Nonstructural Seismic Fragility Data: What the Records Can—and Cannot—Support

Public experimental data can make seismic research easier to reproduce and extend. Its usefulness depends on understanding what each record represents. A database row may describe one damage observation from a specimen that appears elsewhere in the same dataset. Before fitting a fragility model, researchers need to reconstruct those relationships, understand missing fields, and establish which observations provide independent evidence.

What the Nonstructural Element Database organizes

NIST’s Nonstructural Element Database, or NED, organizes observations and seismic fragility models for nonstructural building components. Its relational structure connects experimental records to published references and connects supporting observations to fragility models. That traceability helps researchers investigate where a model’s evidence came from.

The repository separates components, references, experimental observations, fragility models, and fragility curves into different data files. Its experimental schema defines a specimen as the identifier or name recorded in the source reference. It also provides fields for loading protocol, demand metric, damage description, prior damage, repairs, and inspection sequence. These definitions are documented in the official database schema.

These distinctions matter because finding a component category is only the beginning. Researchers still need to determine which records describe comparable configurations, how damage was observed, and whether the available measurements match their research question.

A dated snapshot, with clearly defined counts

An inspection of the repository’s experimental observation file on 30 September 2026 produced the following counts. They describe that repository snapshot; the public database can change as its maintainers add or revise records.

The counts were calculated directly from the records. The 135 pairs are a useful grouping aid, but they are not a verified count of independent physical specimens. Similarly, 13 reference identifiers do not establish 13 independent test campaigns. Publications can reuse, summarize, or extend earlier experiments.

Why specimen history changes the interpretation

The inspected records illustrate the issue. Partition specimen JAB A-4, under reference Mosqueda-2016, appears with a story drift ratio of 0.0026 and prior damage recorded as “No.” Another observation for the same label records a drift ratio of 0.0078 and prior damage as “Yes.” Those entries should trigger a review of the original testing sequence.

Treating both rows as unrelated specimens would discard information about their shared history. A random row-level split could also place one observation in a model’s training data and another in its evaluation data. The evaluation would then include evidence associated with a specimen the model had already encountered.

A research dataset therefore needs an explicit hierarchy: source document, campaign where identifiable, physical specimen, loading sequence, inspection, and observation. Reconstructing that hierarchy may require reading reports, figures, appendices, and test logs. A distinct label pair is a starting point for that work.

A blank field preserves a question

Missing history deserves its own category. Within the inspected gypsum partition subset, 146 rows had blank prior-damage fields and blank prior-damage-repair fields. All 535 rows had a blank specimen-inspection-sequence field. These counts describe the inspected export’s contents; they do not establish that the underlying information is absent from the original research.

The same principle applies to damage labels. A numeric rank needs its associated description and source definition. Researchers should not assume that an identical rank has an identical meaning across publications, or convert an empty field into a convenient default without documenting the decision.

Where the source resolves an ambiguity, the curated dataset should retain both the original value and the interpretation, with a precise citation. Where it does not, the uncertainty should remain visible.

Preparation before fitting a model

A practical preparation workflow can focus on five tasks:

• Preserve the snapshot. Record the retrieval date, repository revision where available, downloaded files, and file hashes so another researcher can reproduce the counts.
• Trace source relationships. Identify records that reuse earlier experiments, share specimens, or describe successive observations within one campaign.
• Check measurement meaning. Keep the demand metric and unit attached to every value. Review loading direction, protocol, amplitude, and measurement location before combining records.
• Reconstruct damage history. Document previous loading, observed damage, repairs, replacements, and the order of inspections wherever the sources support them.
• Define the analysis population. State which component details and observation types are included, why others are excluded, and what evidence supports each transformation.

Researchers also need to establish how the records were selected. Were all loading steps documented, or only steps associated with particular damage observations? Are observations without damage available? When a threshold is only known to fall between two test levels, that interval should be preserved. A collection of damage observations does not automatically provide a complete denominator for every probability estimate.

What a defensible first project can produce

An initial project can produce a source-linked data inventory, a documented component subset, and a benchmark that keeps related observations together. These are useful research outputs because they expose assumptions that would otherwise become hidden inside a model.

Predictive evaluation should then distinguish performance on new observations of familiar configurations from performance on genuinely unseen specimens, campaigns, or installation details. Any grouping rule should follow reconstructed source relationships. Where relationships remain unresolved, the evaluation should explain the resulting limitation.

Broader consequences require additional evidence. At review, NED’s README described an actively developing resource that did not yet contain consequence models or historical-event data. Experimental component records therefore do not, by themselves, establish a building’s reopening time. NIST’s functional-recovery research program treats nonstructural performance and recovery consequences as connected areas requiring further work.

For this snapshot, the immediate opportunity is concrete: reconstruct the provenance and histories behind a clearly bounded partition subset, publish the unresolved questions, and make subsequent analysis reproducible. Any later claim should state which additional evidence was obtained and which uncertainties remain.

Official sources and data

• NIST Nonstructural Element Database repository, including the project status, architecture, schema, and canonical data.
• NIST NED canonical experimental observation file, the dated source used for the snapshot counts above.
• NIST: Functional Recovery of Buildings and Lifelines, including current nonstructural systems research.