BRENNEN

Industries / Construction and property / Engineering and architecture / Geotechnical engineering

Earthquakes and liquefaction

This work predicts how the ground will behave in an earthquake and designs against it, including liquefaction, in which shaking turns loose, waterlogged sand into something close to a liquid.

6

Milestones achieved, from 1964 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1964

    Niigata's buildings tip over

    On 16 June 1964 a magnitude 7.5 earthquake struck Niigata, Japan. Loose, saturated sand under the city liquefied: apartment buildings at Kawagishi-cho tilted far over while staying almost undamaged, and the spans of the Showa Bridge fell when its foundations moved sideways. The University of Washington's soil liquefaction site says this earthquake and the Alaska earthquake of the same year brought liquefaction "to the attention of engineers and seismologists". Japan's Headquarters for Earthquake Research Promotion counts 26 dead; the University of Washington page gives 28.

    LessonA building can come through the shaking whole and still be lost when the ground under it gives way.

    depts.washington.eduhp1039.jishin.go.jp

    Credit

    Japan's Headquarters for Earthquake Research Promotion and the University of Washington, for the accounts

  2. 1971

    A simple liquefaction check

    H. B. Seed and I. M. Idriss published the "simplified procedure" in 1971. It compares the stress an earthquake puts on a layer of soil with the layer's resistance, read from the blow count of the standard penetration test, in which a sampler is hammered into the bottom of a borehole. A 1999 report by the US National Institute of Standards and Technology (NIST) says the procedure, revised several times since, is used "in the U.S. and throughout much of the world", and that versions based on the cone test and on shear wave velocity have followed.

    nehrpsearch.nist.gov

    Credit

    H. B. Seed and I. M. Idriss

  3. 1985

    Lake clay multiplies the shaking

    On 19 September 1985 a magnitude 8.1 earthquake centered about 250 miles (400 km) from Mexico City did its worst damage in the city. James Beck and John Hall of Caltech reported that a soft clay layer 50 to 100 feet (15 to 30 m) thick under the old lake bed amplified the shaking: a record on the lake bed peaked at 17% of gravity, four to five times the 3 to 4% recorded on firm ground. Damage was concentrated in buildings 6 to 15 stories tall, whose natural sway matched the ground's two-second rhythm. Beck and Hall wrote that as many as 10,000 people may have died; the Berkeley Seismological Laboratory gives official figures of 9,500 to 35,000.

    LessonMexico City's code already gave lake-bed buildings the largest design forces, but, Beck and Hall wrote, "no one anticipated the amount of amplification".

    calteches.library.caltech.eduearthquakes.berkeley.edu

    Credit

    Jorge Prince and colleagues at UNAM, who published the lake-bed record two days after the earthquake, and James Beck and John Hall, for the analysis

  4. 1995

    Treated ground holds in Kobe

    The 17 January 1995 Hyogoken-Nanbu earthquake, magnitude 6.9, liquefied much of Port Island and Rokko Island, artificial islands built of hydraulic fill (soil pumped in as a slurry) over soft clay in Kobe's harbor. A US geotechnical reconnaissance team found that sites densified beforehand, with sand compaction piles (Murayama's method of 1958) or vibrating rods (Saito's method of 1977), "sustained significantly less deformation and damage than did the adjacent untreated ground".

    LessonTreatment reduced damage without preventing it everywhere, and the team called for work relating settlement to how much ground was treated and how far beyond each structure.

    geerassociation.orggeerassociation.org

    Credit

    the reconnaissance team, whose report the Geotechnical Extreme Events Reconnaissance Association hosts

  5. 2011

    Land judged too damaged

    The Canterbury earthquakes of 2010 and 2011 caused widespread liquefaction in Christchurch, New Zealand. On 23 June 2011, on advice from geotechnical engineers, the government divided the city's residential land into four zones. The red zone, around 5,000 properties at first, was land "unlikely to be able to be rebuilt on for a considerable period of time", and the Crown offered to buy out insured owners at their 2007 rating value. By June 2023 Land Information New Zealand had handed 600 hectares along the Ōtākaro Avon River Corridor to the city council, after reconfiguring about 5,500 property titles.

    dpmc.govt.nzlinz.govt.nzpar.nsf.gov

    Credit

    the New Zealand government and the Canterbury Earthquake Recovery Authority

  6. 2020

    Liquefaction data goes public

    The Next-Generation Liquefaction (NGL) database went online in 2020. It holds site investigations, earthquakes and field observations from around the world under one shared layout, and a working group checks each upload against its source documents. In 2021 Mertcan Geyin, Brett Maurer and colleagues published about 15,000 cone-test case histories from three Canterbury earthquakes, drawn from the open New Zealand Geotechnical Database. They note that the widely used cone-based model of Boulanger and Idriss (2014) was trained on 255 case histories, so their set enlarges the data available "by at least 50x".

