IEEE Access Manuscript | Adaptive PTW-EEWS

Overview

Keywords: earthquake early warning, spectral acceleration, IDA-PTW, Feature Dichotomy, XGBoost, Java-Sunda subduction

Scope & Paper Differentiation

	Frontiers in Earth Science	IEEE Access (this page)
Dataset	Sunda v2 Geomean, IDA-filtered Java-Sumatra (95°–115°E)	Java-Sunda Trench curated, event-grouped
N traces	34,033	25,058
N events	329	338
N stations	388 BMKG	125 IA-BMKG
Distance range	2–1,438 km (median 384 km)	1–600 km (median ~124 km)
Magnitude range	M 4.0–6.9	M_w 1.7–6.2
Primary novelty	V_S30 Phase 1/2 expansion (+24.6% ΔR²)	Saturation-aware 4-stage pipeline + sigma decomposition
Composite R²	0.7309 (IDA Operational)	0.729 (103-period mean)

Key Metrics

0.729

Composite R²

0.840

Mean R² anchors

0.988

Stage 0 AUC

93.01%

Stage 1 Acc

91.09%

Damaging Recall

99.87%

Routing Fidelity

99.44%

GT Compliance

25,058

Traces

Core Tables

Table III-C · Intensity Class Distribution

Class	PGA threshold	N	%	PTW
Weak (MMI I–III)	<0.025 m/s² (<2.5 gal)	~6,522	26.0%	3 s
Moderate (MMI IV)	0.025–0.10 m/s² (2.5–10 gal)	~7,368	29.4%	4 s
Strong (MMI V)	0.10–0.50 m/s² (10–50 gal)	~7,711	30.8%	6 s
Severe/Damaging (MMI VI+)	≥0.50 m/s² (≥50 gal)	~3,457	13.8%	8 s
Total	—	25,058	100.0%	—

Source: reports/performance/intensity_correlation_metrics.csv (per-period raw counts 7,356 / 8,309 / 8,698 / 3,903 at T = 0.0 s, normalized to unique-trace total).

Table 11 · Fixed-Window Benchmark vs. IDA-PTW

Method	PTW (s)	R² PGA	R² Sa(0.3s)	R² Sa(1.0s)	R² Sa(3.0s)	Composite R²
Fixed	2	0.6749	0.8487	0.7901	0.7703	0.6774
Fixed	3	0.6941	0.8532	0.7994	0.7834	0.7014
Fixed	5	0.7181	0.8595	0.8073	0.7916	0.7289
Fixed	8	0.7357	0.8643	0.8136	0.7987	0.7423
Fixed	10	0.7475	0.8670	0.8197	0.8142	0.7536
IDA-PTW Operational	3–10	0.7548	0.7381	0.6992	0.7291	0.7286
Full-Wave (ceiling)	~341	0.8207	0.8110	0.7827	0.8163	0.8131

Source: reports/analysis/benchmark_results_fixed.csv (Fixed-window anchor rows) + reports/performance/xgboost_103_all_baselines.csv (IDA-PTW and Full-Wave per-period + 103-mean).

Table 15 · Sigma Decomposition

T (s)	R²	RMSE (log₁₀)	τ	φ	σ_total	W±0.5 (%)	W±1.0 (%)
PGA	0.832	0.245	0.324	0.619	0.698	62.9%	89.2%
0.1	0.848	0.235	0.372	0.708	0.800	54.9%	83.6%
0.3	0.876	0.201	0.470	0.768	0.900	43.9%	77.1%
0.5	0.866	0.215	0.495	0.754	0.902	41.1%	74.8%
1.0	0.830	0.282	0.515	0.691	0.862	41.8%	75.3%
3.0	0.820	0.315	0.450	0.542	0.705	58.9%	87.5%
5.0	0.808	0.340	0.406	0.493	0.639	65.2%	90.6%
Mean	0.840	0.262	0.458	0.598	0.755	54.4%	83.3%

Source: reports/performance/spectral_r2_performance.csv (R² per period) + reports/analysis/residual_report.md (RMSE, bias) + inter-/intra-event REML decomposition on pooled 5-fold residuals. Ratio φ²/σ²_total = 62.8% indicates intra-event (site-path) dominance.

