Last verified 19 August 2026. This page is the overview; each step links to its own page.
1. Translate — turn the requirement into something you can screen against
Function, constraints, objective, free variables. A constraint has a property, an operator, a value, a unit and a condition.
Where it goes wrong: adjectives. “Good corrosion resistance”, “reasonably light”, “cost-effective”. None of them can be screened against, so in practice they are waved through and reappear at the end as an argument nobody can settle. If you cannot write it as a comparison, it is not a constraint — it is a qualitative requirement, which is a real thing that belongs in the record, recorded as itself rather than converted into a score.
The test: could a colleague, given only your constraint list, run the same screen and get the same survivors? If not, the translation is not finished.
2. Retrieve — assemble candidates, and name where each came from
Standards, handbooks, mill datasheets, aggregators, computational databases, suppliers, colleagues, and increasingly a model. All legitimate. All different.
Where it goes wrong: the shortlist arrives without its provenance. Six months later the question is not which material you chose but where a particular number came from, and by then the tabs are closed.
The habit: record the source and the date at the moment you read it. Databases are revised, datasheets are reissued, standards get new editions. A value with no retrieval date cannot be reproduced, because nobody can tell which version you saw.
Also record what you rejected. A shortlist without its rejections is a conclusion, not an analysis, and “why didn’t you use the obvious alternative” is the first question a reviewer asks.
→ Where open material data comes from, and its limits
3. Assess — decide how far each number can be defended
This is the step that did not exist thirty years ago, because there was nothing to assess: the data came from one place and that place was the answer.
Now a single table can mix a specification minimum, a datasheet typical value, a DFT result and a model output, rendered identically. They are not equally defensible and nothing in the presentation says so.
The distinction that costs the most: a specified minimum versus a typical value. One is a floor the standard guarantees; the other is what a supplier’s production usually does. Design to the second where the first belongs, and your margin is not the margin you think it is.
The distinction that is easiest to miss: a number with no resolvable source. Under the E-ladder that is E0, and it does not belong in documentation whatever the interface looked like when you found it.
→ Six evidence levels · how far to trust computed data
4. Compare — screen, then rank
Screening is a gate: a candidate clears a constraint or it does not. Ranking orders the survivors on a single objective, usually through a material index that separates the material term from the geometry so you can rank before the design is fixed.
Where it goes wrong, twice:
Precision beyond the inputs. If two candidates differ by less than the width their sources carry, they are not separated, and presenting an order is a fabrication performed by arithmetic. Record the tie.
Derived values that outrank their inputs. An index computed from a certified modulus and a predicted density is a predicted index. Arithmetic does not create evidence — it propagates the weakest link.
→ Material indices and where the exponents come from
5. Export — write the record while you still know the answers
Ashby’s fourth step is documentation. In practice it is the step most often intended and least often completed, because by the time anyone writes it up the cheap information has evaporated.
A record that survives contains seven things: the requirement as constraints, what was considered including rejections, where each number came from, how strong each number is, the date, who decided, and what the analysis did not examine.
Fifteen minutes, at the moment of deciding. Not an afternoon, later, from memory.
→ How to document a material selection decision · build one in the browser
The shape of the whole thing
| Step | What it produces | The usual failure |
|---|---|---|
| Translate | testable constraints | adjectives that were never screened |
| Retrieve | candidates with named sources | a shortlist with no provenance |
| Assess | an evidence level per value | a minimum and a typical value in one column |
| Compare | screened, ranked survivors | a ranking more precise than its data |
| Export | an auditable record | written later, from memory, or not at all |
Nothing here is difficult. It is all cheap at the moment of doing and expensive afterwards, which is exactly the profile of work that does not get done.
What this page does not cover
- Process selection. Which manufacturing route suits the material is a parallel problem with its own method.
- Design codes. ASME, EN 13445, Eurocode and their equivalents impose requirements that govern over anything here.
- Qualification and testing. Selecting a material is not qualifying it.
- Cost modelling beyond substituting cost into an index.
- Sustainability and end-of-life, which increasingly belong in the constraint list and are not treated here.