Automating Mill Test Certificate (MTC) and Material Test Report (MTR) Extraction for Steel and Metals Manufacturing

If you buy steel, aluminum, titanium, or any structural alloy in volume, your receiving dock is buried in PDFs. Every coil, every billet, every plate, every pipe joint arrives with a Mill Test Certificate (MTC) — also called a Material Test Report (MTR) or Certificate of Conformity (CoC) — that proves the material is the grade you ordered, was melted at the heat the supplier claims, and meets the chemistry, mechanical, and dimensional specs called out in your purchase order.

These documents are the legal backbone of metals supply chain traceability. In aerospace, they prove AS9100 conformance. In oil and gas, they prove API 5L or 5CT compliance. In automotive, they prove IATF 16949 traceability. In construction, they prove AISC weldability. Lose them, mis-key them, or fail to match them to incoming inventory — and the cost is not measured in clerical errors. It is measured in shut-down production lines, blocked customer shipments, failed audits, and in the worst case, recalled finished goods.

And yet, in most metals manufacturers, casting houses, service centers, and fabricators we have looked at, MTC handling is still entirely manual. A receiving clerk opens the supplier's PDF, eyeballs the chemistry table, types the heat number into the ERP, files the certificate in a network share named \\fileserver\MTRs\2026\ and hopes it never needs to be found again.

This guide is for quality engineers, supply chain managers, and digital transformation leaders in metals manufacturing who want to fix that. We will cover why MTR processing is uniquely hard, what every field on a typical mill cert means, how AI document extraction changes the economics, and exactly how to build an automated MTC ingestion pipeline using DocumentIQ.

Why Mill Test Certificate Processing Is So Hard

A mill test certificate looks like a structured document. In reality, it is the single most fragmented piece of paperwork in the metals supply chain.

Consider what shows up in your receiving inbox in a single week:

An EN 10204 3.1 certificate from a German steel mill, two pages, dense chemistry and mechanical tables, multi-language headers (German + English), heat numbers with embedded slash separators
An ASTM A370 mill test report from a US service center, one page, with the chemistry expressed in weight-percent and mechanical properties in ksi
An aerospace AMS 2750 cert from a heat-treat house, four pages, including pyrometry traceability, furnace charts, and operator signatures
A Chinese SGCC test report, scanned at 200 DPI, partly handwritten, with a red company chop overlapping the heat number
A 3.2 certificate from an Indian seamless pipe mill where the third-party witness (typically Lloyd's, TUV, or DNV) has counter-stamped the document and added a separate inspection annex

There is no industry standard for layout. EN 10204 defines the content (inspection types 2.1, 2.2, 3.1, 3.2) but not the format. ASTM A370 specifies the mechanical test methods but says nothing about how to lay them out on a page. The result is that every mill, every service center, every heat-treat shop, and every third-party inspection agency produces certificates that look fundamentally different — and your receiving team needs to extract the same set of values from all of them.

Now add the volume problem. A mid-sized steel service center receives between 200 and 2,000 MTRs per week. Aerospace tier-1 suppliers can see 5,000 per week across all incoming raw material. Each MTR contains 40 to 120 discrete data points that need to make it into the ERP, QMS, or material traceability system. Multiply 1,000 certs × 60 fields = 60,000 manual data-entry events per week, every week.

The error rate in manual MTR keying is consistently measured at 2% to 5% across the metals industry — and that is not just a clerical statistic. A single mis-keyed heat number can mean the wrong material gets released to a critical production order. A single missed chemistry deviation can mean a non-conforming part ships to an aerospace customer. The downstream cost of a wrong number on a mill cert is orders of magnitude higher than the cost of typing it correctly.

