The Liquefaction Hazard Engine computes one triggering procedure: Boulanger & Idriss (2014), CPT-based, deterministic. It draws the result against the two published reference curves that share its CPT axis, for visual comparison and nothing more, and returns a third on its own SPT axis. It draws on 8,634 in-situ test records from the Next Generation Liquefaction (NGL) database. This page states exactly what that means: what the engine computes, what it plots, what the data record is, and where it applies.
The engine does not implement a proprietary model, and it does not implement several models at once. It computes one published, peer-reviewed triggering procedure. It then draws that result against the two additional published boundary curves that share its CPT axis, and returns a third on the SPT axis, so the engineer can see, visually, where other established procedures would draw the line.
Semi-empirical, CPT-based liquefaction triggering procedure, computed in its deterministic
form. It is the only triggering procedure the engine computes: every cyclic stress ratio,
cyclic resistance ratio, and factor of safety comes from it. The factor of safety is
CRR divided by CSR at the engineer's own input point, with both terms at
Mw = 7.5 and σ′v = 1 atm; it is
returned as user_factor_of_safety alongside
factor_of_safety_method, and it is null when fines content is not supplied,
because the published closed form is defined in clean-sand-equivalent space.
The headline risk label is a different computation. It comes from a staged screening method that scores each throughgoing layer against eight susceptibility questions and rolls the borings up to one site label. Boulanger & Idriss (2014) sets the triggering numbers; the screening method sets the label.
Boulanger, R.W. & Idriss, I.M. (2014), “CPT and SPT Based Liquefaction Triggering Procedures,” Report UCD/CGM-14/01, UC Davis.
These are plotted for visual comparison. The engine does not compute them, does not run their equations against the site's input, and does not average them with the implemented procedure.
An SPT-based relationship, evaluated from the paper's own closed form (eq. 20) at a probability of liquefaction PL = 15%, moment magnitude Mw = 7.5, effective overburden stress σ′v = 1 atm, and fines content FC = 0%. Its abscissa is clean-sand-equivalent SPT N1,60, not cone tip resistance, so it is returned on its own SPT axis and is not drawn on the CPT plot: overlaying it there would put two different measurements on one x-axis. Described as probabilistic in the original publication; the curve is a single fixed slice of that model.
Cetin, K.O. et al. (2004), “Standard Penetration Test-Based Probabilistic and Deterministic Assessment of Seismic Soil Liquefaction Potential,” J. Geotech. Geoenviron. Eng., 130(12).
A CPT-based triggering correlation, probabilistic in the literature, plotted here as a reference boundary curve for comparison against the implemented procedure, not computed by the engine.
Moss, R.E.S., Seed, R.B., Kayen, R.E., Stewart, J.P., Der Kiureghian, A. & Cetin, K.O. (2006), “CPT-Based Probabilistic and Deterministic Assessment of In Situ Seismic Soil Liquefaction Potential,” J. Geotech. Geoenviron. Eng., 132(8).
An earlier semi-empirical triggering procedure from the same research lineage as the implemented procedure. Plotted as a reference boundary curve for comparison, not computed.
Idriss, I.M. & Boulanger, R.W. (2008), Soil Liquefaction During Earthquakes, EERI Monograph MNO-12, Earthquake Engineering Research Institute.
Published triggering procedures do not always agree. At real sites, with real CPT and SPT profiles, Boulanger & Idriss, Moss et al., and Idriss & Boulanger (2008) can draw different boundaries for the same soil column and the same shaking, and Cetin et al. (2004) can disagree again on the SPT axis. This is not a defect in any one procedure. It is what happens when independent research groups build triggering correlations from overlapping but non-identical case-history records, using different functional forms and different corrections.
Most tools hide that disagreement before the engineer ever sees it, by picking one model as the house standard and presenting its output as the answer. The engine computes one procedure, Boulanger & Idriss (2014), and draws the site's position against the two reference curves that share its axis: Idriss & Boulanger (2008) and Moss et al. (2006). Cetin et al. (2004) is returned on its own SPT axis rather than drawn on the CPT plot. The engine does not average the reference curves into the result. It shows them, so the engineer can see where the published literature would disagree with the implemented result.
That visual disagreement is information: it tells the engineer where the case-history record underlying these procedures is thin, where the site sits near a triggering threshold, and where professional judgment, not another decimal place, is what the decision actually turns on.
The engine draws on 8,634 in-situ test records (670 CPT soundings and 7,964 SPT samples) across 354 sites and 31 earthquakes, from the Next Generation Liquefaction (NGL) database (Brandenberg et al., 2020). Each record carries the test profile, groundwater depth, and shaking parameters documented for that site.
