The science of
the cutting room.

The standard factory estimate treats spreading speed as a flat rate — metres per minute, nothing more. It isn't. A spreading machine accelerates from rest, reaches cruise speed, then decelerates to stop at the marker end. For short markers, the machine never reaches cruise speed at all — it is still decelerating before it gets there. This means short-marker throughput is systematically overestimated by any model that ignores the ramp. The physics is a trapezoid, not a rectangle.

There is one published academic paper on spreading time prediction. It uses regression on lay length and ply count — no machine physics. The motion model behind a real spreader has never appeared in the literature. A physics-based implementation of this model, calibrated against real production data, achieves a prediction error small enough to replace manual estimates in production planning.

Every fabric has a preferred spreading mode. One-way (1W): the machine lays in one direction and returns empty. Zigzag (ZZ): it lays continuously in both directions without returning. The per-ply cycle roughly halves in ZZ mode — but the job-level throughput gain is 35–40%, not 2×. Roll changes, setup, and loading time are mode-independent and dilute the saving. The actual figure depends on lay length, average roll length, and roll loading time. Without a calculator, it is a guess.

ZZ also generates higher end loss, not lower. The fabric must fold at both ends, and the machine releases additional material at each turn to maintain the fold while reversing. ZZ trades fabric for throughput. This trade-off has never been formally modelled in the academic literature. It is known qualitatively in industry and ignored quantitatively everywhere.

Before a single ply is spread, someone must decide: which size combinations go on each marker, how many plies to spread per lay, which fabric rolls to assign to which jobs, and how many cutting sections to open. Together these decisions form the cut order plan. The problem is combinatorially hard — formally equivalent to a bin-packing variant with demand constraints, table length limits, and knife height limits. Confirmed NP-hard.

It has been studied academically for over 20 years. The most recent work (2026) applies Particle Swarm Optimization to 32 real factory cases and reports 1.25% cost savings against genetic algorithm baselines at 98.6% less computation time. Most factories still plan manually, with a spreadsheet and a lot of experience.

Nesting — placing pattern pieces onto a fabric marker to minimise waste — is among the oldest applied combinatorics problems in manufacturing. The foundation is the No-Fit Polygon (NFP): a geometric construction that transforms the question "do these two pieces overlap?" from a polygon intersection test into a point-in-polygon test, orders of magnitude faster. Every commercial nesting system built since 2000 uses NFP-based placement combined with a metaheuristic sequencing algorithm.

On plain wovens without matching constraints, good automation achieves 85–92% marker efficiency — competitive with skilled human markers. The exception is pattern-matched fabrics: stripes, plaids, checks. Matching introduces a phase-alignment constraint across every seam pair. The algorithm must simultaneously satisfy non-overlap, grain tolerance, and repeat-phase alignment simultaneously. The computation cost is high and the efficiency penalty is real. Vision systems have narrowed this gap significantly. For simple stripe matching, algorithms are competitive with skilled human markers today. For complex plaids — where both warp and weft repeats must align across multiple seam pairs — the human expert's ability to trade off less visible seams still produces better outcomes. The frontier is moving.

After the marker is approved, the CNC cutter must determine the order in which to cut each piece and the path the cutting head will follow. The problem is formally a Generalized Travelling Salesman Problem (GTSP): minimize total non-cutting travel while visiting every pattern contour exactly once. NP-hard. Commercial systems solve it with genetic algorithms and achieve 15–30% reductions in non-productive travel versus naive sequencing.

The cutting head also controls blade angle continuously — the knife must remain tangential to the cut path at all times. At high speeds and tight curves, the required angular velocity can exceed the rotation actuator's physical limit, causing positional lag and notch errors. High-end systems pre-compute curvature and pre-rotate the blade before the head reaches the curve. This layer of the process is entirely inside the machine's embedded software. The factory operator does not see it.

Fabric waste in a cutting room has two distinct causes that require two distinct responses. Preventable waste — end loss, end bits, splice overlaps — is owned by the planner and the operator. It is reducible through better roll assignment, tighter machine precision, and process discipline. Non-preventable waste — defective fabric removed from the spread — is owned by the supplier. No amount of better planning recovers it.

