elastic-wave-structure-emerge

Reusable arrival information across seismic models

Code and model release for “Reusable arrival information emerges across seismic models despite domain-dependent geometry.”

The central finding is function–geometry decoupling across recorded AE populations: arrival information remains transferable while training-associated cross-model agreement about waveform similarity can weaken or reverse. The study compares PNSN, SeismicXM, PhaseNet, EQTransformer and original self-supervised SeisLM. Each model contributes one released checkpoint; different training histories are contrasts, not isolated causal factors. Parameter scaling is context, not the axis of comparison.

Main Figure 2 now brings the five-model external comparison forward. Original-trained arrival readouts remain informative in glass beads and Min-U-Sil, while focal learning-delta relationships change. Geometry is measured on annotation-centred windows and timing on displaced windows: this is a population-level comparison, not a sample-wise causal relation. Native picking supplies direct cross-scale evidence. RF and NCF are supporting controls separating information accessibility, a pretraining contribution and absolute predictive performance; they do not independently establish function–geometry decoupling.

The positive conclusion is:

Arrival information remains transferable across populations with different global learning-delta geometry.

Compact PNSN already supports native cross-scale arrival transfer. The controls define the scope: random-network shared components also carry arrival information, standardized random RF features remain competitive, and the NCF period-wise median remains stronger than the neural ensemble. Native AE medians coexist with substantial error tails. Each route uses one released checkpoint; five models do not imply five independent development lineages or a magnitude-stratified generalization test. All favorable, unfavorable and null outcomes are retained.

Contents

• experiments/2026-08-26-cross-scale-waveform-transfer/: primary RF, AE, NCF, matched-random, convergence, ensemble, residual, baseline and uncertainty programs.
• experiments/2026-08-29-pnsn-final-state-dispersion/: PNSN final-state and endpoint dispersion readout controls.
• experiments/2026-08-29-phasenet-eqtransformer-ae-transfer/: native-output AE tests for the PhaseNet and EQTransformer results in main Fig. 3.
• experiments/2026-08-29-phasenet-eqt-ncf-readouts/: supplementary architecture-native NCF readouts for PhaseNet and EQTransformer.
• experiments/2026-08-30-four-model-ae-convergence/: common-input four-model learning-delta matrix, bootstrap and correspondence-permutation analysis.
• experiments/2026-09-07-nc-robustness/: RF factorial, displaced AE localization, expanded geometry, calibration sensitivity and synthetic boundary controls.
• experiments/2026-09-22-independent-validation/: executed five-model/external-material/shared-component experiments, locked plans and portable replay launcher.
• scripts/check_independent_validation.py: data-free recomputation from derived matrices/predictions; generates SI Tables 25–30 and current main Figure 2 (historical asset name Fig4_independent_validation).
• docs/INDEPENDENT_VALIDATION.md: source acquisition, environment, replay and interpretation limits.
• scripts/check_robustness_results.py: offline recomputation of the added predictions, matrices and uncertainty, with exact generation/checking of SI Tables 16–20.
• model_definitions/: the exact PNSN and SeismicXM PyTorch definitions needed by the programs.
• models/: the four released upstream checkpoints and SeisBench metadata used in the manuscript.
• scripts/check_reported_results.py: smoke/regression proof that recomputes the headline RF, AE and NCF values from deposited artifacts and asserts the primary convergence statistics.
• scripts/check_manuscript_consistency.py: submission gate that verifies checkpoint lineage and checks the artifact-derived claims in the main text and Supplementary Information, including the exploratory four-model matrix.
• reproducibility/: the complete curated archive of formal checkpoints, run configurations, histories, predictions, validation-selection records, uncertainty outputs, locked manifests and derived figure values.
• docs/: data contract, execution map, model provenance and licensing notes.

Installation

git xet install
git clone https://huggingface.co/cangyeone/elastic-wave-structure-emerge
cd elastic-wave-structure-emerge
conda env create -f environment.yml
conda activate elastic-wave-structure-emerge
python scripts/verify_release.py

When the manuscript repository is checked out beside this repository, run the cross-repository submission gate with:

python scripts/verify_release.py --full \
  --manuscript-dir ../overleaf-paper \
  --check-submission-tex

environment.yml describes the original study environment. The independent extension's separately recorded environment and replay steps are in docs/INDEPENDENT_VALIDATION.md; do not assume identical package versions across study stages. CPU and Apple MPS are supported; backend changes are not claimed to be numerically bit-identical.

