The datum reference frame: how three surfaces decide where your part is

2026-08-24 · long read · arithmetic only, worked step by step

Every position, perpendicularity and profile callout on a drawing is measured from somewhere. That somewhere is the datum reference frame — a coordinate system of three theoretically perfect, mutually perpendicular planes meeting at a common origin, established from three real, imperfect surfaces on the part in front of you. This note shows what the frame is, how it is constructed from measured points, and why changing nothing but the order of the letters in a feature control frame can turn a passing hole into a failing one. There is a complete worked example, and every number in it can be checked with a pocket calculator.

1 · What a datum reference frame actually is

A drawing says a hole must sit within Ø0.25 of where it belongs. Where it belongs is a location — and a location needs an origin and three directions to be measured from. On a machined part there is no origin painted on the material. So geometric dimensioning and tolerancing does something quietly clever: it builds the coordinate system out of the part itself, and that frame is what locates the tolerance zones for every geometric control on the drawing.

A datum reference frame (DRF) is three mutually perpendicular theoretical exact planes that intersect at a common origin, together with the three axes formed where those planes meet. That is the whole definition. It is a perfect grid — perfectly flat planes, perfectly perpendicular to one another — derived from surfaces that are none of those things.

Two ideas do all the work here, and conflating them is the single most common source of confusion:

• The datum features are the real, tangible physical surfaces on the manufactured part: a face, a bore, a shoulder. They have form error. You can touch them.
• The datums are the theoretically perfect version derived from those surfaces: a perfect datum plane, a perfect datum axis, a perfect reference point. You cannot touch them; they are geometry, not material.

The datum reference frame is the reference system assembled from those perfect datums. Take it away and a position tolerance means nothing — you would be measuring from wherever the part happened to be lying on the table. So the DRF matters far more than its share of drawing ink suggests: it is the thing that makes the other numbers mean something, and if two people build it differently they will honestly report two different results from one part.

2 · Six degrees of freedom, and what constraining one means

Put a part on a bench. It can slide three ways and turn three ways: translation along X, Y and Z, and rotation about X, Y and Z. Six degrees of freedom. To measure the part against a fixed reference system you must remove all six — the part has to have exactly one place it can be.

Nothing is being clamped here. Constraining a degree of freedom means the datum feature determines it: given the surface, only one answer is left for that particular slide or turn. Whether you achieve that with a physical fixture or with arithmetic inside inspection software makes no difference to the result.

The order in which the datums are referenced decides which surface gets to determine what. The primary datum goes first and takes the freedoms it can; the secondary datum takes what is left that it can control; the tertiary datum takes whatever remains. All three are chosen to satisfy the part's functional requirements in assembly or use. Each is constrained by the ones ahead of it, never the other way round. That hierarchy is the entire reason datum order changes answers, and section 8 puts a number on it.

3 · The 3-2-1 rule

For the most common case — three nominally perpendicular flat faces — the accounting has a nickname. The 3-2-1 rule is shop language for how three planar datum features between them arrest six degrees of freedom, and it also describes how you would build a fixture to do the same thing physically. It is a locating convention rather than a clause you will find worded that way in the standard, but the freedom count behind it is exact.

Why three points first? A flat plate resting on two points can still rock. Three non-collinear points of contact settle it: the part can no longer tip about either horizontal axis, and it can no longer sink into the plate. One translation and two rotational degrees gone in a single move — which is why the primary datum feature is the most influential surface on the drawing and should be the largest, flattest, most functionally important of the three.

The part can still slide two ways and spin about the vertical axis. Push it against a second face at two points and the spin goes along with one translation. Touch a third face at one point and the last slide is gone: the part has exactly one position in space.

One caution about the nickname: the standard does not establish a datum plane from literally three probed points. It establishes it from contact with the whole datum feature. The three points are the minimum number that pins the plane down, and they are how a fixture works, but a real evaluation uses everything you measured on that surface.