What's Actually on a Mill Test Certificate

Before we talk about extraction, it is worth being precise about what we are extracting. A complete MTC carries these data classes:

Identity and Traceability

Heat number / cast number — the unique identifier for a single melt of steel; the keystone of all downstream traceability
Lot number / batch number — sub-identifier for material processed together
Coil ID, slab ID, billet ID, pipe joint ID — physical piece identifiers
Customer purchase order number — your PO that triggered the order
Mill order number / supplier reference — the supplier's internal order ID
Material grade / specification — e.g., ASTM A572 Gr. 50, EN 10025-2 S355J2+N, API 5L X65, AMS 5510
Dimensions — diameter, wall thickness, length, width, gauge
Quantity / weight — pieces, theoretical weight, actual weight

Chemistry (Heat Analysis)

A chemistry table typically lists 8 to 25 elements with their measured weight-percent values:

Carbon (C), Manganese (Mn), Silicon (Si), Phosphorus (P), Sulfur (S)
Chromium (Cr), Nickel (Ni), Molybdenum (Mo), Vanadium (V), Copper (Cu)
Aluminum (Al), Niobium (Nb), Titanium (Ti), Boron (B)
Nitrogen (N), Hydrogen (H), Oxygen (O) — for vacuum-degassed grades
Carbon equivalent (CE / CEV / Pcm) — derived weldability metric
Sometimes a calculated field like "CE = C + Mn/6 + (Cr+Mo+V)/5 + (Ni+Cu)/15"

Mechanical Properties

Yield strength (R_eH, R_p0.2, YS) — usually in MPa or ksi
Tensile strength (R_m, UTS) — MPa or ksi
Elongation (A, El%) — percent, with gauge length (typically A5 or A50mm)
Reduction of area (Z, RA) — percent
Hardness — Brinell (HB / HBW), Rockwell (HRB / HRC), Vickers (HV)
Impact / Charpy V-notch energy — Joules at test temperature (e.g., 27J at -20°C)
Fracture appearance / shear area — percent ductile

Heat Treatment

Heat treatment condition — as-rolled, normalized, normalized & tempered, quench & tempered, annealed, solution-annealed
Heat treatment temperatures — austenitizing, tempering, holding times
Furnace charge number / pyrometry record (aerospace and nuclear)

Surface and Dimensional Tests

Ultrasonic test result — pass/fail, acceptance level (e.g., EN 10160 S2/E3)
Hydrostatic test pressure and duration (pipe)
Eddy current / magnetic particle / dye penetrant results
Dimensional inspection — OD, ID, wall thickness, ovality, straightness

Standards and Approval

Reference standards — every applicable spec the material claims compliance with
Certificate type — EN 10204 2.1, 2.2, 3.1, or 3.2; or ISO 10474 equivalent
Inspector signature, name, and date
Third-party witness stamp (for 3.2 certs) — TUV, Lloyd's, DNV, BV, SGS, ABS
Mill metallurgist or quality manager signature

Pulling 60 to 120 of these data points off a free-form PDF and dropping them into a structured table — at scale, across hundreds of mills, with traceability you can defend in an audit — is the actual problem.

How Most Manufacturers Handle MTRs Today

Three patterns dominate, and each has a structural ceiling.

Pattern 1: Pure manual keying

A receiving clerk opens the PDF, reads the certificate, and keys the heat number, grade, and a handful of "critical" fields into the ERP. The full certificate is filed (often just dragged into a folder by date), and the rest of the data — chemistry, mechanicals, dimensions — never makes it into the system at all.

Ceiling: 50–80 certs per FTE per day. 2–5% error rate. Most chemistry and mechanical data is invisible to downstream systems, which means engineering can't query "show me all the heats below 0.25% C from the last quarter" without manually reopening hundreds of PDFs.

Pattern 2: Template-based OCR

A second wave of manufacturers deployed traditional OCR (often ABBYY FlexiCapture or Kofax) with templates per supplier. The setup looks promising: define a template once, and every certificate from that mill flows through automatically.

The reality is messier:

Templates break the moment a mill changes layout — and most mills redesign their certificate templates every 12 to 24 months. Worse, some mills are non-deterministic: their certificate generator emits slightly different layouts for different product families.
Long tail of small suppliers — you might have 20 mills supplying 80% of volume and 200 service centers supplying the rest. Building templates for 220 suppliers is a multi-year IT project.
Multi-page tables and continuation logic — chemistry tables that span pages routinely break grid-detection parsers.
Scanned and image-only PDFs — OCR can read the text but cannot reconstruct row-column relationships from a noisy scan.

The end state is a template library that costs a quality engineering team 10–20 hours a week to maintain, with manual fallback for everything that doesn't fit.