The engine searches this record for the test profiles most similar to the site under assessment and surfaces them alongside the computed result. This is a statement about the depth of the case-history record behind the engine, not a benchmarking claim, not a validation claim, and not a comparison against any other tool.
The NGL database is the same field-observation record the geotechnical community already builds on. GeoLiquefy cites it; GeoLiquefy does not own it.
The implemented triggering procedure computes the cyclic stress ratio (CSR), the cyclic resistance ratio (CRR), and the factor of safety against triggering. That arithmetic belongs entirely to Boulanger & Idriss (2014): the engine does not substitute an AI estimate for any of it, the reference curves are not recomputed by the AI either, and the memo model is handed the finished numbers to quote rather than any arithmetic to perform. Where a number could not be computed, the memo says so instead of estimating one.
AI's role is to draft the assessment memo and the engineering-readable explanation that accompanies it: which in-situ test records are most similar to the site in question, how the site's result sits against the published reference curves, and what the numbers mean in plain language. Every number that appears in that memo comes from the triggering procedure and the test-record search, not from the AI.
The draft is reviewed and sealed by the engineer of record. This is a computation and drafting instrument that sits in front of published methods, not a system that outputs a verdict on its own authority.
The implemented triggering procedure is calibrated on field case histories drawn predominantly from a limited set of well-instrumented earthquakes and regions. Applying it outside those conditions requires local validation: the procedure carries no automatic license to extend to geology or seismicity it was not built from.
The engine is a screening and assessment instrument. It does not replace site investigation, site response analysis, or the judgment of a licensed engineer. It exists to make published methods faster to apply and easier to defend, not to remove the steps that require a stamp.
Output quality is bounded by input quality. A CPT or SPT profile with sparse sampling, unclear groundwater conditions, or uncertain fines content will produce an assessment no more reliable than that profile deserves, regardless of which procedure is run against it.
Every output is a deterministic point estimate. The engine does not currently propagate input uncertainty, run Monte Carlo simulation, or produce confidence intervals. A factor of safety is a single number computed from a single set of inputs, not a probability of outcome.
The engine assesses triggering: whether liquefaction is expected to initiate. It does not model lateral-spread displacement or settlement. Consequence assessment beyond triggering remains the engineer's work.
The case-history record that underlies published triggering procedures is concentrated in a relatively small number of well-instrumented earthquakes, mostly outside India. GeoLiquefy is running an open research initiative to validate the implemented triggering procedure against Indian SPT and CPT profiles with known liquefaction outcomes.
This is research in progress, not a completed validation. We are not claiming Indian-specific calibration today. If your organization holds SPT or CPT profiles with documented liquefaction outcomes (from Bhuj, Sikkim, or any other Indian earthquake) and is willing to contribute them to this record, we would like to hear from you.
Contribute Data →Tiwari, A., Moss, R. & Gupta, A.K. (2026). On why the factor of safety compresses a multidimensional problem into a single, false-precision number, and how matching a site to documented in-situ test records in the NGL database restores the information that compression discards.
Download the Paper →Critical decisions require professional engineering review. Our platform augments, not replaces, human expertise.
Every assumption behind an assessment (which procedure, which test records, which input values) is documented and visible to the engineer, not buried in the output.
The engine shows where the site's result sits against the published reference curves, instead of hiding the comparison or collapsing it into a single score.
The AI-drafted memo is a starting point. The engineer of record reviews the reasoning behind every number and seals the final work.
No. The Boulanger & Idriss (2014) procedure computes the triggering result. AI drafts the assessment memo and the plain-language explanation around that result; it does not compute the number.
One implemented triggering procedure: Boulanger & Idriss (2014), CPT-based and deterministic. Two published boundary curves share its CPT axis and are drawn alongside it for visual comparison, not computed into the result: Idriss & Boulanger (2008) and Moss et al. (2006). Cetin et al. (2004) is SPT-based and is returned on its own axis rather than drawn on the CPT plot. The headline risk label is not a triggering output at all: it comes from a separate staged screening method.
No. Every output is a deterministic point estimate. The engine does not propagate input uncertainty, run Monte Carlo simulation, or produce confidence intervals.
No. The engine assesses triggering, whether liquefaction is expected to initiate. It does not model lateral-spread displacement or settlement.
8,634 in-situ test records (670 CPT soundings and 7,964 SPT samples) from 354 sites across 31 earthquakes in the Next Generation Liquefaction (NGL) database (Brandenberg et al., 2020).
No. It is a screening and assessment instrument. It does not replace site investigation, site response analysis, or the judgment of a licensed engineer. The engineer of record reviews the reasoning behind every number and seals the final work.