4 · Datum feature, datum, datum feature simulator

A third term completes the picture, and it explains how perfect geometry gets extracted from an imperfect surface. A datum feature simulator is the boundary that contacts the datum feature. In a fixture it is a physical thing — a surface plate, an angle plate, a precision pin, a chuck. In software it is the same idea computed rather than machined: the theoretically perfect plane laid against the measured points. Either way it touches the high points of the real surface, and the datum is taken from the simulator, not from the average of the material.

So the chain runs: datum feature (real surface) → datum feature simulator (perfect boundary in contact with it) → datum (the theoretical exact plane, axis or point) → datum reference frame (three of those, mutually perpendicular, with an origin).

One practical consequence catches people out: form tolerances applied to a datum feature — flatness on A, for instance — still apply, and they limit how far the real surface may depart from its own perfect counterpart. A datum feature with sloppy form is a wobbly foundation for everything referenced to it, and no amount of care downstream repairs that.

5 · Datum feature symbols and the feature control frame

Two pieces of drawing notation carry all of this.

The datum feature symbol identifies which physical surface is which. It is a capital letter in a square box, connected by a leader to a triangle that sits on the feature. Where the triangle lands decides what kind of datum you get, and this is the part worth being deliberate about:

• On a surface — the triangle touches the surface outline or an extension line clear of a dimension line. The datum is a plane.
• On a feature of size — the triangle aligns with the dimension line arrow of a diameter or width. The datum is the datum axis of a bore or pin, or the centre plane of a slot. Yes, a centreline can absolutely be a datum: that is exactly what this placement means.

Letters are assigned in any order that suits the drawing; I, O and Q are skipped because they are too easily confused with the numerals 1 and 0. Datum letters are not required to appear in alphabetical order in a feature control frame, and assuming they do is a reliable way to get the wrong answer.

The feature control frame is the box that states the geometric control, and its compartments read left to right: the geometric characteristic symbol, then the tolerance zone (its shape, size and any material condition modifier), then the datum references in order of precedence, which establish how the other features on the drawing are located and oriented relative to the frame. A frame reading position, Ø0.25, A, B, C means: primary A, secondary B, tertiary C. The same three letters written A, C, B is a different instruction and a different frame — and that stated datum structure is what decides whether the controlled feature will pass inspection.

6 · Worked example, part one: building the frame

Here is a complete construction. It is a hypothetical part with numbers chosen so that every step comes out clean; nothing in it is a real customer job. Take a calculator and follow along.

The part. A machined rail, nominally 300.000 × 60.000 × 20.000 mm.

• Datum feature A: the top face, 300 × 60. Primary.
• Datum feature B: the long side face, 300 × 20. Secondary.
• Datum feature C: the end face, 60 × 20. Tertiary.
• The feature under test: a Ø10 through hole, basic dimensions 240.000 in X and 30.000 in Y from the datum origin, controlled by position Ø0.25 to A, B, C at regardless-of-feature-size.

The measurement. The rail is scanned sitting on a table, not aligned to the machine's X and Y. The scanner reports everything in its own machine coordinate system. All figures are millimetres.

These are representative points standing in for the fitted results of thousands of scanned points on each surface. Datum feature A comes out level in the machine frame because the rail is lying on the table and the scanner was set up square to it — an idealisation chosen so the rotation you have to follow is a single one, in the horizontal plane. Nothing about the recipe changes when A is tilted; the dot products simply carry three non-zero components instead of two.

Step 1 — the primary datum: a plane, and the Z direction

Two vectors along the surface give its normal by cross product.

Here A₂ − A₁ = (90.000, 30.000, 0.000) and A₃ − A₁ = (30.000, 70.000, 0.000). The cross product is (30·0 − 0·70, 0·30 − 90·0, 90·70 − 30·30) = (0, 0, 5400). Dividing by 5400 gives the unit normal n_A = (0, 0, 1), and the datum A plane is z = 40.000.

That normal becomes the Z axis of the frame. Three degrees of freedom are now settled: the part cannot move along Z, and it cannot rotate about X or Y.