    doi.orgpar.nsf.gov

    Credit

    Scott Brandenberg, Jonathan Stewart and colleagues, for NGL; Mertcan Geyin, Brett Maurer, Brendon Bradley, Russell Green and Sjoerd van Ballegooy, for the Canterbury set

Watch for

  1. A design target for liquefaction

    US building design has moved to risk-targeted shaking levels, while liquefaction checks still rest on a uniform shaking level. Andrew Makdisi, Steven Kramer, Brett Maurer and six co-authors, working within the Building Seismic Safety Council's update of the national seismic provisions, propose that liquefaction checks for new buildings aim at a maximum acceptable chance of triggering of 2.5% in 50 years, through a new liquefaction-targeted peak ground acceleration.

    Achieved whenthe council's recommended provisions or the ASCE 7 standard adopt a liquefaction-targeted design level.

    usgs.gov

    Claimed by

    Makdisi and colleagues, Earthquake Spectra, July 2026

  2. Cone tests in rapid estimates

    The US Geological Survey's ground failure product estimates liquefaction in near real time after an earthquake from shaking maps and surface clues, and has no way to use site tests. Davis Engler, Eric Thompson, Brett Maurer and colleagues published a method that adds cone tests to it. Tested on the 1989 Loma Prieta earthquake with 373 Northern California cone tests, it predicted where liquefaction happened more accurately than the map-only model.

    Achieved whenthe live USGS product uses local test data and its accuracy on a later earthquake is reported by someone outside the team.

    usgs.gov

    Claimed by

    Engler and colleagues, Seismological Research Letters, September 2025

Open

  1. 01

    Case histories from old reports

    A reader that drafts entries for the NGL database from published reconnaissance reports and papers, including ones in Japanese and Chinese, for a person to check before upload: the site, its borehole or cone test data, the groundwater depth, the earthquake and what was seen at the surface.

    The brief
    Credit · unclaimed
  2. 02

    Scoreboard for liquefaction models

    A public benchmark for liquefaction prediction in which models are tested on whole earthquakes they never saw in training, with fixed splits, the Boulanger and Idriss method as the baseline, and a page where anyone can submit results.

    The brief
    Credit · unclaimed

Open 01 · the brief

Case histories from old reports

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A reader that drafts entries for the NGL database from published reconnaissance reports and papers, including ones in Japanese and Chinese, for a person to check before upload: the site, its borehole or cone test data, the groundwater depth, the earthquake and what was seen at the surface.

Why it is still open

Data reach NGL through an upload form, and a working group checks each upload against its source documents; the most used cone-based model was trained on 255 case histories. A 2026 US Geological Survey study used Google's Gemini 2.5 Pro to pull over 2,300 geocoded shaking reports out of records of the 1957 Daly City earthquake, but this research found no one doing the same for liquefaction case histories.

doi.orgusgs.gov

Built with what exists

The NGL database with its schema and Python tools in DesignSafe's Jupyter notebooks; reconnaissance reports on the GEER Association's site; a vision-language model that can read scanned tables and charts; a person checking every draft.

  1. First step

    Pick ten reviewed NGL sites whose source reports are online, draft them from the reports alone and compare.

  2. Done looks like

    For 50 sites already reviewed in NGL, a published field-by-field table showing the drafts match the reviewed values at 95% or better; then 100 new drafts submitted, with the share the working group accepts.

Open 02 · the brief

Scoreboard for liquefaction models

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public benchmark for liquefaction prediction in which models are tested on whole earthquakes they never saw in training, with fixed splits, the Boulanger and Idriss method as the baseline, and a page where anyone can submit results.

Why it is still open

A December 2025 paper says machine-learning liquefaction studies "remain fragmented: sampling strategies vary widely across studies without a unified benchmark." New models keep arriving on their own data, such as a 2026 conference abstract from the Willis Research Network and University College London built on 5,879 New Zealand cone-test locations. This research found no public leaderboard or agreed test split for the Canterbury or NGL data.

arxiv.orgmeetingorganizer.copernicus.org

Built with what exists

The Canterbury set of about 15,000 cone-test case histories on DesignSafe (doi 10.17603/ds2-tygh-ht91); the NGL database; Python with scikit-learn; a leaderboard hosted on GitHub or Hugging Face.

  1. First step

    Download the Canterbury set, hold out the February 2016 earthquake, and score the Boulanger and Idriss method on it.

  2. Done looks like

    Splits that hold out at least two whole earthquakes; published scores for the Boulanger and Idriss method and at least three published machine-learning models; submissions from at least two groups outside the builders.