Pattern 3: Outsourced data entry

Some manufacturers send the entire MTR inbox to an offshore BPO that keys the data into structured form and returns CSV files. This works, but typical turnaround is 24 to 72 hours, costs $1.50–$4.00 per certificate, and introduces a custody-of-data step that compliance teams in aerospace and defense will not allow.

This is where LLM-based document extraction reshapes the economics.

What AI-Powered Extraction Changes for MTRs

The breakthrough is the same one we have written about for bills of lading, freight invoices, and certificates of analysis: an LLM does not need a template. Hand it any mill test certificate — German, Chinese, Indian, Brazilian, US-domestic, EN 10204 2.2 or 3.2, single-page or fifteen-page — and it will identify the heat number, grade, chemistry, mechanicals, and dimensional fields from context, without per-supplier configuration.

For MTR processing specifically, this means:

Every mill format works on day one. No templates. No mapping per supplier. No update when a mill rebrands its cert layout.
Multi-language certs are handled natively. A German Schmelznummer, a Chinese 炉号, a Spanish Número de colada, an Italian Colata, and an English Heat No. all map to the same field.
Chemistry and mechanical tables become structured rows. The LLM understands that "C 0.18 / Mn 1.35 / Si 0.28" is a chemistry row — even when the table layout is unusual or spans pages.
Cross-document matching becomes a database operation. Once the heat number is structured on both the MTR and the receiving inspection record, matching is a join, not a manual lookup.
Confidence scoring routes only ambiguous certs to humans. Clean certs flow straight through to the ERP. Anything below a confidence threshold goes into a review queue for the quality team.

Production benchmarks from manufacturers running DocumentIQ for MTR ingestion are consistent: 97–99% field-level extraction accuracy on first pass, 90%+ touchless processing rate, and 80–95% reduction in receiving cycle time for material certification.

Building an MTR Extraction Pipeline in DocumentIQ

Here is the exact configuration most metals manufacturers, service centers, and aerospace tier-1s use to run MTR ingestion at scale. These are real field definitions and prompt patterns deployed today.

Step 1: Create a project per certificate type — or one master project

There are two viable patterns:

A. One project per major certificate family — separate projects for "Steel Plate MTRs," "Pipe MTRs," "Forged Bar MTRs," "Aerospace Heat-Treat Certs." Useful when the fields you care about vary significantly between product families.

B. One master MTR project with a superset of fields, where each field's instruction tells the LLM to return null if not applicable. Simpler to operate. This is what most teams settle on after the first quarter.

Set the project-level extraction prompt to provide domain context:

"These are mill test certificates for metallic raw materials. Values may be expressed in metric (MPa, °C, mm, kg) or imperial (ksi, °F, in, lb) units — preserve the original unit in the extracted value. Chemistry is reported in weight-percent unless otherwise stated. Heat numbers are alphanumeric and may contain dashes, slashes, or letter prefixes. Reference standards include EN 10025, EN 10204, ASTM A36/A572/A516/A106, API 5L/5CT, AMS, ASME SA-grade variants. Some certificates are bilingual."

Step 2: Define the field schema

A practical MTR schema runs 40–80 fields. Here are the highest-leverage ones with their extraction instructions:

Identity fields:

heat_number (text) — "Extract the heat number, cast number, or melt number — typically labelled Heat No., Cast No., Heat/Schmelz Nr., Colada, or 炉号. Return exactly as written including dashes and prefixes."
lot_number (text) — "Extract the lot number, batch number, or charge number if present; null if not."
customer_po (text) — "Extract the customer purchase order number — usually labelled Customer PO, P.O. No., Order No., or Käufer-Bestellnummer."
mill_order_number (text) — "Extract the supplier's internal mill order or works order reference."
material_grade (text) — "Extract the full material grade designation including any heat treatment suffix (e.g., S355J2+N, A572 Gr. 50, X65 PSL2, AMS 5510)."
material_specification (list) — "Extract all referenced material standards as a JSON array (e.g., ['ASTM A516', 'ASME SA-516', 'EN 10028-3'])."
product_form (text) — "Identify the product form: plate, sheet, coil, bar, tube, pipe, forging, casting, wire."
nominal_dimensions (text) — "Extract nominal dimensions as given (e.g., '12mm × 1500mm × 6000mm', 'OD 168.3 × WT 7.11mm', '3/4 inch dia × 12 ft')."

