Cutting stock problems, nesting optimisation, bin packing, trim loss, raw-material waste, cutting patterns, column generation, branch-and-price, guillotine cutting, kerf, defects and usable leftovers all belong to one industrial question: how do we turn expensive stock material into the demanded parts while wasting as little as possible and still producing a plan the factory can actually cut? A paper mill slits giant rolls into customer widths. A steel service centre cuts coils and plate. A furniture factory nests panels on wood sheets. Glass, textiles, stone, aluminium, foam and composite manufacturers face related versions of the same combinatorial problem.
The deceptively simple objective “use less material” quickly becomes a deep cutting-stock optimisation problem. Parts have dimensions and quantities. Machines lose width to kerf. Some patterns require too many knife positions or rotations. Grain direction can forbid orientation. Defects remove usable regions. Sheets may require guillotine cuts. A leftover can be scrap today or valuable stock tomorrow. Setup time can make the mathematically lowest-waste plan operationally worse than a slightly wasteful plan using fewer distinct patterns. Current industrial reviews emphasise exactly these realities: raw-material reuse, defects, trim distance, demand uncertainty, pattern similarity and other plant constraints decide whether an optimisation model is useful outside a benchmark.
The mathematical breakthrough is often not a single clever packing heuristic but the way the problem is represented. In the classical one-dimensional cutting stock formulation, one variable can represent how many stock rolls use one feasible cutting pattern. There may be millions or billions of possible patterns, so writing them all down is impossible. Column generation begins with only some patterns, solves a restricted master linear programme, uses dual prices to ask which new pattern would improve the solution, and solves a knapsack-like pricing problem to generate that pattern. Branch-and-price adds integrality. Modern 2025–2026 reviews show that practical cutting still builds on this foundation while adding reusable remnants, two-dimensional geometry, industrial machine rules and uncertainty.
Quick Read
A cutting stock problem starts with stock objects—rolls, bars, sheets or plates—and demanded items smaller than the stock. A feasible cutting pattern describes one legal way to extract items from one stock object. The optimisation chooses how many times to use each pattern so demand is met while material, trim loss, cost, setup count or other objectives are minimised.
In one dimension, pattern generation is closely related to knapsack. In two dimensions, the geometry becomes harder: rectangles can be placed at different coordinates and orientations subject to overlap, edge, trim and cutting constraints. Guillotine cutting requires every cut to run from one boundary of the current rectangle to the opposite boundary, creating a hierarchical slicing structure. Free nesting can be more material-efficient and may require CNC or other flexible cutting machinery.
Classical Gilmore–Gomory column generation avoids enumerating every pattern. The master problem decides pattern usage; dual variables tell the pricing subproblem how valuable each demanded item currently is; the pricing problem searches for a stock pattern whose dual value exceeds its stock cost. If no improving pattern exists, the LP relaxation is solved. Integer production plans then require rounding, heuristics, branch-and-price or related exact methods.
Modern industry adds another question: should a remainder be thrown away or stored as usable leftover? A 2026 review of cutting stock with usable leftovers surveys one-, two- and three-dimensional formulations and notes that uncertainty and practical implementation remain important research directions. A 2025 survey of two-dimensional industrial cutting identifies raw-material reuse, defects, trim distance, pattern similarity and demand uncertainty among practical requirements across industries.
One-sentence answer: Mathematics improves the world by converting raw-material cutting into an explicit combinatorial optimisation problem, allowing factories to search enormous spaces of feasible patterns and choose production plans that use less virgin material, create less scrap and still respect the physical rules of real machines.
1. A Scrap Bin Is Often a Mathematical Symptom
Walk through a manufacturing plant and the waste can look physical: narrow strips beside a slitter, irregular sheet offcuts near a laser cutter, timber ends, textile remnants. But many of those shapes were determined before the machine moved. They are consequences of which orders were grouped together, which stock size was opened, which orientation was chosen and which pattern was scheduled.
Some waste is unavoidable. Kerf consumes material. Defects remove regions. Demand dimensions may simply not tile stock dimensions. The optimisation job is not “zero waste” as a slogan. It is to distinguish unavoidable loss from loss created by poor combinations.
That distinction turns the scrap bin into feedback. If trim loss remains high, ask whether stock assortment, order batching, pattern generation, machine constraints or business rules are causing it. Mathematics helps move the investigation upstream.
2. Cutting and Packing Are Closely Related but Not Identical Jobs
Bin packing asks how to place items into a minimum number of bins or containers. Cutting stock asks how to satisfy item demands by cutting stock objects according to patterns. The geometry can be mathematically similar, and terminology often overlaps.
The production context changes the model. Cutting may lose kerf. Items may have demand quantities greater than one. Stock objects can have different sizes and costs. A leftover can return to inventory. A pattern can require a machine setup and be repeated hundreds of times.
The distinction matters because a packing solution can be geometrically excellent and operationally impossible to cut. Cutting optimisation must preserve the manufacturing process, not only the final arrangement.
3. One-Dimensional Cutting Stock Starts With Length or Width
Suppose stock bars have length L and demanded pieces have lengths l₁,…,lₘ with quantities d₁,…,dₘ. A feasible pattern a=(a₁,…,aₘ) satisfies Σlᵢaᵢ≤L after accounting for any kerf or edge rules.
