IDA-PTW EEWS — Final Results (Task 15)

TL;DR — Ringkas — The final operational pipeline achieves composite R² = 0.7091 across 103 spectral periods under event-grouped 5-fold CV on 23,537 Indonesian accelerograms. Stage 0 URPD delivers AUC = 0.9136 with 100% Damaging Recall in 0.5 s, shrinking the near-field blind zone from 38 km to 11 km (Balanced) / 4 km (Aggressive). Stage 1 intensity classifier achieves 79.64% overall / 81.68% balanced accuracy under an MMI-hybrid 3-class discretisation via XGBoost+LightGBM soft-voting ensemble with SMOTE and Optuna tuning. The 103-period Stage-2 spectral regressor is class-conditional, marginalised over the Stage-1 posterior; oracle upper bound (Stage-1 perfect) is R² = 0.7779. 99.44% Golden Time Compliance. Retrospective validation on Cianjur 2022 and Sumedang 2024 yields 100% Stage-0 Damaging Recall. Pipeline operasional final mencapai composite R² = 0,7091 pada 103 periode spektral di bawah event-grouped 5-fold CV pada 23.537 accelerogram Indonesia. Stage 0 URPD memberikan AUC = 0,9136 dengan 100% Damaging Recall dalam 0,5 s, mengurangi near-field blind zone dari 38 km ke 11 km (Balanced) / 4 km (Aggressive). Stage 1 intensity classifier mencapai 79,64% overall / 81,68% balanced accuracy dengan diskretisasi 3-kelas MMI-hybrid via ensemble soft-voting XGBoost+LightGBM dengan SMOTE dan tuning Optuna. Regressor Stage-2 pada 103 periode adalah class-conditional, dimarginalisasi terhadap posterior Stage-1; oracle upper bound (Stage-1 sempurna) adalah R² = 0,7779. 99,44% Golden Time Compliance. Validasi retrospektif pada Cianjur 2022 dan Sumedang 2024 menghasilkan 100% Damaging Recall di Stage 0.

Headline Performance (Task 15 Final)

Performa Utama (Task 15 Final)

Metric	Metrik	Value	Nilai	Context	Konteks
Composite R² (103 periods, marginalised)	Composite R² (103 periode, marginalised)	0.7091	0,7091	Leading operational metric	Metrik operasional utama
R² operational band (T ≤ 2 s, 42 periods)	R² operational band (T ≤ 2 s, 42 periode)	0.6962	0,6962	Short-to-medium structural periods	Periode struktural pendek-menengah
R² oracle upper bound (all periods)	R² oracle upper bound (semua periode)	0.7779	0,7779	Stage-1-perfect theoretical ceiling	Plafon teoritis Stage-1 sempurna
Stage-1 OOF overall accuracy	Akurasi overall OOF Stage-1	79.64%	79,64%	MMI-hybrid 3-class ensemble	Ensemble 3-kelas MMI-hybrid
Stage-1 OOF balanced accuracy	Balanced accuracy OOF Stage-1	81.68%	81,68%	Safety-relevant (minority Damaging)	Relevan keselamatan (minoritas Damaging)
Stage-0 URPD AUC	AUC URPD Stage-0	0.9136	0,9136	Near-field sub-second discriminator	Diskriminator sub-detik near-field
Blind zone reduction (Balanced)	Reduksi blind zone (Balanced)	38 → 11 km	38 → 11 km	−71% with 100% Damaging Recall	−71% dengan 100% Damaging Recall
Blind zone reduction (Aggressive)	Reduksi blind zone (Aggressive)	38 → 4 km	38 → 4 km	−89% with FAR = 5%	−89% dengan FAR = 5%
Golden Time Compliance	Golden Time Compliance	99.44%	99,44%	Traces with latency budget ≤ P-S	Trace dengan budget latensi ≤ P-S
σ_total (T = 1.0 s, Al Atik decomposition)	σ_total (T = 1,0 s, dekomposisi Al Atik)	0.376	0,376	τ = 0.135, φ = 0.351 (intra-event dominant)	τ = 0,135, φ = 0,351 (intra-event dominan)

Source: reports/performance/ida_ptw_task15.json. Script: review_deliverables/task15_stage1_ensemble_optuna.py. CV: GroupKFold(5) by event_id, random_state = 42.