Table 9 · PTW Selection Logic

Intensity	Primary criterion	PTW	N	%
Weak (MMI I–III)	PGA < 0.025 m/s²	3 s	6,522	26.0%
Moderate (MMI IV)	0.025 ≤ PGA < 0.10 m/s²	4 s	7,368	29.4%
Strong (MMI V)	0.10 ≤ PGA < 0.50 m/s²	6 s	7,711	30.8%
Severe/Damaging (MMI VI+)	PGA ≥ 0.50 m/s²	8 s	3,457	13.8%
Total	—	—	25,058	100.0%

Source: routing logic derived from Table III-C stratification; PTW assignment per Stage 1 intensity gate and Stage 1.5 distance-aware refinement.

Why Adaptive? — Research Prediction vs. Operational Warning (§VI.B)

Two Distinct Research Objectives Share the Same Dataset

Objective (i) — Research-Grade Prediction

Objective (ii) — Operational Warning

Why Composite R² Is the Wrong Metric for EEWS

Periode	Fixed 3s R²	Fixed 15s R²	Gap	Wilcoxon p
Sa(1.0 s)	0.6454	0.7198	+7.4%	< 0.001
Sa(3.0 s)	0.6089	0.6748	+6.6%	< 0.001
Sa(5.0 s)	0.5966	0.6654	+6.9%	< 0.001

Three Architectural Capabilities Fixed-2s Cannot Replicate

Sub-second near-field detection

Catalog-independent operation

Saturation resilience for M ≥ 7

Latency Per Class — IDA-PTW Is Not Uniformly Slower

Class	%	Alert need	Fixed-2s latency	IDA-PTW latency
Weak	26.0%	No alert	2 s	3 s
Moderate	29.4%	Advisory	2 s	4 s
Strong	30.8%	Protective action	2 s	6 s
Severe	13.8%	Critical alert	2 s	8 s
Dataset-weighted mean	100%	—	2.0 s	4.9 s

✓ The Verdict

SOTA Benchmark — Head-to-Head vs Published EEWS Methods

Comparison matrix across 12 evaluation dimensions

Study / Ref	Region	Method	N traces	M range	Window	Periods	Best R²	σ	Sub-sec alert	Autonomous (no cat.)	Multi- station	Saturation- aware	M≥7 tested
Wu & Kanamori 2008 [23]	Taiwan	P_d threshold	~1k	4–7	3 s	PGA only	—	—	❌	❌	❌	❌	✓
Colombelli et al. 2015 [29]	Japan (K-NET)	Peak amplitude + cumul. integrals	~8k	4–7.3	3–15 s	PGA	—	—	❌	❌	❌	✓ (partial)	✓
Hoshiba & Aoki 2015 [84]	Japan	PLUM (data assim.)	—	—	Real-time	PGV spatial	—	—	❌	✓ (network)	✓	—	✓
Jozinović et al. 2020 [31]	Italy (INGV)	Multi-station CNN	~5k	3.5–6.5	10 s	PGA/PGV	~0.75	—	❌	❌	✓	❌	❌
Münchmeyer et al. 2021 [77] TEAM	Italy	Transformer DL	~15k	3–7.5	10 s	PGA/PGV/MMI	~0.80	—	❌	⚠️ (partial)	✓	❌	✓ (partial)
Fayaz & Galasso 2022 [32] ROSERS	NGA-West2	Deep NN	~21k	4–7.9	3 s	5 periods	>0.85	—	❌	❌ (cat. dist)	❌	❌	✓
Dai et al. 2024 [35]	Japan (K-NET)	XGBoost 11 features	~8k	4–7.4	1–10 s	~10 periods	>0.84	—	❌	❌ (cat. dist)	❌	❌	✓
Shokrgozar & Chen 2025 [33]	NGA-West2	GradCAM Explainable DL	~17.5k	4–7.9	3 s	111 periods	UHS-valid	—	❌	❌ (cat. dist)	❌	❌	✓
Lin & Wu 2024 [85] P-alert Hualien	Taiwan	CAA cumul.	Event-based	Mw 7.4	15 s	Mag.	—	—	❌	⚠️	✓	✓	✓ (failed)
IDA-PTW (this) [—]	Java-Sunda subduction	4-stage XGBoost	25,058	1.7–6.2	0.5–8 s adaptive	103 periods	0.840 (anchor)	0.755	✅ 0.5 s (AUC 0.988)	✅ Stage 1.5 99.87% fid.	❌ (future via MQTT)	✅ Feature Dichotomy 5×	❌ (dataset M≤6.2)