Data status

The independent external test uses the open, ungated AEsNet archive, DOI 10.5281/zenodo.14639594, CC BY 4.0. Glass beads and Min-U-Sil are separate populations. Its raw recordings remain at the source; event selections and derived results are included with attribution. This public external source does not remove the original AE archive's access restriction.

Raw training and evaluation data are not included in this release. The datasets are cited or governed by their original sources; no HDF5 corpus, NumPy waveform array, preprocessed input trace, feature cache or descriptor cache is redistributed here.

The receiver-function arrays are publicly downloadable from the repository associated with Gong et al., which does not declare a data license. The AE records are held in the restricted-access AEbagging Zenodo deposition (DOI 10.5281/zenodo.15110538); its record metadata specify CC BY 4.0, and access requests are handled through Zenodo. The ambient-noise correlations and dispersion targets are publicly available without gating from SeisDispFusion-NCF revision afcd805 (DOI 10.57967/hf/9114) under Apache-2.0. That processed release derives its NCF waveforms from the CC BY 4.0 Mendeley Data source; both NCF releases should be cited. These source datasets are cited rather than duplicated here.

The repository does include the products needed to audit the reported computations without confusing them with source data: path-sanitized locked split manifests, source-file hashes, formal configurations, trained downstream checkpoints, histories, validation selections, test predictions, bootstrap/permutation outputs, aggregate tables and non-waveform figure values. It also includes a compact NCF implicit-geometry summary containing only split sizes and six correlations; the source waveforms and feature cache used to derive it remain excluded. These products are publicly distributed as derived reproducibility artifacts under this repository's GPL-3.0 license, not as copies of the cited input datasets. The historical manifests use ${RF_DATA_DIR}, ${AE_TRAIN_H5}, ${AE_TEST_H5} and ${NCF_H5} placeholders in place of machine-local paths.

After obtaining the datasets, create the locked, leakage-audited manifests with explicit paths:

python experiments/2026-08-26-cross-scale-waveform-transfer/prepare_manifests.py \
  --rf_dir /path/to/rf/data \
  --rf_reference_split /path/to/split_indices.npz \
  --ae_train /path/to/AE_train.h5 \
  --ae_test /path/to/AE_test.h5 \
  --ncf_h5 /path/to/ncf_disp_dataset.h5

See docs/DATA.md for the expected schemas and data-source boundary, docs/REPRODUCIBILITY.md for the execution order, and reproducibility/README.md for the artifact map. The exact main-figure generators are deposited under scripts/manuscript_figures/; they consume the archived derived results plus locally obtained cited inputs.

Checkpoints

The fifth route, SeisLM BASE, is fetched from its pinned official release rather than mirrored: SHA-256 6e97c53c562e9ba6bc96df85d8d208b89ff84e80a25468825f29bb5bace5eacb, 11,359,568 parameters in the full pretraining module. The inspected upstream repository has no declared redistribution license. See docs/INDEPENDENT_VALIDATION.md for the official source, exact revision and acquisition steps.

For PNSN and SeismicXM, the retained count is the frozen representation path used by the downstream analyses and excludes source decoders and target readouts. python scripts/audit_model_parameters.py strictly reloads both checkpoints and reproduces the detailed component-level counts. The reproducibility archive also contains all 338 saved downstream checkpoints from the completed primary runs: 240 formal paired runs and 98 temporal, residual and specialized-baseline runs. Supplementary architecture-native runs that did not save a head checkpoint retain their complete histories, metrics and predictions. Interrupted, smoke-test and quarantine runs are excluded. Full model and license provenance is in docs/MODELS.md.

License and citation

Versioned citation: use the immutable Git commit recorded in the manuscript's Code availability. DataCite metadata for repository DOI 10.57967/hf/10158 still record version 0a61cf7 (checked 23 September 2026), although its landing URL is the repository home. That DOI must not be presented as a newly registered identifier for this independent-validation release. No new DOI has been minted.

The repository's derived code is released under GPL-3.0. PhaseNet and EQTransformer checkpoint files retain their upstream MIT terms and attribution. See docs/MODELS.md before redistribution. The earlier deposit is identified by DOI 10.57967/hf/10158; cite the current code using its pinned Git revision, together with the manuscript and the five upstream model publications listed in docs/MODELS.md.