Sumber: reports/performance/ida_ptw_task15.json. Skrip: review_deliverables/task15_stage1_ensemble_optuna.py. CV: GroupKFold(5) berdasarkan event_id, random_state = 42.

Pipeline Optimisation Journey

Perjalanan Optimisasi Pipeline

The composite R² = 0.7091 was reached through a staged feature-engineering programme; each addition contributed an independently measurable gain under the same GroupKFold 5-fold CV protocol.

Composite R² = 0,7091 dicapai melalui program feature-engineering bertahap; setiap penambahan memberikan kontribusi yang terukur di bawah protokol CV GroupKFold 5-fold yang sama.

Stage	Tahap	Method Added	Metode Ditambah	M	R²	ΔR²
task 11	task 11	Class-conditional marginalisation (baseline)	Class-conditional marginalisation (baseline)	34	0.6074	—
task 12	task 12	+ Site features (VS30, log-distance)	+ Fitur site (VS30, log-distance)	36	0.6641	+0.057
task 14	task 14	+ Dai multi-window (2, 5, 10 s post-P)	+ Dai multi-window (2, 5, 10 s pasca-P)	66	0.6966	+0.033
task 15	task 15	+ XGB+LGB ensemble, SMOTE, Optuna, frequency features	+ ensemble XGB+LGB, SMOTE, Optuna, fitur frequency	90	0.7091	+0.013

Total improvement: +0.102 across four feature-engineering generations. Largest single contribution: site features (+0.057). Dai continuous-time-window adapted from K-NET Japan contributes the second-largest gain (+0.033), validating cross-regional transferability with ~0.10 degradation attributable to higher Sunda-arc site-path variability.

Total peningkatan: +0,102 dalam empat generasi feature-engineering. Kontribusi tunggal terbesar: fitur site (+0,057). Continuous-time-window Dai yang diadaptasi dari K-NET Jepang memberikan kontribusi terbesar kedua (+0,033), memvalidasi transferabilitas lintas-regional dengan degradasi ~0,10 yang dapat diatribusikan ke variabilitas site-path Sunda yang lebih tinggi.

90-Feature Architecture Breakdown

Rincian Arsitektur 90-Fitur

Family	Keluarga	M	Source	Sumber	Examples	Contoh
Time-domain rosers_ptw3 (base)	rosers_ptw3 time-domain (dasar)	34	`rosers_features_ptw3.csv`	`rosers_features_ptw3.csv`	τc, Pd, Pa, Pv, CAV, Ia, TP, dist_*	τc, Pd, Pa, Pv, CAV, Ia, TP, dist_*
Prior site metadata	Metadata site prior	2	`metadata_recalibrated.csv`	`metadata_recalibrated.csv`	metadata_vs30, metadata_log_dist	metadata_vs30, metadata_log_dist
Dai multi-window (2, 5, 10 s post-P)	Dai multi-window (2, 5, 10 s pasca-P)	30	`rosers_features_multi_window.csv`	`rosers_features_multi_window.csv`	Pa_w2s, Pv_w5s, Pd_w10s, Pav, Pad, Pvd	Pa_w2s, Pv_w5s, Pd_w10s, Pav, Pad, Pvd
Frequency-domain (3 s PTW, Z + H channels)	Frequency-domain (PTW 3 s, kanal Z + H)	24	`freq_features_ptw3.csv`	`freq_features_ptw3.csv`	spectral_centroid, bandwidth, rolloff85, band_1-5Hz	spectral_centroid, bandwidth, rolloff85, band_1-5Hz
Total	Total	90	merged on (trace_name, event_id) inner join				merged pada (trace_name, event_id) inner join

Four-Stage Pipeline at a Glance

Pipeline 4-Stage Sekilas

Each stage delivers a single targeted contribution. Stage outputs feed downstream; Stage 0 is independent for sub-second near-field alerts.

Setiap tahap memberikan satu kontribusi yang ditargetkan. Output tahap mengalir ke hilir; Stage 0 independen untuk alert near-field sub-detik.