Step 2 — the secondary datum: a direction, made perpendicular to A

The B face gives a direction. Subtract its two points:

B2 − B1 = (230.4000, 67.2000, 0.0000)

Its length is √(230.4² + 67.2²) = √(53084.16 + 4515.84) = √57600 = 240.000 exactly. Divide through and the unit direction is u_B = (0.9600, 0.2800, 0.0000).

Now the critical move, and the numerical meaning of datum precedence. The secondary datum is not allowed to disturb the primary. Whatever component of u_B points along n_A gets stripped out before the direction is used:

In this case u_B · n_A = 0.96(0) + 0.28(0) + 0(1) = 0, so nothing is removed and X = (0.9600, 0.2800, 0.0000). The third axis follows from the other two:

Y = n_A × X = (0, 0, 1) × (0.96, 0.28, 0) = (−0.2800, 0.9600, 0.0000)

Check it: X·Y = 0.96(−0.28) + 0.28(0.96) = −0.2688 + 0.2688 = 0. Perpendicular, as required. Two more freedoms gone — the spin about Z, and one slide.

Step 3 — the tertiary datum: the last slide

C contributes position only. Its datum plane is constructed perpendicular to both A and B and pushed against the C face until it touches. The part's material lies to +X along the frame's own X direction, so the simulator stops at whichever point on that face sits lowest along that axis — step 4 works it out as C1, at X = 0, against C2 at X = 0.030. The origin of the frame is then the point lying on all three datum planes at once: machine point (150.0000, 100.0000, 40.0000).

Verify that it really is on all three. It sits at z = 40.000, so it is on plane A. Its offset from B1 is (−19.2000, −5.6000, 10.0000), and the Y component of that is (−19.2)(−0.28) + (−5.6)(0.96) + 0 = 5.376 − 5.376 = 0, so it is on plane B. Its offset from C1 is (0, 0, 10.0000), whose X component is zero, so it is on plane C.

Step 4 — converting any measured point into the frame

Run the two C points through it as a sanity check. C1 minus the origin is (0, 0, −10.0000), giving frame coordinates (0.0000, 0.0000, −10.0000). C2 minus the origin is (−16.7712, 57.6084, −10.0000), giving

X = (−16.7712)(0.96) + (57.6084)(0.28) = −16.100352 + 16.130352 = 0.0300
Y = (−16.7712)(−0.28) + (57.6084)(0.96) = 4.695936 + 55.304064 = 60.0000

So the end face runs 60.000 mm across and drifts 0.030 mm sideways as it goes. That is a squareness error of 0.030 mm between datum features C and B. Hold on to it — section 8 is about what it costs.

7 · Worked example, part two: evaluating the hole

The hole axis meets the A plane at machine point (372.0300, 196.0400, 40.0000). Take the axis to be perpendicular to datum A, so this single point describes it — on a hole that leans, the worst point along the axis governs. Subtract the origin:

p − O = (222.0300, 96.0400, 0.0000)

Then three dot products:

X = 222.0300(0.96) + 96.0400(0.28) = 213.1488 + 26.8912 = 240.0400
Y = 222.0300(−0.28) + 96.0400(0.96) = −62.1684 + 92.1984 = 30.0300
Z = 0.0000

The hole sits at (240.0400, 30.0300) against basic dimensions of (240.000, 30.000). The deviations are ΔX = +0.0400 and ΔY = +0.0300.

√(0.0400² + 0.0300²) = √(0.0016 + 0.0009) = √0.0025 = 0.0500. Doubled, the position is Ø0.1000 mm against a tolerance of Ø0.25 — 40 % of the allowance used. The hole passes inspection comfortably.

8 · Datum order: the same data, a different answer

Now change one thing. Not the part, not the scan, not a single measured number — only the drawing, which now reads position Ø0.25 to A, C, B — the same three datums, arranged in a different precedence order. Datum feature C is promoted to secondary and B demoted to the last datum.

