BRENNEN

Industries / Industrial equipment and services / Industrial services / Process safety consulting

Consequence and risk modeling

Consequence and risk modeling works out how far a release, fire or explosion could reach and how often it might happen, so that plants, buildings and towns can be kept a safe distance apart.

6

Milestones achieved, from 1978 to 2024.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1978

    Canvey Island's risk study

    On 20 June 1978 the UK Health and Safety Executive (HSE) published its investigation of potential hazards from the oil, gas and chemical installations on Canvey Island in the Thames estuary. The UK Atomic Energy Authority's Safety and Reliability Directorate did the study for HSE. A minister's written answer to Parliament called it "the first such inquiry to be carried out in this country and probably in the world." HSE gave its evidence as individual risk (the chance that a person at a given spot is killed) and societal risk (the chance of accidents that kill many people). HSE's C. Nussey wrote in 1992 that this avoided "sterile discussion" of the worst possible accident, but that the assessments were expensive because they were only partly computerised. After a 1981 public inquiry criticised HSE for not using quantified risk assessment, HSE built its own method, later known as RISKAT.

    LessonStating risk as numbers kept the argument off the worst imaginable accident, but the method spread only once it could be computed cheaply.

    api.parliament.ukicheme.org

    Credit

    Health and Safety Executive; UKAEA Safety and Reliability Directorate

  2. 1980

    Six Rotterdam plants quantified

    The Rijnmond public authority, the local government of the Rotterdam and Europoort area, had a detailed pilot study, called COVO, made of six hazardous installations. It used the "classical" method: list every way the equipment could fail, put a probability and a consequence on each, and sum them as risk to life. The results were drawn as risk contours on a map and as F-N curves, which plot how often accidents killing N or more people are expected. The method proved far too laborious to repeat for every installation in the area, so the authority commissioned simplified methods from the consultancy Technica. RIVM, the Dutch national institute for public health and the environment, says Dutch rules now prescribe one program, DNV's Safeti-NL, for calculating these risks.

    LessonA full risk study of each plant costs too much to use across a whole region.

    icheme.orgrivm.nl

    Credit

    Rijnmond Public Authority; E. F. Blokker, R. A. Cox and colleagues

  3. 2002

    Seven teams, one ammonia store

    In the EU's ASSURANCE project, seven teams each did a risk analysis of the same ammonia storage plant: a 15,000 t refrigerated tank, a ship terminal, a feed pipeline and ten pressurised tanks. Their estimates of how often pipelines and loading arms would fail differed by up to four orders of magnitude, a factor of 10,000. The causes included different pipe lengths, different generic failure data, valves and flanges left out, and whether part-time use was counted. Once the teams used common assumptions, the spread fell to one or two orders of magnitude for most scenarios.

    LessonA calculated risk depends as much on the analyst's assumptions as on the plant.

    backend.orbit.dtu.dk

    Credit

    European Commission Joint Research Centre, Ispra, and Risø National Laboratory (K. Lauridsen, I. Kozine, F. Markert, A. Amendola, M. Christou, M. Fiori)

  4. 2005

    A cloud that detonated

    The vapour cloud that exploded at the Buncefield oil depot in England on 10 December 2005 produced overpressures generally above 200 kPa across the whole cloud, as high over car parks and grassland as over tank farms and tree belts. The explosion was called unprecedented at the time. FABIG, an industry information group on fire and blast, notes that earlier explosions of large low-lying clouds in the open, such as Newark (1983), Naples (1985) and Saint Herblain (1991), should have warned hazard assessors. Later joint industry projects confirmed that dense vegetation can speed up a flame until it detonates, and FABIG's guidance keeps ground-level vegetation belts under 2 m deep.

    LessonA model that assumes a cloud in the open can only burn slowly may leave out the worst case.

    fabig.com

    Credit

    the Buncefield Explosion Mechanism joint industry projects, as summarised by FABIG

  5. 2016

    A 20-ton chlorine release

    In the Jack Rabbit II trials, run for the US Department of Homeland Security at the US Army's Dugway Proving Ground in Utah from August 2015 to September 2016, pressurised liquid chlorine was released and measured as it moved through a mock urban area and over open ground. The series ended with the full discharge of a 20-ton chlorine road tanker. An international exercise then compared 17 dispersion models with the measurements, and HSE found that DRIFT, the model it uses, gave reliable predictions for large chlorine releases.

    dhs.govckanprod.data-commons.k8s.ucar.eduhse.gov.uk

    Credit

    DHS Science and Technology Directorate; US Army Dugway Proving Ground; HSE (model comparison)

  6. 2024

    Twenty-one teams model ammonia

    From 2021 to 2024, 21 modelling groups from North America and Europe ran their dispersion models against two older series of ammonia field trials, Desert Tortoise and FLADIS. They returned 27 sets of results, from simple plume models to CFD (computational fluid dynamics, which divides the air into millions of small cells). The exercise, coordinated by Joe Chang of the RAND Corporation and Simon Gant of HSE, is being used to plan Jack Rabbit III, a future series of large ammonia releases. The inputs, each group's results and the comparison plots are published for download.

    admlc.com

    Credit

    RAND Corporation; HSE; the participating groups

Watch for

  1. Fast stand-ins for CFD

    S. Mackie, C. Bloodworth and C. Coffey built a surrogate model, a fast statistical stand-in trained on the inputs and outputs of CFD runs, from a Gaussian process model and a recurrent neural network. For one example leak modelled with FLACS-CFD, it gave cloud volumes and a spread of outcomes similar to the CFD predictions, cheaply enough to cover tens of thousands of scenarios.

    Achieved whenanother group reproduces the match on other kinds of leak, and the surrogate's risk results are checked against full CFD runs or field measurements.

    cetjournal.it

    Claimed by

    S. Mackie, C. Bloodworth and C. Coffey, Chemical Engineering Transactions, June 2025

Open

  1. 01

    An assumption ledger for QRAs

    A tool that reads a quantitative risk assessment (QRA) report and lists every number it rests on, with its page and the data source the report names: failure frequencies, hole sizes, pipe lengths, part-time factors, ignition probabilities and weather cases. It also lines up two reports side by side, so the assumptions behind a gap in their answers can be seen.

    The brief
    Credit · unclaimed
  2. 02

    A standing leaderboard for dispersion

    An open scoring harness and public leaderboard built on the published ammonia exercise files. It recomputes the exercise's comparisons for all 27 result sets, and lets anyone submit a new model, including fast machine-learning surrogates, to be scored on the same trials with the same statistics.

    The brief
    Credit · unclaimed

Open 01 · the brief

An assumption ledger for QRAs

Credit · unclaimed

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

What to build

A tool that reads a quantitative risk assessment (QRA) report and lists every number it rests on, with its page and the data source the report names: failure frequencies, hole sizes, pipe lengths, part-time factors, ignition probabilities and weather cases. It also lines up two reports side by side, so the assumptions behind a gap in their answers can be seen.

Why it is still open

In ASSURANCE, estimates of how often the same pipelines would fail differed up to 10,000-fold, and the causes were found only by breaking each team's work into small pieces in a coordinated benchmark. On 19 November 2026, GitHub and Crossref searches found no tool that extracts the assumptions in a QRA.

backend.orbit.dtu.dk

Built with what exists

QRA reports that are public in planning and permit files; an AI reviewer of long documents that cites the passage behind each item, such as the Claude-based reviewer in the Note "Seven more flagged" or an agent built with Cursor's Agent SDK (Note: "Findings that cite notes"); pdfplumber to pull tables from PDFs.

  1. First step

    Collect five QRA reports that are public on planning or permit portals, and extract their numbers by hand to make the answer key.

  2. Done looks like

    On 10 public QRA reports, the tool finds at least 95% of the numbers that two reviewers extract by hand, each with the right page, and invents no more than 2%. For one pair of reports on similar facilities, a reviewer confirms every difference the tool lists.

Open 02 · the brief

A standing leaderboard for dispersion

Credit · unclaimed

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

What to build

An open scoring harness and public leaderboard built on the published ammonia exercise files. It recomputes the exercise's comparisons for all 27 result sets, and lets anyone submit a new model, including fast machine-learning surrogates, to be scored on the same trials with the same statistics.

Why it is still open

The exercise's results are posted as zipped files and plots, not set up for new entries. Surrogate models are arriving (see Watch for), and REALM, a December 2025 benchmark for neural surrogates, says such models are mostly judged on "simplified, low-dimensional proxies"; its abstract lists propulsion and fire safety scenarios but no gas dispersion trials.

admlc.comarxiv.org

Built with what exists

The ADMLC downloads of inputs, results and trial definitions; Python with pandas and matplotlib; ALOHA (free, from EPA and NOAA) and OpenFOAM as reference entries; GitHub Actions to rescore each submission.

  1. First step

    Download the zipped results for Desert Tortoise trial 1 and recompute its predicted-versus-measured scatter plot.

  2. Done looks like

    The harness reproduces the published arc-maximum comparisons (the highest concentration on each measuring arc) for all 27 result sets, and at least two new models are scored and posted, one of them a machine-learning surrogate.