Materiascope is a materials engineering tool that answers property and selection questions and shows where every number came from — which database, which entry, and whether it was measured, computed or predicted. It reads open computational sources and cites them. It does not decide for you, and it does not give you a number it cannot attribute.
Same alloy, two standards, two product forms. Open the full page
Materials ProjectOQMDOMDBOPTIMADE
Every one of them is public and free. We tell you which number came from which, and when we took it — retrieved 2026-08-19. The full list, with licences and what each one cannot tell you
You need a property for a grade you have not used before. Twenty minutes later you have four tabs open: a supplier PDF, a MatWeb entry, a forum thread, and a chat window that gave you a figure to three significant digits with no source at all.
You will probably pick the right material. The question that arrives six months later is not which material — it is where did this number come from, and by then the tabs are closed.
Three steps, and the third one is the point: the tool stops exactly where professional responsibility starts.
A question in plain language, the way you would ask a colleague. No query syntax, no schema, no formula lookup first.
The agent screens open databases, collects candidates, traces every value to an entry, and drops what it cannot attribute.
You get a Material Decision Record: constraints, candidates, rejections with reasons, values with levels, and what is missing. The signature is still yours.
The middle figure is deliberately not a number. Run time has not been measured yet, and a page that grades other people's numbers does not get to publish an unmeasured one of its own. That is what E0 means.
Not a bibliography at the bottom. The database, the entry identifier and the method, beside the value, where you can click it. If we cannot attribute a number, we do not show it.
That rule is what makes the gaps visible too. A property we could not source is written as
not established, with the reason — a paywalled standard, an entry with no stated method —
instead of being quietly left out. An empty cell is a bug. A named gap is something you can act
on.
A specification minimum from a standard, a DFT result from an open database, and a machine-learned prediction are three different kinds of claim. They are labelled as three different kinds of claim.
We publish the grading scale we use as an open specification, so you can apply it to us and to everything else you use. Six levels, from a value a standard specifies down to a value nothing supports.
A material decision record you can put into your design documentation: what you asked, what was considered, what was rejected and why, what each value rests on, and what the analysis does not cover. Written to be read by somebody who was not there.
That last part is the test. A chat log answers the person who was in the room. A record answers the reviewer who arrives afterwards, and that is the person the question actually comes from.
A tool that rules is a liability. Your name goes on the drawing, and the question that arrives six months later arrives to you. So the agent does not choose — it screens, traces, and writes down what it rejected and why. What you get is the material for a decision that survives review, in a form you can hand to someone who was not there.
Where a value below the top of the scale is used as a design allowable, the record writes that down as your acceptance, with the date. Not ours. That is the difference between a tool taking responsibility it cannot carry and one that hands you the evidence and gets out of the way.
Materials Project, OQMD, OMDB and other providers reachable through OPTIMADE — roughly 27.3 million structures across 45 aggregated sub-databases, all public and free. The full list, with licences and what each one cannot tell you.
We have no proprietary database and we are not trying to build one. The public projects are better funded than we are and have a fifteen-year head start. Two of the four state no licence at all, and that is on the page as well.
Our coverage is strong for inorganic crystalline materials and weak for anything formulated. Polymers, elastomers, composites, concretes and coatings are outside our data layer. Commercial grade designations bridge incompletely to composition-indexed databases, so if you need a specification minimum for a named grade we will tell you that we do not have it and point you at the standard.
This is the section a competitor would not write, and it is the one that saves you the most time. An engineer who finds the limits after paying is a refund and a bad review.
Every capsule is one entry from the knowledge base: the vertical extent is its published strength range and the colour is its material class. It plots 37 entries and shows 25 of them — the rest do not publish both a density and a strength, so they are left out rather than estimated. The empty regions are the coverage gap described above, drawn rather than asserted.
The conformance framework we publish is one we currently fail. The failures are listed, with dates and with what each one would take to fix, in our own assessment.
If that reads oddly on a homepage: it is easier to check than to claim. A checklist with a built-in win is detected in three minutes, and after that nothing else on the page is believed either.
The notes below are tests with their methodology, their raw results, and a closing section saying what each test does not establish.
230 / 250 MPa
the same sheet, measured in two directions
2026-08-20
82 gaps named
properties across 27 grades we could not source at all
2026-08-20
1971
the year the AI-designed compound was first reported
2026-08-19
Engineers who have to justify a choice. You need a candidate, the numbers behind it, and a document that survives the question “why this one”.
Researchers and lecturers. Faster literature orientation, sources you can cite, and an academic plan priced so a cohort can use it.
Consultants and small design offices. An hour recovered is an hour billed, and the output is an artefact you can hand to a client.
A general model gives you a number and no way to check it. Here the database, the entry and the method sit beside the value, and the value is labelled as specified, measured, computed or predicted. Ask both the same property question and compare what comes back — that comparison is the entire argument, and it takes two minutes.
You should not, and you are not asked to. The tool does not rule on anything: it assembles evidence, grades it, and leaves the decision with you. A value with no resolvable source is graded E0, and E0 never satisfies a constraint.
Stay with it for the selection itself. This is faster in the phase before you know what you are looking for, and what comes out is a record with working links that you can send to somebody else.
Both, unevenly, and this is the sharpest limit we have. The open computational databases are indexed by composition and structure, not by grade designations. Grade pages are built from datasheets and standards instead, and where the specification minimum sits behind a paywall the page says so and points you at the document rather than guessing.
Yes. Your question goes to a hosted model. If your project data cannot leave your network, this is not the right tool for you today — that is a straight answer rather than a subscription you would not be able to use.
That is what the record is for. Read its limitations section first: it lists what the analysis does not cover, and that section is part of the record rather than a footnote to it.
Every value carries the document it came from, so an error is traceable to a source rather than to a black box. That does not make a wrong number harmless. It makes it findable, which is the difference between a correction and a recall.
It is free, and we read it. What you pay for is not having to write the queries, and a record with the sources attached that you can paste into a document with a list of references.
The free plan gives you three complete reports and the whole knowledge base, with no card. If the sources do not hold up, you will know in the first week.