Chemistry — table-style extraction:

chemistry (list) — "Extract the chemical composition as a JSON array of objects, one per element. Each object has: element (chemical symbol — C, Mn, Si, P, S, Cr, Ni, Mo, etc.), value (numeric weight-percent), unit (default 'wt%'). Include only elements with a numeric value. If carbon equivalent (CE, CEV, Pcm) is given, include it as a separate object with element 'CE' or 'CEV'."

The model returns a clean structured array, e.g.:

[
  {"element": "C",  "value": 0.18, "unit": "wt%"},
  {"element": "Mn", "value": 1.35, "unit": "wt%"},
  {"element": "Si", "value": 0.28, "unit": "wt%"},
  {"element": "P",  "value": 0.012,"unit": "wt%"},
  {"element": "S",  "value": 0.008,"unit": "wt%"},
  {"element": "CEV","value": 0.42, "unit": "calculated"}
]

Mechanical properties — also table-style:

mechanical_properties (list) — "Extract mechanical test results as a JSON array of objects with: property (yield_strength, tensile_strength, elongation, reduction_of_area, hardness, charpy_impact), value (numeric), unit (MPa, ksi, %, HB, HRC, J), and condition (e.g., test temperature for impact: '-20°C')."

Heat treatment:

heat_treatment_condition (text) — "Extract the heat treatment delivery condition (as-rolled, normalized, normalized+tempered, quench+tempered, solution-annealed, annealed)."
heat_treatment_details (text) — "Extract any specific heat treatment parameters: austenitizing temperature, tempering temperature, holding time, quenching medium."

Non-destructive testing:

ut_result (text) — "Extract ultrasonic test result and acceptance level (e.g., 'Pass, EN 10160 S2/E3', 'Accept ASTM A578 Level A')."
ndt_other (list) — "Extract any other NDT results: hydrostatic test, MPI, dye-penetrant, eddy current — as an array of {test, result, standard} objects."

Certificate metadata:

certificate_type (text) — "Identify the inspection certificate type per EN 10204: '2.1', '2.2', '3.1', or '3.2'. If a 3.2 certificate, also identify the third-party witness in a separate field."
third_party_witness (text) — "If this is an EN 10204 3.2 certificate, extract the name of the witnessing inspection agency (TUV, Lloyd's, DNV, BV, SGS, ABS, etc.); null if not applicable."
inspector_name (text) — "Extract the name of the mill metallurgist or quality manager who signed the certificate."
certificate_date (date) — "Extract the certificate issue date in YYYY-MM-DD format."

Step 3: Choose an extraction mode

For MTRs, per-field mode is almost always the right choice. The certificate carries 50+ structured values, many in tables, and the cost of an error on a single field is high. Per-field mode dispatches one focused LLM call per field, in parallel across fields for each document, and consistently scores 1.5–3 percentage points higher on accuracy than batch mode on dense technical documents.

For lower-criticality use cases — for example, ingesting historical MTRs into an archive for retrospective search — batch mode is fine and roughly half the credit cost. See the complete IDP guide for how DocumentIQ prices both modes per billing plan.

Step 4: Use annotations for tricky mill formats

A handful of mills have certificate layouts that consistently confuse first-pass extraction — for example, a Chinese mill whose chemistry table is rotated 90° on page 2, or a heat-treat house that splits the mechanical table across two appendices. For these, open one representative cert in the DocumentIQ PDF viewer and use the annotation tool:

Draw a bounding box around the correct heat number on page 1
Map it to the heat_number field
Repeat for chemistry and mechanicals if needed

These few-shot annotations are injected as examples in subsequent extractions for similar certificates. Two or three annotations per problem mill typically lifts that mill's accuracy to the same level as everyone else.

Step 5: Wire confidence scoring into your QMS

Every extracted field comes back with a confidence score between 0 and 1. Set thresholds that match the risk profile of each field:

Heat number, grade, certificate type → require confidence ≥ 0.95; anything below routes to a quality reviewer
Chemistry and mechanicals → confidence ≥ 0.90; route the entire document to review on any element with lower confidence
Optional fields (inspector name, customer PO if missing) → accept ≥ 0.70 or null

The result is a workflow where 85–95% of MTRs flow straight through to the ERP with no human touch, and only the genuinely ambiguous ones reach the quality team — where their attention adds the most value.