Where IDA-PTW wins vs competitors (niche analysis)

WIN #1 ↗

Only method with sub-second alert

WIN #2 ↗

Only catalog-independent spectral prediction

WIN #3 ↗

Only saturation-aware algorithmic design

WIN #4 ↗

Most spectral periods for subduction zone

WIN #5 ↗

First EEWS with formal sigma decomposition

WIN #6 ↗

Largest Indonesian subduction dataset for EEWS ML

WIN #1 — Sub-Second Near-Field Alert (Stage 0 URPD)

📘 Analogi:

The Physics

How Stage 0 URPD Solves It

Why Spectral Centroid Works

🎯 Concrete Evidence — Retrospective Case Studies

Why Competitors Cannot Do This

ROSERS (Fayaz 2022): 3 s minimum
Dai et al. 2024: 1–10 s range, min 1 s
TEAM (Münchmeyer 2021): 10 s window
Colombelli 2015: 3–15 s adaptive
Shokrgozar 2025: 3 s

💡 Operational Impact

WIN #2 — Catalog-Independent Autonomous Operation (Stage 1.5)

📘 Analogi:

The Problem

How Stage 1.5 Solves It

🎯 Quantitative Evidence

Why Competitors Fail This Criterion

Method	Distance input	Autonomy?
ROSERS [32]	Requires catalog	❌
Dai et al. [35]	Requires catalog	❌
Shokrgozar [33]	Requires catalog	❌
TEAM [77]	Network triangulation	⚠️ Partial
Colombelli [29]	Requires catalog	❌
IDA-PTW	P-wave features only	✅ Full

Why This Matters for InaTEWS Specifically

💡 Operational Impact

WIN #3 — Saturation-Aware Algorithmic Design (Feature Dichotomy)

📘 Analogi:

The Physics of Saturation

🎯 The Tohoku 2011 Empirical Kill-Shot

The Feature Dichotomy — IDA-PTW's Algorithmic Solution

Class	Examples	Behaviour at M≥7	Stage 2 weight
Saturating	τ_c, P_d, P_v, P_a, TP, centroid	Plateau — unreliable	0.2× (5× down-weighted)
Non-saturating	CAV, CVAD, CVAV, Arias, PIv	Monotonic — accumulates during rupture	1.0× (full weight)

Empirical Validation

Experiment 3: Saturation Test Evidence

Period	Fixed-3s R²	Fixed-15s R²	Gap	Wilcoxon p
Sa(1.0 s)	0.6454	0.7198	+7.4 pp	< 0.001
Sa(3.0 s)	0.6089	0.6748	+6.6 pp	< 0.001
Sa(5.0 s)	0.5966	0.6654	+6.9 pp	< 0.001

💡 Operational Impact for Future Megathrust

WIN #4 — Most Spectral Periods Validated on Subduction Data

📘 Analogi:

Why 103 Periods?

Competitive Landscape — Period Coverage

Study	Dataset	Periods	Applicable to subduction?
Wu 2008 [23]	Taiwan	PGA only	⚠️ Mixed
Colombelli 2015 [29]	Japan K-NET	PGA only	✅ Japan subduction
Jozinović 2020 [31]	Italy	PGA/PGV	❌ Crustal
TEAM 2021 [77]	Italy	PGA/PGV/MMI	❌ Crustal
ROSERS 2022 [32]	NGA-West2	5 periods	❌ Crustal
Dai 2024 [35]	Japan K-NET	~10 periods	✅ Japan subduction
Shokrgozar 2025 [33]	NGA-West2	111 periods	❌ Crustal (not transferable)
IDA-PTW	Java-Sunda	103 periods	✅ First 100+ on subduction

Why Subduction-Specific Matters

🎯 Performance Evidence

WIN #5 — First EEWS with Formal Sigma Decomposition

📘 Analogi:

The Al Atik et al. (2010) Framework

IDA-PTW Decomposition Results

T (s)	τ	φ	σ_total	φ² / σ²_total
PGA	0.324	0.619	0.698	79%
0.3 s	0.470	0.768	0.900	73%
1.0 s	0.515	0.691	0.862	64%
3.0 s	0.450	0.542	0.705	59%
Mean	0.458	0.598	0.755	62.8%

🎯 The Critical Finding: φ > τ

Why This Decomposition Matters Operationally

Why No Prior EEWS Did This

💡 Policy Implications for BMKG

WIN #6 — Largest Indonesian Subduction Dataset for EEWS ML

📘 Analogi:

Dataset Size Comparison

Study	Region	N traces	Tectonic regime
Colombelli 2015 [29]	Japan K-NET	~8k	Subduction
Dai et al. 2024 [35]	Japan K-NET	~8k	Subduction
Jozinović 2020 [31]	Italy	~5k	Crustal
TEAM 2021 [77]	Italy	~15k	Crustal
Shokrgozar 2025 [33]	NGA-West2	~17.5k	Crustal
ROSERS 2022 [32]	NGA-West2	~21k	Crustal
IDA-PTW (this)	Java-Sunda	25,058	Subduction

Why Indonesian-Specific Validation Matters

🎯 Indo-Pacific Seismic Belt Impact

Data Provenance and Reproducibility

💡 Scientific Legacy

Where competitors win (honest limitations)

IDA-PTW's Unique Niche — Summary

Data Provenance

Manuscript artefact	Source file
N traces = 25,058; N events = 338; 103 periods	`experiments/rosers_features_ptw3.csv`
Intensity class distribution (Table III-C)	`reports/performance/intensity_correlation_metrics.csv`
Fixed-window benchmark (Table 11 Fixed-N rows)	`reports/analysis/benchmark_results_fixed.csv`
IDA-PTW per-period R² + Full-Wave ceiling (Table 11)	`reports/performance/xgboost_103_all_baselines.csv`
Composite R² = 0.729 (103-period mean)	`reports/performance/xgboost_103_all_baselines.csv` column mean
Information ceiling (Table 12)	`reports/validation_evidence_report.md` §4 + `reports/performance/comparison_r2_table.csv`
Saturation test N=1,204 (Table 13)	`reports/analysis/saturation_test_results.csv`
P-arrival sensitivity (Table 14)	`reports/analysis/p_arrival_sensitivity.csv`
Per-period R² curve (Table 15 R² column)	`reports/performance/spectral_r2_performance.csv`
Residual statistics μ, σ, RMSE (Table 15)	`reports/analysis/residual_report.md`
Zero-bias finding (\|μ\| < 0.004 log₁₀)	`reports/analysis/residual_report.md` §1
103-period R² curve visual (fig)	`reports/figures/per_period_r2_full.png`
Seismicity map (fig)	`reports/figures/seismicity_map.png`
Validation dashboard per-period (fig)	`reports/figures/validation_dashboard_T*.png`

Claims Pending Artefact Export

#	Claim	Location	Target artefact
1	Stage 0 URPD AUC = 0.988	Abstract, Table 4, Table 17	`reports/performance/stage0_urpd_auc.csv`
2	Stage 0 SHAP importances	Table 3, §IV.B	`reports/analysis/stage0_shap_values.csv`
3	Stage 1 accuracy 93.01%, Damaging Recall 91.09%	Table 7, Abstract, Table 17	`reports/performance/stage1_intensity_gate_metrics.csv`
4	Stage 1.5 routing fidelity 99.87%	Table 8, Abstract, Table 17	`reports/performance/stage15_distance_routing.csv`
5	Stage 1.5 variants C0–C4 R² table	Table 8 §IV.D	`reports/analysis/stage15_variants.csv`
6	Cianjur 2022 / Sumedang 2024 / Garut 2022 case studies	§IV.B	`reports/case_studies/{cianjur,sumedang,garut}_*.json`
7	Golden Time Compliance 99.44%	§V.F, Abstract	`reports/analysis/golden_time_per_trace.csv`

Research Narrative

1. Motivation & Research Context

2. Problem Framing — Four Compounding Failure Modes

Failure Mode 1 — The Near-Field Blind Zone

Failure Mode 2 — Magnitude Saturation (The Fixed-Window Paradox)

Failure Mode 3 — Distributional Bias in Fixed-Window Training

Failure Mode 4 — Catalog Dependency and the Composite-R² Illusion

3. Research Questions & Hypotheses

RQ	Research question	Hypothesis	Evaluated in
RQ1	Can a saturation-aware ML pipeline predict Sa(T) across 103 structural periods from ≤8 s of P-wave data on Indonesian subduction records?	H1: Non-saturating energy features dominate; Feature Dichotomy.	§V.A–C, Table 11, Fig 2–4
RQ2	Can the pipeline operate autonomously without BMKG catalog dependency?	H2: Autonomous distance ΔR² loss < 0.002.	§IV.D, Table 8
RQ3	How much irreducible aleatory uncertainty governs prediction, and what is the primary reduction pathway?	H3: φ > τ; V_S30 is binding constraint.	§V.E, Table 15
RQ4	Does adaptive windowing provide operational advantages beyond composite R²?	H4: Three architectural capabilities Fixed-2s cannot replicate.	§VI.B (Why Adaptive?)

4. The IDA-PTW Framework

Stage 0 — Ultra-Rapid P-wave Discriminator (URPD): Gradient Boosting on 7 spectral features from a 0.5-s window, AUC 0.988; reduces blind zone from 38 km to 11 km (human protection) / 4 km (infrastructure).
Stage 1 — XGBoost Intensity Gate (2.0 s): 93.01% accuracy, Damaging Recall 91.09%; routes traces to PTW ∈ {3, 4, 6, 8} s based on the Feature Dichotomy paradigm.
Stage 1.5 — XGBoost Epicentral Distance Regressor (+0.1 s): 99.87% routing fidelity; enables fully autonomous operation without catalog dependency (H2 confirmed).
Stage 2 — Ensemble of 412 XGBoost Spectral Regressors (4 PTW × 103 periods): anchored on non-saturating features (CVAD, CAV, Arias Intensity) with 5× down-weighted saturating features.

5. Feature Dichotomy — The Intellectual Core

Physical grounding: Two classes of P-wave information

❌ SATURATING features (nucleation-phase only)

✓ NON-SATURATING features (ongoing energy accumulation)

Algorithmic implementation: Per-stage feature weighting

Stage	Objective	Saturating features	Non-saturating features	Rationale
Stage 0 URPD (0.5 s)	Binary intensity flag	Used (spectral centroid dominant)	CAV only (cumulative in 0.5 s)	0.5-s window too short for CAV/Arias to accumulate; saturation acceptable for binary flag
Stage 1 Gate (2.0 s)	4-class intensity	Used at full weight (42 features)	Used at full weight	Classification decision is binary-like (not magnitude regression); saturation tolerable
Stage 2 Regression (3–8 s)	Full Sa(T) regression	5× down-weighted	Full weight — anchors prediction	Prevents saturating features from dominating gradient; critical for M_w ≥ 7 generalisation