Stage	Tahap	Purpose	Tujuan	Method	Metode	Key Metric	Metrik Kunci
Stage 0 — URPD	Stage 0 — URPD	Near-field blind-zone discriminator (0.5 s, binary Damaging alert)	Diskriminator near-field blind-zone (0,5 s, alert Damaging biner)	Gradient Boosting, 7 features	Gradient Boosting, 7 fitur	AUC 0.9136	AUC 0,9136
Stage 1 — Intensity Gate	Stage 1 — Intensity Gate	3-class intensity routing (Low < 2 gal, Med 2–50 gal, High ≥ 50 gal)	Routing intensitas 3-kelas (Low < 2 gal, Med 2–50 gal, High ≥ 50 gal)	XGB + LGB ensemble, SMOTE, Optuna	Ensemble XGB + LGB, SMOTE, Optuna	Acc 79.64% / Bal 81.68%	Acc 79,64% / Bal 81,68%
Stage 1.5	Stage 1.5	Distance regression ablation (negative result)	Ablasi regresi jarak (hasil negatif)	XGBoost over 5 C0–C4 feature sets	XGBoost pada 5 feature set C0–C4	R² ≤ 0 (all variants)	R² ≤ 0 (semua variant)
Stage 2 — Spectral Regressor	Stage 2 — Spectral Regressor	103-period Sa prediction, marginalised over Stage-1 posterior	Prediksi Sa 103-periode, dimarginalisasi terhadap posterior Stage-1	Class-conditional XGB + LGB blend	XGB + LGB blend class-conditional	R² = 0.7091	R² = 0,7091

Authoritative Source Files

File Sumber Otoritatif

review_deliverables/task15_stage1_ensemble_optuna.py — main training pipelinepipeline training utama
review_deliverables/task15a_extract_freq_features.py — frequency-domain feature extractorextractor fitur frequency-domain
review_deliverables/task13_extract_multiwindow_features.py — Dai multi-window feature extractorextractor fitur Dai multi-window
reports/performance/ida_ptw_task15.json — summary JSONringkasan JSON
reports/performance/ida_ptw_task15.csv — 103-period per-period R² resultshasil R² per-periode untuk 103 periode
reports/performance/ida_ptw_task15_anchors.csv — anchor periods (T = 0.303, 1.010, 3.030 s)periode anchor (T = 0,303, 1,010, 3,030 s)
reports/performance/ida_ptw_task15_per_fold.csv — per-fold accuracy and R²akurasi dan R² per-fold
reports/performance/ida_ptw_task15_optuna_trials.csv — Optuna hyperparameter search loglog pencarian hyperparameter Optuna
manuscript_draft_IEEE_v2.md — final manuscript draft, IEEE Access targetdraft manuskrip final, target IEEE Access
IEEE_cover_letter.md — IEEE Access cover lettercover letter IEEE Access

Retrospective Event Validation

Validasi Retrospektif Event

Two destructive near-field events held out from training demonstrate real-world applicability:

Dua event near-field destruktif yang ditahan dari training menunjukkan aplikabilitas dunia nyata:

Cianjur 2022-11-21, $M_w$ 5.6 — 100% Stage-0 Damaging Recall100% Damaging Recall di Stage-0
Sumedang 2024-01-01, $M_w$ 5.7 — 100% Stage-0 Damaging Recall100% Damaging Recall di Stage-0

Both events demonstrate that the IDA-PTW framework detects destructive near-field shaking within the Golden Time budget on events the model has never seen during training.

Kedua event menunjukkan bahwa framework IDA-PTW mendeteksi getaran near-field destruktif dalam budget Golden Time pada event yang belum pernah dilihat model selama training.

Submission Status

Status Submission

Target venue:Target venue: IEEE Access (Q1)
Manuscript: manuscript_draft_IEEE_v2.md — 924 lines, final (task 15 numbers integrated)Manuskrip: manuscript_draft_IEEE_v2.md — 924 baris, final (angka task 15 terintegrasi)
Cover letter: IEEE_cover_letter.md — framework-first narrativeCover letter: IEEE_cover_letter.md — narasi framework-first
Reproducibility bundle: review_bundle.zip (scripts + CSVs + feature files) — upload to Zenodo pendingBundle reproducibility: review_bundle.zip (script + CSV + file fitur) — upload ke Zenodo pending
Status: ready for submission pending Zenodo DOIStatus: siap submit menunggu DOI Zenodo

IDA-PTW EEWS — Final Operational Framework (Task 15)

IDA-PTW EEWS — Framework Operasional Final (Task 15)