The consequences follow mechanically from equation 2. A is still primary, so the Z axis is unchanged. But now it is C whose direction is orthogonalised against A and becomes an axis of the frame, and B that is reduced to supplying a translation. The end face's 0.030 mm of drift over 60 mm, which the first frame ignored entirely, is now the thing that sets the frame's rotation.

The angle is small and easy to compute:

tan φ = 0.030 / 60.000 = 0.0005 → φ = 0.02865°

Rotating the axes by φ and re-expressing the hole gives (the cosine of such a small angle is 0.999999875, and its sine 0.0005):

X′ = 240.0400(0.999999875) − 30.0300(0.0005) = 240.039970 − 0.015015 = 240.024955
Y′ = 240.0400(0.0005) + 30.0300(0.999999875) = 0.120020 + 30.029996 = 30.150016

The origin does not move, and it is worth seeing why. Datum feature C now contacts its simulator along its whole face, so the frame's X = 0 plane still passes through the same corner. Datum feature B is demoted to supplying one translation, and the rotation has tipped it away from the frame — across the face's full 300 mm run from the C plane, its Y′ climbs from 0 at the corner to 0.150 at the far end. The simulator touches the lowest point, which is that same corner. (The two tabulated B points sit at X = 20 and X = 260, well inside the face; they stand in for the fitted line, not for its ends.) Same origin, pure rotation.

If you would rather skip the trigonometry, the whole effect is visible in one multiplication — the hole is 240 mm from the origin, and swinging it through 0.0005 radians moves it 240 × 0.0005 = 0.120 mm sideways. Added to the 0.030 it already had, that is 0.150.

Now the deviations are ΔX = +0.024955 and ΔY = +0.150016:

√(0.024955² + 0.150016²) = √(0.00062275 + 0.02250480) = √0.02312755 = 0.152077
position = 2 × 0.152077 = Ø0.3042 mm

Three times the position error, and a pass turned into a fail, from a drawing edit. Nobody measured anything wrong. Both numbers are correct answers to different questions.

The engineering lesson underneath the arithmetic is about lever arms. Datum feature C is only 60 mm long, and it is being asked to control the rotation of a 300 mm part. A small angular error over a short feature becomes a large positional error a long way away — 0.030 mm at the end face turned into 0.120 mm at the hole, a fourfold amplification, purely because the hole is four times further out than the datum feature is long. Datum feature B is 300 mm long and controls the same rotation five times better.

9 · When the datum features are not flat faces

Planes are the easy case. Real parts bring three complications, and the accounting stays the same in all of them, whatever the datum structure: six freedoms, removed in order of precedence.

Features of size. Point the datum feature symbol at a diameter and the datum becomes a datum axis, taken from the perfect cylinder that contacts the bore or pin. A cylinder as primary removes four freedoms at once — two translations and two rotations — leaving only the spin about that axis and the slide along it, so the count becomes 4-1-1. A shaft-like part is often A (the bore axis), B (a face for axial position), C (a keyway or hole for clock); different datums can come from pins, slots, holes or faces, depending on how the part is functionally located. This is also where datum shift enters: reference a feature of size at a material boundary modifier rather than regardless of feature size, and the part may move within its simulator by the difference between the actual and boundary size — legitimate, but it must be intended.

Inclined datum features. A face at a basic angle still yields a datum plane; the frame's planes stay mutually perpendicular and the inclined feature's simulator is held at the basic angle to them. Equation 2 is doing exactly this work — a measured direction never becomes an axis unmodified, it is always brought into line with what the higher-precedence datums already decided.

Datum targets. When a surface is too rough, wavy or large to trust as a whole — a casting skin, a forging, a weldment, a panel that flexes — the drawing can nominate specific points, lines or areas of contact instead. Datum targets are drawn as circles split by a horizontal line, located by basic dimensions, and a classic pattern is the 3-2-1 arrangement made explicit: A1, A2, A3 on the primary surface, B1 and B2 on the secondary, C1 on the tertiary. The point is repeatability: two inspections that touch the same nominated spots build the same frame, where two that use "the whole surface" of a rough casting will not. That is why cast and forged parts can be measured to consistent numbers at all — a subject running throughout our note on structured light in the forge and foundry.