Step 6: Cross-check against specification limits

This is where the structured data starts paying compound dividends. Once chemistry and mechanicals are in a structured table, you can validate every incoming heat against the spec limits before the material is released to production:

Is C ≤ 0.23 wt% for ASTM A516 Gr. 70?
Is Mn within 0.85–1.20 for an EN 10025 S355J2?
Is yield strength ≥ 345 MPa and tensile in 470–630 MPa?
Is Charpy at -20°C ≥ 27 J (longitudinal)?

In legacy workflows, this comparison happens (if it happens at all) when a quality engineer manually opens the PDF and eyeballs the numbers against a spec book. In a DocumentIQ pipeline, it happens automatically the moment the cert lands. Out-of-spec heats are flagged before the material is unloaded from the truck.

Step 7: Build the searchable traceability layer

Pipe the structured output into your QMS, ERP (SAP, Oracle, Microsoft Dynamics), or a dedicated material traceability database. Now your quality and engineering teams can run queries that were physically impossible before:

"Show me every plate from heat 4-A7821 still in inventory."
"Find all heats supplied between 2024 and 2026 with C above 0.20 wt% and CEV above 0.45."
"Which open production orders are consuming material from supplier X that had a Charpy deviation last quarter?"
"List every 3.2 cert witnessed by TUV for X65 pipe."

This is the same RAG pattern we described in Chat with Your PDFs — your quality team starts asking natural-language questions of an MTR archive that, until recently, was just a folder of opaque PDFs.

Compliance and Audit Considerations

For metals manufacturers operating under AS9100, IATF 16949, ISO 9001, API Q1, PED 2014/68/EU, or nuclear codes, MTR handling is not a back-office workflow — it is a regulated process. Three points matter when introducing AI extraction into the chain.

1. The original PDF remains the legal record

DocumentIQ does not replace the original mill certificate. The PDF as supplied — including signatures, witness stamps, and watermarks — remains the legal record of conformance and is stored unmodified in Azure Blob Storage (or S3, switchable via environment variable) with immutable retention policies. The extracted structured data is a derived index used for traceability and search; it is never the legal proof on its own.

2. Every extraction is traceable to its source

Each extracted field carries a reference to: the source document ID, the page number, the bounding region (when annotation was used), the model used, the model version, and the timestamp. If an auditor asks "where did this chemistry value come from?", you can trace it back to the exact page of the exact PDF in one click.

3. Human review remains the gatekeeper for critical fields

The confidence-threshold routing described above means a human reviewer touches every cert that does not pass a high-confidence bar. For aerospace and nuclear customers, many quality teams require 100% human review of EN 10204 3.2 certificates as a matter of policy — DocumentIQ supports this by pre-filling the extracted values for the reviewer to confirm or correct, which still cuts review time by 70–85% versus a blank form.

ROI: What This Actually Costs and Saves

Let us put rough numbers to it for a mid-sized service center or fabricator.

Assumptions:

1,000 MTRs processed per week
60 fields extracted per cert
Manual baseline: 4 minutes per cert for the critical subset, 8 minutes for full extraction
Loaded labor cost: $35/hour (US Midwest receiving/QA clerk)

Manual baseline cost (full extraction):

1,000 certs × 8 min = 133 hours/week
133 × $35 = $4,655/week
$242,000/year, plus a 2–5% error rate that produces downstream cost in production stoppages, wrong-material releases, and audit findings

DocumentIQ AI extraction (per-field mode, mid-tier plan):

1,000 certs × ~2.5 credits/cert (60-field cert in per-field mode on GPT-4o or Claude 3.5 Sonnet) = 2,500 credits/week
At typical mid-tier credit pricing, this lands in the $400–$900/week range, or $20k–$45k/year
Quality team time for review queue: roughly 5–10 hours/week (high-confidence pass-through means only ambiguous certs reach a human)
Net annual cost (software + review labor): $40k–$60k

Annualized savings: $180k–$200k, plus the downstream value of:

2–5% error rate dropping to <0.5% on critical fields
Full chemistry and mechanicals now structured, queryable, and validate-able
80–95% reduction in receiving cycle time
Audit response time measured in seconds, not days

Use the ROI calculator with your own volumes — the math is usually compelling well below 500 certs/week.

