← analyzer · failure modes · recoater strike

Why a metal 3D print fails halfway through the build

The most common way a laser powder bed fusion (LPBF) build dies is not a laser problem or a powder problem. It is mechanical: part of the geometry rises above the powder bed and the recoater hits it. The blade or roller strikes the raised material, the build stops — and if the blade chips, it drags defects through every other part on the plate for the remaining layers.

It usually happens hours in, on a plate that looked fine at layer 200. That is why it is expensive: a mid-size metal build is roughly $10,000–$20,000 of machine time, gas, labour and powder, and a peer-reviewed cost study puts build-failure risk at 26–38% of LPBF unit cost (Baumers et al.) — an independent number, unlike the widely-quoted “70% of metal AM failures are distortion,” which has no primary source and should not be cited.

What actually happens

Each melted layer solidifies and contracts while the material underneath it is already solid. The new layer cannot shrink freely, so it pulls — and the part accumulates tensile stress near its top surface, layer after layer. That stress has to go somewhere:

It lifts a corner off the plate if adhesion or the support anchors give way first. The corner rises into the recoater’s path.
It curls an unsupported edge upward where new material is being melted over loose powder rather than over solid metal. Powder is a poor heat sink; the overhanging edge runs hotter, contracts harder, and rolls up.
It bends a tall thin feature sideways — and a thin wall does not need to lift far to be in the way.

All three end at the same place. The failure people describe as “it pulled off the supports” and the failure they describe as “the blade crashed” are generally the same event, seen from opposite ends.

What the geometry tells you before you print

Recoater strike is one of the four failure modes that is predictable from geometry alone, because every one of its precursors is a measurable shape property. Four gates carry it:

Gate	What it measures	Threshold and where it came from
`overhang`	Per-face downskin angle from the build plate, area-weighted, with plate-contact faces excluded	Two tiers, not one line. Below 20° unsupported faces genuinely fail; at or above 30° no support is needed. The stock 45° rule rejects a band six independent studies printed unsupported — IN718 at 25° (Herzog 2022), AlSi10Mg built at 30° and failed at 15° (Calignano 2022, Metals 12:2083), CoCr built at 30° and failed at 20°, NiTi at 24°. The measured failure boundary is 15–20°.
`slenderness`	Height over thickness of each free-standing thin feature	Watch at h/t 40, fail at 75. The previous 8/12 rule of thumb was wrong by 4–8×: Herzog 2022 (J. Laser Appl. 34:012015) built a 0.3 mm × 15 mm wall (h/t 50) with no deviation from design, while 0.3 × 30 mm (h/t 100) failed to reach full height and 0.4 mm bars (h/t 75) bent.
`section_jump`	Rate of cross-sectional area change per millimetre of Z	A sharp area increase going up is new material starting over powder — the unsupported shoulder, and the classic corner-lift site. Reported with its Z height so you can find it on the model.
`support_burden`	Support volume, plate-versus-part landing, and drop height, from geometry alone	Banded against 100 real parts. What it does not judge is support design — strut sizing, perforation, removal access.

Two further gates — bottom_warp_screen and heat_accumulation — localize where the stress concentrates, as self-normalized indices rather than physical units.

The honest limit on warp. This screen localizes distortion risk. It emits no displacement numbers, and the reason is published: the quantitative inherent-strain solve behind it was measured against both NIST AM-Bench benchmark bridges and under-predicts deflection by roughly 100×. It is useful for ranking and location, never for absolute millimetres, and the report says so in place. Anyone quoting you sub-millimetre warp predictions from an STL alone should be asked what they validated against.

What it cannot see

Neighbouring parts and plate position. The screen receives one part in isolation. Spatter-induced defects are experimentally established (ORNL) to depend on what else is on the plate, so they are structurally out of reach here.
Process excursions. NASA’s combustion chamber that failed after nine seconds passed inspection; it died from witness-line fractures at build restarts, “none of which were detected by the machine logs alone.” Screening owns design-stage risk. Excursions belong to in-situ monitoring.
Recoater compliance and recoat direction. A soft rubber blade tolerates what a hard steel one will not, and a wall’s angle to the recoat direction matters. Neither is in the STL; both ride in the finding text instead of the number.

Scope, stated once. This is a necessary-condition checker, not a build-outcome predictor. Geometry can say a part is at high risk. It cannot say a build will succeed — that also depends on powder, machine, layout, parameters and the operator. Of the 18 LPBF failure modes in the tool’s published taxonomy, 4 are fully geometry-predictable and 12 are informed by at least one of the 20 gates; the rest are out of reach of geometry and are listed as such in every report.

Check a part

Every gate named on this page runs automatically on an uploaded STL or STEP at amprintanalyzer.com — free, no sign-up, and the geometry is never modified. Findings are pinned to the model and each one is labeled by its evidence class: measured (a direct geometric fact), screening (a dimensionless index that localizes risk), or estimate (a planning number with its assumptions stated).

If the screen disagrees with something you have actually measured on a printed part, that is the useful case — say so through the feedback strip. Disagreements are the calibration data.