10 · Which fit rule builds the datum

Everything above assumed the datum plane and the datum direction come out of the measured points in a defined way. They do — but the definition in the standards and the default in the software are frequently not the same, and that is a quiet source of disagreement between two labs measuring one part.

Both ASME Y14.5 and ISO 5459 establish datums by contact: the simulator is a perfect surface progressed into the datum feature until it touches, with all the material on one side, and ISO 5459 is explicit that the association is a minimax fit constrained to lie outside the material. Many CMM and scanning packages will nonetheless build an alignment from a least-squares plane, which threads through the middle of the surface instead of resting on its high spots.

On a good surface the two agree to within a fraction of the form error and nobody notices. On a bowed or wavy datum feature they do not — and a tilt on the primary datum propagates into every position measured from the frame, amplified by the same lever arm that made section 8 dramatic. The difference between the two rules, worked from first principles with numbers you can check, is the subject of our note on Gaussian versus Chebyshev fitting.

A second wrinkle: a bowed or convex datum feature can rock on its simulator, so more than one contacting plane is valid. Both standards have machinery for resolving that — candidate datum sets in ASME Y14.5.1, constrained minimax in ISO — and most software has a setting for it. The instruction is the same either way: find out what your software does, write it into the inspection method, and make sure your customer's lab does the same. A disagreement about the alignment rule looks exactly like a disagreement about the part.

The frame is not free of measurement error either. Every datum feature is sampled, fitted and reported with its own uncertainty, which flows into every result the frame produces — see measurement uncertainty for how that budget is built. And because the alignment is re-established each time the part is loaded, its repeatability is a real component of what a gage R&R study attributes to your process: if two operators seat the same casting differently against the same datum features, the frame moves, and so does every number on the report.

11 · Common questions, and the short version

What is the difference between a datum and a datum feature?

The datum feature is the real surface on the manufactured part. The datum is the theoretically perfect plane, axis or point derived from it. You inspect the datum feature; you measure from the datum.

What are the three types of datum?

By geometry: datum planes, datum axes and datum points — with a centre plane as the plane form derived from a feature of size. By role in a feature control frame: primary, secondary and tertiary, which is about precedence rather than shape.

Does datum order have to be alphabetical?

No. Precedence is the left-to-right order in the feature control frame and nothing else. A, C, B is a perfectly valid and materially different callout from A, B, C, as section 8 shows in numbers.

Can a centreline be a datum?

Yes — attach the datum feature symbol to the dimension line of a feature of size and you get a datum axis or centre plane. What you may not do is attach it to a centreline drawn between two unrelated surfaces and expect it to mean anything; the datum has to come from a tangible feature.

How many datums do I need?

As many as the control needs, and no more. Not every geometric control requires a full three-plane frame: flatness needs none, perpendicularity needs one, and a position in a single direction may only need two. Referencing a third datum you do not need adds an over-constraint that can fail parts for no functional reason.

Further reading

• ASME Y14.5 — dimensioning and tolerancing: datum reference frames, degrees of freedom, datum feature simulators, datum targets.
• ASME Y14.5.1 — mathematical definition of dimensioning and tolerancing principles.
• ISO 5459 — geometrical product specifications: datums and datum systems.
• ISO 1101 — geometrical tolerancing: form, orientation, location and run-out.
• ISO 8015 — fundamentals, concepts, principles and rules of GPS.
• JCGM 100 (GUM) — evaluation of measurement data, for the uncertainty that rides along with every alignment.

Related notes: Gaussian vs Chebyshev — the fitting rules that decide what a datum plane actually is — and Gage R&R from first principles — how much of the variation you see is the part rather than the setup. Worked demonstrations live in the examples.

If your alignments are rebuilt by hand on every part, and nobody can say which fit rule produced them, that is a solvable problem.