HAZOP and PHA Worksheets as AI Training Data for Process Safety

Why public HAZOP datasets do not exist

Public HAZOP data is scarce because worksheets are proprietary engineering records that describe how a specific plant can fail. A search for "hazop dataset" returns a split result: process-safety papers, plus computer-vision work that borrows the term to describe safety analysis of image datasets. The process-safety corpus that does surface annotated equipment, materials, states and consequences from HAZOP text and split it 8:1:1 for entity extraction [1], which shows the records are usable NLP data but not that you can download them.

Industry interest runs the other way: practitioners want to reuse HAZOP knowledge through scenario databases instead of leaving it locked in study reports [2], and trade press treats worksheets as managed process-safety data [3]. That is the opening for licensed sourcing. The data exists at almost every covered facility; it is simply held privately and needs rights review, redaction and a license before it can train anything.

What a HAZOP worksheet actually contains

A HAZOP worksheet is a structured table that follows the guide-word method described in IEC 61882: the team divides a system into nodes, states the design intent, combines guide words with parameters to generate deviations, and records the consequences and follow-up actions. In practice that produces rows with a stable set of columns, which is why the data suits supervised learning better than most engineering text.

The common columns are:

• Node: a line or vessel segment, usually with a P&ID reference, design intent and operating envelope (for example, "feed line from T-101 to R-201, 3 to 5 m3/h at 6 barg").
• Guide word and parameter: combinations such as NO + flow, MORE + pressure, REVERSE + flow, OTHER THAN + composition.
• Deviation, causes and consequences: the scenario chain, often with consequences stated both unmitigated and mitigated.
• Safeguards: alarms, interlocks, relief devices, procedures and SIS functions credited against each scenario.
• Severity, likelihood and risk rank: values from the operator's own risk matrix, which differs between companies.
• Recommendations: an action ID, owner, due date and, in the better systems, a close-out status.

Procedural HAZOPs examine operating steps instead of line segments, and IEC 61882 applies the method to procedures and sequences as well as continuous process lines [7]. Ask suppliers which variant a study used, because the two produce different row semantics and should not be mixed without a label.

HAZOP, What-If and LOPA records compared

Each PHA method yields a different supervision signal, so specify which you need before you request data. The OSHA Process Safety Management standard (29 CFR 1910.119 [6]) accepts several techniques, including What-If, checklist, HAZOP, FMEA and fault tree analysis, so facilities choose different methods for different units. LOPA is typically a follow-on that takes high-consequence HAZOP scenarios and quantifies independent protection layers.

LOPA data is small in row count but dense in numbers: initiating event frequencies, probability of failure on demand for each layer and the residual gap against a tolerable frequency. Check that IPL credits follow the operator's documented rules; inconsistent credit is a known quality problem in real LOPA records and makes good evaluation material if labeled as such.

Revalidation cycles make longitudinal data

PSM requires covered facilities to update and revalidate each PHA at least every five years (29 CFR 1910.119(e)(6)). A facility that has run PSM for decades may therefore hold several analyses of the same units. Those cycles are among the most valuable things a supplier can offer.

Revalidations range from a full redo to updating the original study for process changes only, so the delta between cycles shows how analysts handle management of change, new incident learning and retired equipment. For an AI team, paired cycles let you train or test "what should change in this node given these MOCs" rather than generating worksheets from scratch. Ask for the revalidation method per cycle, because a change-only update looks deceptively sparse next to a full redo.

Facilities in Program 3 of EPA's Risk Management Program must also perform a process hazard analysis (40 CFR 68.67) that closely mirrors PSM, so a single site may hold studies scoped for both regimes. Record which regime each study served.

Context you must license alongside the rows

A worksheet row without its node context is a weak training example. "NO flow: pump P-201 trips" means little unless the model knows what P-201 feeds, what the design intent was and which safeguards sit on that line. Plan to license one of three context levels:

1. Node descriptions only: text design intent, equipment tags and operating ranges. Easiest to redact and enough for many SFT tasks.
2. Marked-up P&ID extracts: drawings with node boundaries highlighted. Richer, but drawings carry more security-sensitive detail and need separate review; digitizing them is its own problem.
3. Supporting process safety information: relief sizing basis, alarm setpoints and SIS cause-and-effect matrices. Rarely needed for training and harder to release.

Structured exports help. Many operators keep PHAs in dedicated PHA software or an asset performance management system that stores analyses as linked records (systems, nodes, deviations, causes, consequences, safeguards) rather than a Word table [4]. A database export preserves those links; a PDF report destroys them, and re-parsing merged cells is a common source of misaligned cause and consequence pairs.

Illustrative record schema

The schema below shows a target structure for a licensed HAZOP row with node context and close-out outcome. Use it as a request template and adapt field names to the supplier's export.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "study_id": "PHA-2023-U12",
  "method": "HAZOP",
  "variant": "continuous",
  "cycle": 3,
  "revalidation_type": "update_for_changes",
  "node": {
    "node_id": "N-07",
    "description": "Reactor feed line, storage tank to reactor inlet",
    "design_intent": "Transfer liquid feed at 3-5 m3/h, 6 barg max",
    "pid_refs": ["PID-[REDACTED]-104"]
  },
  "guide_word": "MORE",
  "parameter": "flow",
  "deviation": "High feed flow to reactor",
  "causes": ["Flow control valve fails open"],
  "consequences": ["Reactor temperature excursion; potential relief to flare"],
  "safeguards": ["High-flow alarm", "High-temperature SIS trip", "Relief valve to flare"],
  "severity": 4, "likelihood": 2, "risk_rank": "B",
  "risk_matrix_ref": "operator_matrix_v2",
  "recommendations": [{
    "rec_id": "R-031",
    "text": "Verify SIS trip setpoint against runaway onset temperature",
    "status": "closed",
    "closed_date": "2024-05",
    "resolution": "Setpoint lowered; MOC reference recorded"
  }],
  "redaction": {"site": "pseudonymized", "chemicals": "class_only", "people": "removed"}
}

Ship the operator's risk matrix definition with every study. Without it, severity 4 at one company and severity 4 at another are different labels, and pooled models learn noise.

Security, ownership and redaction checks

HAZOP worksheets describe credible release scenarios at named facilities, so redaction is about security as much as privacy. Pseudonymize site names, unit numbers, tag prefixes and geographic hints; consider generalizing specific chemicals to hazard classes where your task allows; and remove attendee lists, which name the facilitator, scribe and operators. Ask whether any study touches chemicals that trigger additional security programs, and exclude those units if in doubt. Engineering drawings for some processes may also raise export-control questions, covered in export-controlled technical data checks.

Ownership is the other trap. Facilitation and engineering consultancies often hold the worksheets they produced, but the analysis usually belongs to the operator that commissioned it; confirm in the engagement terms rather than assuming a consultancy can license client studies. Buyers in that position may find the context on engineering consultancies as data holders useful, and on manufacturing operators for the facility side.

Document provenance in a machine-readable form. A Croissant-RAI record can capture provenance and labeling metadata in a machine-readable vocabulary [5], which simplifies your own internal review when legal, privacy and security approvers ask what the data is.

Acceptance checklist for a HAZOP or PHA dataset

Use this checklist when a sample arrives, before you accept a full delivery. It pairs with the general dataset acceptance testing process.

• Method and variant labeled per study (HAZOP continuous, procedural, What-If, LOPA).
• Node context present for at least the sampled rows, with design intent, not just a tag.
• Cause, consequence and safeguard alignment intact; spot-check rows where PDF parsing may have shifted cells.
• Risk matrix definition attached and consistent within each study.
• Recommendation outcomes linked by ID, with status and close date where available.
• Cycle and revalidation type recorded when multiple studies cover the same unit.
• Redaction log listing what was pseudonymized or generalized, with a reviewed sample.
• Duplicate and template rows flagged; boilerplate "operator training" safeguards copied across hundreds of rows inflate apparent coverage.
• Rights statement confirming the operator, not only the facilitator, approved release.

Building evaluations, not just training sets

Real worksheets are most valuable as evaluation data, because generated HAZOPs are hard to grade without expert reference sets. A strong benchmark holds out complete nodes, gives the model the node description and design intent, and scores recall of the expert team's deviations and causes, plausibility of credited safeguards and whether recommendations target the real gap. Keep held-out studies from the same site out of training to avoid leakage through shared tag names and boilerplate.

Expect expert disagreement. Two competent teams will not produce identical worksheets, so treat the reference as one acceptable answer and add a reviewer rubric. Related industrial records extend the signal: HSE incident and near-miss reports show which scenarios actually happened, PLC, SCADA and DCS alarm logs show how safeguards behaved, and FMEA worksheets cover equipment-level failure analysis that HAZOP deliberately skips. Field inspection reports can link recommendations to verified condition. The wider industry-specific operational data guide and the AI data hub cover adjacent categories.

How SourceX approaches HAZOP and PHA data requests

SourceX sources operational datasets, including engineering records and documents, from US companies on request; it does not hold HAZOP data in stock, and a request does not guarantee a match. You describe the worksheets you need, and SourceX looks for US businesses that hold them; every release is approved by the supplying company. Each dataset is rights-reviewed for ownership and consents, personal details are removed or replaced before delivery with the method recorded and a sample checked, and delivery happens only after an executed license and supplier approval. You can start a request on the SourceX buyer page.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Request HAZOP and PHA worksheet data

Describe the methods, node context and outcome fields you need, and SourceX will look for US companies that hold matching records and manage the licensing process if a supplier agrees. Nothing is contracted until the supplier approves, and pricing and allowed uses are set per deal in the license. Submit a HAZOP data request.