Materials failure analysis, against your own thirty years of it.
You have seen this fracture surface before, and the report that proves it is somewhere in thirty years of your own archive.
Three places this work actually breaks
The precedent is in your archive, not your memory
Environmental stress cracking, hydrogen embrittlement, fatigue striations, you have written them up dozens of times. The characterisation work is done. Finding which report it is done in is what costs the afternoon.
“Thirty years of cause-and-origin reports sit in folders only the senior partner can navigate.”
Micrographs and data sheets live apart from the reports
The image is on a share drive, the tensile data is in a spreadsheet, the material data sheet is a PDF from a supplier who has since revised it, and the conclusion that ties them together is in a report nobody has opened since it shipped.
“My report archive is my edge. It cannot end up training a tool my competitors use.”
A material property you cannot source is a liability
A yield strength, a glass transition temperature, a supplier’s revised spec, if the number in your report cannot be traced to the document it came from, the number is the weak point, not the analysis built on it.
“If it will not survive cross-examination, it is worse than useless, it loses the case.”
Three steps, and you approve the sources
Load the materials archive
Metallurgical and polymer failure reports, micrograph sets, tensile and impact data, supplier data sheets and their revisions. You approve every source; nothing outside them is read.
Describe the failure, not the filename
“Brittle fracture with beach marks near a weld toe.” “Crazing on a glass-filled nylon housing.” Answers come back with the report and page attached.
Compare against what you found before
Run the comparator across the archive to see which prior failures share the morphology, the material family and the service conditions, and where they diverge.
The comparison you would do by hand, run once
It lines up the prior failures that share this morphology and shows you where they differ, then stops for your read on which are genuinely comparable.
Input Fracture surface, PA66-GF30 housing, batch 4471
- Two are environmental stress cracking, same agent
- One shares the morphology but not the service temperature
Cited, or it tells you
No claim about what caused the failure in front of you, that is the metallurgy and it is yours. These are properties of the software, checkable on your own archive.
Cited to the report, not to a summary
A hardness value, a composition, a conclusion about a fracture mode: each one names the report and page it was taken from, so you can open the original and see the conditions it was measured under. An uncited number is the dangerous kind.
It tells you rather than interpolating
Ask about a material or a service condition your archive has not characterised and it says so. It will not reason across to the nearest alloy it has seen and present the result in the same tone as a measured one.
Related: Fire & explosion investigation → · Building & transportation → · Expert witness preparation →
Before you send us a document
Can it read micrographs and identify the failure mode?
No, and be sceptical of anything that says it can. It finds the reports, notes and data in your archive where a failure mode like the one you describe was already characterised by your own people, and it cites the page. Reading the fractograph is yours.
What about supplier data sheets that have been revised?
If both revisions are in your sources, both are findable, and each answer names the document it came from, so you can see which revision a value belongs to. That is exactly the case where an uncited number is dangerous.
Does it write the failure analysis report?
No. It retrieves and organises what is already in your file so you can review it. The analysis and the opinion stay yours. What it produces is checkable, because every line names the document and page behind it.
Could our archive end up training a tool a competing lab uses?
No. Your reports stay in your own tenant and we do not train on them. A lab’s value is that it has already characterised a failure mode a hundred times and written down what it found; handing that to a shared model would be handing over the reason clients call you rather than the lab down the road. Nothing about how this works requires it.
Test it on a failure you already characterised
Send one report you know well and describe the morphology from memory. Twenty minutes, and you will see whether it finds the right report and names the page.