Empirical validation: SHAP analysis

Theoretical mechanism (Festa et al. 2008)

6. Validation Strategy — Five Independent Experiments

7. Validation Findings

Experiment 1 — Fixed-Window Benchmark vs. IDA-PTW Adaptive (Table 11)

Experiment 2 — Information Ceiling Analysis (Table 12)

Experiment 3 — Saturation Test (Table 13)

Experiment 4 — P-Arrival Sensitivity (Table 14)

Experiment 5 — Sigma Decomposition (Table 15)

Stage 0, Stage 1, Stage 1.5 operational metrics

8. Research Gaps Addressed

#	Research gap	Prior literature	IDA-PTW resolution	Evidence
1	No near-field EEWS for Indonesia. No prior work provided sub-second damaging-event detection for Δ < 38 km.	Cremen & Galasso [18]; Minson et al. [19]	Stage 0 URPD: 0.5-s window, AUC 0.988.	Blind zone 38→11 km; Cianjur/Sumedang 100% Recall.
2	No adaptive window validated on subduction data. Existing methods operate on fixed PTW.	Colombelli et al. [29]; Dai et al. [35]	Stage 1 Gate: 42 features, 4 classes.	93.01% acc, 91.09% Recall on 338 events.
3	No autonomous spectral prediction. All prior models require catalog distance.	Boore et al. [48]; Fayaz & Galasso [32]; BMKG [46]	Stage 1.5 Regressor: log₁₀ Δ from P-wave.	99.87% routing; ΔR² < 0.002; 10.65 s alert.
4	No sigma decomposition for Indonesian EEWS.	Al Atik et al. [44]; Kotha et al. [62]	REML decomposition on IA-BMKG data.	τ=0.458, φ=0.598, σ=0.755; V_S30 roadmap.
5	No comprehensive multi-experiment validation.	Dai et al. [35]; Fayaz & Galasso [32]	5 experiments, 25,058 traces, GroupKFold.	Tables 11–15; H1–H4 tested.
6	No algorithmic formalisation of Feature Dichotomy.	Wu et al. [22]; Lancieri & Zollo [24]; Festa et al. [25], [65]	5× down-ranking via XGBoost feature_weights.	SHAP inversion confirms; algorithmic primitive.

Additional meta-contribution: Reframing of EEWS success metrics (§VI.B)

Summary:

9. Theoretical Significance Beyond IDA-PTW

10. Operational Implications for InaTEWS

11. Limitations & Boundary Conditions

12. Future Directions

Synthesis:

Validated Figures with Narration

1 Java-Sunda Study Area — Seismicity & Station Network Map

reports/figures/seismicity_map.png

Corresponds to Table III-C (Intensity Class Distribution)

2 Stage 2 Per-Period Accuracy Curve (R² vs Period)

reports/figures/per_period_r2_full.png

Corresponds to Table 15 (Sigma Decomposition)

3 Complete R² Spectrum Against 103 Periods (Zhu 2024 Horizontal Features)

reports/figures/xgboost_103_curve.png

Corresponds to Table 11 (IDA-PTW Composite R²)

4 Fixed-Window vs IDA-PTW vs Full-Wave vs Newmark-Beta (Information Ceiling)

reports/figures/comparison_fixed_vs_ida_vs_fullwave.png

Corresponds to Table 11 (Fixed vs IDA-PTW)

5 Stage 2 Intensity Gating — Breaking the R² > 0.8 Barrier

reports/figures/xgboost_gated_stage2_comparison.png

How to read the chart

The three curves

🟢 Green — Full-Wave S-Phase Upper Bound (Theoretical Limit)