For deployments at scale or where the workflow integrates with custom QMS, ERP, or PLM systems (SAP S/4HANA, Oracle Cloud, Aras Innovator, Siemens Teamcenter), Algoscale's data engineering team and AI consulting practice handle the integration, validation, and change-management side.

Common Pitfalls — and How to Avoid Them

A few patterns we have seen trip up first-time MTR automation projects:

Trying to extract too many fields on day one. Start with the 15 critical ones (identity + grade + spec + chemistry array + mechanicals array). Add the rest after the pipeline is stable. The temptation to "extract everything" leads to schema sprawl and review-queue fatigue.
Treating chemistry and mechanicals as flat fields rather than arrays. If you define carbon_pct, manganese_pct, silicon_pct, …, niobium_pct as separate fields, every new element on a new cert breaks your schema. Define chemistry as a list field — see step 2 above.
Skipping confidence thresholds and pushing everything to the ERP. Without confidence routing, the 1–3% of low-confidence extractions pollute the downstream system and erode trust. Wire confidence into your release process from day one.
Not capturing the certificate type. EN 10204 2.2 (non-witnessed) and 3.2 (third-party witnessed) carry very different commercial and legal weight. Always extract certificate_type and third_party_witness — both for filtering and for audit.
Forgetting to back-index historical archives. Most teams have a network share with 5–15 years of past MTR PDFs. Once the extraction pipeline is running, batch-process the archive too. The retrospective traceability value is often as high as the go-forward value.

Adjacent Use Cases on the Same Platform

If you are building MTR automation, the next 90 days usually surface adjacent document types that benefit from the same pipeline:

Bill of materials extraction — pulling structured BOM data from engineering drawings and supplier quotes
Quality certificate compliance — extending the same workflow to ITPs, FAIs, and PPAPs
Warehouse receiving automation — matching MTRs to packing lists and POs on the dock
Invoice digitization — closing the AP loop against the same supplier records
Spec sheet extraction — pulling structured data from supplier technical data sheets
Contract intelligence — extracting price-escalation, force-majeure, and indexation clauses from long-term supply agreements (see also The Hidden Cost of Missing Price Escalation Clauses)

The pattern is consistent: any document where the same set of values needs to be pulled, validated, and indexed against operational data is a candidate for the same extraction layer.

Security and Deployment

Manufacturing data — especially aerospace, defense, and nuclear MTRs — has strict residency and confidentiality requirements. DocumentIQ runs in a single-tenant configuration on Azure Container Apps with the following defaults:

Encryption at rest (Azure Blob Storage with customer-managed keys available)
Encryption in transit (TLS 1.2+)
All secrets in Azure Key Vault, accessed via managed identity
Data never leaves your tenant, never used to train models
Audit logs for every extraction, review action, and user access event

For ITAR-controlled material, regulated aerospace deployments, or air-gapped configurations, Algoscale's data governance consulting and AI agent development team handle the security review, residency design, and compliance documentation needed to clear procurement.

Where to Go Next

If you are evaluating MTR automation, here is a sensible path:

Quantify your baseline. Run the ROI calculator with your weekly cert volume and average minutes per cert. The output is usually surprising even before you factor in error reduction.
Pick one product family. Steel plate, seamless pipe, forged bar — start with one and build the schema there. Expand once it is touchless.
Read the broader context. The intelligent document processing guide covers the underlying technology; the OCR vs LLM comparison explains why this works where template OCR did not.
Compare alternatives. See how DocumentIQ stacks up against ABBYY FlexiCapture, AWS Textract, Rossum, Nanonets, and Azure Document Intelligence.
Look at related case studies. Quality certificate compliance and bill of materials extraction are the closest analogues.

Mill test certificate processing is one of the highest-ROI document automation opportunities in metals manufacturing. The technology to handle it at 99% accuracy and 90%+ touchless throughput exists today; the only real question is how soon your quality team stops keying chemistry tables by hand.

Related reading on DocumentIQ:

Related Algoscale services: