-reported +44% upward bias was traced to V_s target imputation and eliminated by the target-clean protocol.

---

📁 Repository structure

.
├── main_pipeline.m                 ← end-to-end entry point (script)
├── make_all_publication_figures.m
├── write_run_summary.m
├── config/
│   └── default_config.m            ← single source of truth for all settings
├── src/
│   ├── io/                         ← Excel reader, data audit
│   ├── preprocessing/              ← outliers, imputation (predictors only), denoise, resample
│   ├── features/                   ← mRMR, LASSO, scenarios, permutation importance
│   ├── models/                     ← PNN, MLFFNN, DFFNN, CNN1D (train + predict)
│   ├── stacking/                   ← Ridge, I-CNN, Hybrid I-CNN (train + predict)
│   ├── evaluation/                 ← model-results master, deployment selection, QC gate
│   ├── uncertainty/                ← MC-Dropout
│   ├── deployment/                 ← OOD, clipping, geomechanics, empirical baselines
│   ├── figures/                    ← diagnostic figures
│   ├── figures_publication/        ← 300-DPI publication figures
│   └── utils/                      ← logging, safe I/O helpers
├── data/                           ← (confidential well data NOT committed; see data/README.md)
├── results/                        ← generated at runtime (timestamped)
└── docs/
    └── REVISION_NOTES.md           ← changelog of the target-clean revision

---

🔧 Requirements

• MATLAB R2024b or later.
• Toolboxes: Statistics and Machine Learning (mRMR, LASSO, kNN, CV), Deep Learning (CNN1D, I-CNN, Hybrid I-CNN, MC-Dropout), Optimization (grid/Bayesian search), Signal Processing (Savitzky-Golay, resampling). Parallel Computing is optional.

Verify with ver in MATLAB.

---

🚀 Quick start

% From the repository root:
addpath(genpath('src'));
addpath('config');

% Place NEJ-1.xlsx and NEJ-2.xlsx under data/ (not distributed — see data/README.md)

% Run the full pipeline end-to-end (preprocessing → models → multi-seed → deployment)
main_pipeline

% Outputs are written to results/NEJ_excel_locked_final_<timestamp>/

main_pipeline is a script (not a function): it loads default_config() internally, runs the five-seed robustness loop, and uses the canonical seed (42) for deployment and figures.

---

🛠 Configuration

All settings live in config/default_config.m. Key fields:

cfg.data.train_file        = fullfile('data','NEJ-1.xlsx');   % calibrated well (has VS)
cfg.data.blind_file        = fullfile('data','NEJ-2.xlsx');   % blind well (no VS)
cfg.features.candidate_logs = {'GR','RHOB','NPHI','PHIE','VP'};
cfg.features.target         = 'VS';
cfg.features.mrmr_top_k     = 3;                              % S3 = top-3
cfg.stacking.candidates     = {'ridge','icnn','hybrid_icnn'};
cfg.selection.tie_threshold = 0.005;                          % ΔR² practical-tie band
cfg.uncertainty.mc_dropout_T = 200;
cfg.deployment.ood_zthresh   = 3;
cfg.deployment.vs_clip_kms    = [0.20, 4.50];

The five robustness seeds are set in main_pipeline.m (cfg.robustness.seeds = [7 99 123 2026 42], canonical seed listed last).

---

🧪 Reproducing paper results

The reported numbers come from the configuration in config/default_config.m with the five seeds above. Because the field data are confidential, reproduction without them uses statistically-matched synthetic wells (see data/README.md). Expected outcomes:

• Hybrid I-CNN mean R² ≈ 0.82; Ridge and PNN within seed variability.
• All-feature configuration (S2) is the strongest scenario.
• ~13.5% of blind-well samples OOD-flagged; predicted V_s tracking GC shale.

---

🔒 Data availability

The NEJ-1 / NEJ-2 well-log data cannot be redistributed (operator confidentiality) and are excluded from version control. Templates, a synthetic generator, and the final result tables under results/tables/ are provided so the workflow remains reproducible. See data/README.md.

---

📂 Output structure

Each run creates a timestamped folder:

results/NEJ_excel_locked_final_<timestamp>/
├── intermediate/      ← preprocessed + supervised CSVs (gitignored)
├── tables/            ← multi_seed_robustness.csv, feature_selection_scenarios.csv,
│                        all_scenarios_performance.csv, predictions_NEJ-2.csv, per_seed/
├── figures/           ← diagnostic figures
├── figures_publication/ ← 300-DPI publication figures
├── audit/             ← QC artifacts (gitignored)
└── logs/              ← timestamped run log (gitignored)

---

📚 How to cite

@article{Wibowo2026Vs,
  author  = {Wibowo, Rahmat Catur and Handoyo and Kumalasari, Isti Nur and
             Winardhy, Ignatius Sonny and Amijaya, Hendra and Normansyah and Aristoteles and Sarkowi, Muh},
  title   = {Hybrid I-CNN Stacking Ensemble for Shear-Wave Velocity Prediction
             from Conventional Well Logs: A Multi-Seed Robustness Study},
  journal = {Artificial Intelligence in Geosciences},
  year    = {2026},
  doi     = {10.xxxx/xxxxx}
}

A CITATION.cff file is included for GitHub's citation widget — update its title field to match the manuscript title above.

---

📄 License

MIT License — see LICENSE.

---

📧 Contact

Rahmat Catur Wibowo — Geological Engineering, Universitas Lampung, Indonesia · ORCID 0000-0003-2754-1803 · Code issues: https://github.com/rcw3712/VsPrediction-Stacking/issues