⚪ Grey Dashed — Stage-1 Unfiltered (28.2k) Baseline 3s

🔴 Red Solid — Stage-2 Intensity Gated (Top 25k) 3s

The pink shaded band — Accuracy Gain from Intensity Gating

The valley at T ≈ 1 s — Why all three curves dip

Regime	Dominated by	Why predictable
T < 0.1 s	Site / near-surface	Spectral centroid, V_S30, H/V ratios directly relevant
T ≈ 1 s	All three (hardest)	Source + path + site all contribute; non-linear interactions; fundamental period of critical buildings
T > 3 s	Source / path	Metadata (distance, depth) correlates strongly

The R² = 0.80 threshold (orange dotted)

Takeaway for the manuscript

Supports Feature Dichotomy paradigm · Click for Table 9 (PTW Selection)

6 Prediction Accuracy & Error Stratified by PGA Intensity

reports/figures/intensity_vs_accuracy_correlation.png

Corresponds to Table III-C (4-bin intensity stratification)

7 Scientific Validation Dashboard — T=0.3 s Anchor Period

reports/figures/validation_dashboard_T0.300.png

Supports Zero-Bias finding (Table 15 key finding #1)

8 Validation Dashboards — T=1.0 s & T=3.0 s (Side-by-side)

reports/figures/validation_dashboard_T1.000.png + validation_dashboard_T3.000.png

Supports Period-dependent Sigma Pattern (Table 15 key finding #4)

Provenance:

🎓 Monev Supplements — May 2026 Update NEW

Single-Slide PPTX (9 files)

Validation A · GMPE + ML Baseline

6-curve chart with GMPE Atkinson-Boore

Download .pptx

Derivasi Blind Zone

Mathematical defense for 38→4 km claim

Download .pptx

Sitasi Blind Zone Peer-Review

Distribution chart 10 peer-reviewed sources

Download .pptx

4-Lapis Defense Δ R² < 0,02

Four-layer academic defense

Download .pptx

Oracle Bound Hierarchy

Bayesian decision theory bar chart

Download .pptx

Roadmap Disertasi 3-Paper

Strategic dissertation roadmap

Download .pptx

Proses scwfparams PSA5/DRS5

Pasca-pengolahan SeisComP3

Download .pptx

Feature Importance per Stage

3-stage cascade FI

Download .pptx

Validation A · GMPE Baseline (alt)

Alternative GMPE-only version

Download .pptx

Formal DOCX Explanations (9 files)

Penjelasan PSA vs True Sa vs Sa(T)

9 pages · Theoretical comparison · Nigam-Jennings simulation · why scwfparams produces both

Download .docx

Sitasi Blind Zone Peer-Review

10 peer-reviewed citations + BibTeX entries + 3 ready-to-cite templates

Download .docx

Derivasi Matematis Blind Zone

13 pages · First-principles physics derivation 38→11→4 km

Download .docx

Framework Al Atik 2010

8 pages · σ²=τ²+φ² · variance hierarchy · IDA-PTW application

Download .docx

Oracle Bound + Bayesian Theory

8 pages · Risk function · 3 predictor forms derived · investment direction

Download .docx

Feature Importance + Glossary

18 pages · 90-feature glossary + per-stage narrative + interpretation

Download .docx

Narasi Roadmap Disertasi

10 pages · 10-segment narrative · 8-question FAQ · Q1 2027 timeline

Download .docx

Narasi Konsolidasi IDA-PTW

134 pages · 26 narration chapters merged · full presentation script

Download .docx

Narasi Kesimpulan Akhir

8 pages · 3-Pesan Utama (Paradigma · Empirik · Dampak) · closing speech

Download .docx

Generated Figures (22 PNG files)

Downloads

📄 Read Full Manuscript (In-Browser)

902 lines · HTML with figures, tables, math

Open Manuscript ↗

manuscript_IEEE_Access.md

902 lines · Markdown source

Download .md

Manuscript_Audit_Report.docx

18 pages · DOCX

Download .docx