Most factories see one number: total consumption against marker yardage. The split between preventable and non-preventable, and between the planned allowance and actual floor practice, is invisible in their reporting. Operators pad end-loss allowances. Splices are made manually with generous overlaps because it is faster than precision. The over-allowance is buried inside total consumption, and most managers look away to avoid owning the problem. The academic literature has not formally modelled this variance. The gap exists, is measurable, and in most cutting rooms goes unmeasured.

Three applications of AI have demonstrated consistent results in cutting room production: genetic algorithms and simulated annealing for cut order planning and scheduling (20+ years of validated research), computer vision for fabric defect detection (mature commercial products, 99%+ accuracy on trained defect types on solid-color fabrics), and machine learning for demand forecasting upstream of the cutting room.

Everything else is either early-stage research or marketing. Reinforcement learning for real-time scheduling shows genuine promise in simulation; in real factories, sparse reward signals, unplanned machine breakdowns, and sudden order changes have prevented reliable deployment. The gap between what the research literature claims and what is actually running in production is wide — and rarely acknowledged by the vendors selling "AI-powered" cutting rooms.

Tracking cut pieces in most factories starts after cutting, not before it. Labels — barcodes, RFID tags, piece identifiers — are applied to bundles after the cutter finishes, as a separate operation by a separate team. The reason is mechanical: a print-and-apply labeler cannot keep up with a spreader at production speed. The per-ply cycle of a standard applicator — trigger, print, piston extends, label applied, piston returns — takes several seconds. The spreader has already moved on. The problem compounds with marker complexity: more pieces per marker means more labels per ply, more cycles required before the carriage can advance.

The result is a structural decoupling. Spreading is a high-speed operation; labeling is not. Bundle labels carry a summary identifier — order number, size — but not the full digital history of the individual piece. The per-piece record exists in the planning system. It does not travel with the fabric.

Two engineering changes break this constraint. First: the pneumatic piston driving the applicator head is split into two shorter pistons in series. Given equal air pressure, two short strokes are faster than one long stroke — the physics of piston travel and air volume set the cycle ceiling, and halving the stroke length without halving the pressure resets it. Second: rather than waiting for the piston to complete its return before triggering carriage motion, sensors track the piston position in real time and initiate movement the moment the upward stroke begins. The carriage does not wait for the piston; the cycles overlap. Together, these reduce the label-to-label cycle to one second.

At one second per label, synchronised operation becomes viable. The labeler mounts on the spreader carriage and runs in lockstep with the spread. The overhead on spreading time is 20–30%, depending on marker complexity and table length. In return, the post-cut bundle labeling operation — with its separate team, separate pass, and accumulated positional errors — is eliminated entirely. Each piece carries the full digital record of its production: order number, material code, work order, fabric lot and roll, piece name, ply number. Not a bundle identifier that maps to a record elsewhere. Per-piece identity, applied at the moment the fabric is laid.

The paper marker is standard factory practice and has been since the 1960s. A CAD system produces a cut file; a plotter prints it at 1:1 scale on a paper roll; the roll is walked out over the spread and positioned by hand before cutting begins. On a 20-metre table, the marker is 20 metres of paper — consumed once per lay, per job, disposed of after cutting.

The paper marker works. It is readable under any lighting, requires no technology from the cutter beyond a blade, and carries six decades of proven use. It also carries a positioning error that is rarely measured: paper placed by hand on an uneven fabric surface can shift, stretch, and wrinkle. Cutting starts from wherever the paper lands, not from where the nesting algorithm placed the pieces.

Projection marking replaces the paper with a calibrated projector mounted above the table. The cut geometry is displayed directly onto the fabric surface. The cutter follows projected lines. When the cut file changes — a size substitution, a re-optimised nest, an order modification — the projection updates in seconds. No reprint. No re-lay. No paper. The coordinate system is referenced to the table, not to a physical sheet placed by hand, so the geometry does not drift with handling. Because the system communicates directly with the planning layer, it can display additional information — piece identifiers, cut sequence guidance, quality flags — that paper cannot carry.

The engineering constraints are real and rarely documented outside development environments. Projector calibration drifts under factory lighting. Lensing error increases at table edges. Depth-of-field variation across a fabric stack is non-trivial. These are solvable problems; maintaining the solution under three-shift factory conditions requires a calibration discipline that vendor literature does not describe. The paper marker is not disappearing because nobody thought of replacing it. It is disappearing slowly because the replacement is harder to maintain than it looks.