The pattern might cut two 3-metre pieces and one 2-metre piece from an 8-metre bar. Another pattern might cut four 2-metre pieces. The optimiser does not decide one piece at a time; it chooses a portfolio of patterns whose total item production covers demand.
This pattern view is the key abstraction. The factory performs cuts; the optimisation can reason at the level of complete stock-object recipes.
4. A Pattern Is a Production Recipe
For each feasible pattern j, let aᵢⱼ be the number of items of type i produced and let xⱼ be how many stock objects use that pattern. Demand constraints require Σⱼaᵢⱼxⱼ≥dᵢ for each item type.
A simple objective minimises Σⱼxⱼ, the number of identical stock objects consumed. With variable stock costs, minimise Σcⱼxⱼ. With explicit trim-loss costs or leftover value, the objective changes.
The master problem becomes compact in rows—one demand constraint per item type—and enormous in columns because each feasible cutting pattern is a possible variable.
5. The Number of Patterns Explodes Combinatorially
A stock width of 10,000 mm and dozens of order widths can admit millions of integer combinations. Add machine limits on knife count, minimum trim, maximum narrow pieces and the pattern space remains huge.
Enumerating every feasible pattern before optimisation wastes memory and time because most patterns will never appear in a good solution.
This is why cutting stock became a historic home for column generation. Instead of writing every variable, let the optimisation generate only variables that evidence says might improve the current solution.
6. The Linear Programming Relaxation Gives a Lower Bound
Production counts must ultimately be integers: a factory cannot cut 2.37 stock rolls. Relax the constraint temporarily and allow xⱼ≥0 real-valued. The resulting LP can be solved efficiently and provides a lower bound on stock usage or cost.
The fractional solution is not directly manufacturable and is extraordinarily informative. Its dual variables reveal which demand constraints are expensive to satisfy under the current pattern set.
Column generation uses those dual prices to search intelligently for a pattern worth adding. The LP relaxation becomes not merely a bound but a guide to the next variable.
7. Dual Prices Tell the Pricing Problem What Is Scarce
Each item-demand constraint receives a dual value πᵢ in the LP solution. Informally, πᵢ measures how much the objective would improve if that demand constraint were relaxed slightly, under the current basis and formulation.
A new cutting pattern containing items with high dual values is attractive because it supplies scarce demand. The pricing problem searches for a feasible pattern maximising Σπᵢaᵢ subject to stock capacity and machine rules.
If the best pattern’s dual value exceeds the cost of one stock object, it has negative reduced cost in a minimisation master problem and can improve the LP. If no such pattern exists, the restricted master solution is optimal for the full LP relaxation.
8. The Pricing Problem Is Often a Knapsack Problem
For identical one-dimensional stock of length L, choose integer counts aᵢ to maximise Σπᵢaᵢ subject to Σlᵢaᵢ≤L. That is an unbounded or bounded integer knapsack problem depending on demand and pattern rules.
Dynamic programming, branch-and-bound or specialised knapsack algorithms can solve the pricing problem. Practical restrictions—knife count, limits on narrow cuts, minimum leftover width—enter as additional constraints.
This decomposition is elegant: one enormous LP is replaced by repeated coordination between a small master problem and a structured combinatorial subproblem.
9. Column Generation Is an Economic Conversation Between Two Problems
The master says: under my current patterns, these item types are expensive. The pricing problem replies: here is a new stock pattern exploiting those prices. The master reoptimises and prices change.
This continues until no unseen pattern can improve the LP. The algorithm has effectively searched a huge pattern universe without enumerating it.
Gilmore and Gomory’s work made this pattern-generation architecture foundational in operations research. Decades later, column generation remains central because the structural reason for it has not changed: the useful variables are a tiny fraction of all possible variables.
10. A Tiny Column-Generation Example
Stock length is 10. Demands are pieces of length 6 and 4. Obvious patterns include (1,1), (0,2) and (1,0). Pattern (1,1) fills the stock exactly and will usually be attractive when both demands remain.
Add a third item length 3. Now patterns (0,1,2), (1,1,0), (0,0,3) and others compete. Which pattern matters depends on residual demand and dual prices, not only local fill percentage.
A pattern with one millimetre less utilisation can be globally better if it produces the item type currently forcing many otherwise wasteful stock rolls. Local packing efficiency and global order efficiency are not identical.
11. Rounding the LP Can Be Good and Can Be Misleading
The LP solution may use pattern A 2.4 times and pattern B 3.6 times. Rounding up produces a feasible integer plan and may overproduce items and consume extra stock. Rounding down creates shortages.
Classical heuristics round selected variables, fix them, reoptimise remaining demand and repair shortages with additional patterns. The small integer gap common in many cutting-stock instances makes such heuristics surprisingly strong.
“Surprisingly strong” is not a theorem of universal optimality. Exact solutions require integrating integrality into the pattern-generation process.
12. Branch-and-Price Adds Integrality Without Enumerating Every Pattern
Branch-and-bound divides an integer problem into subproblems by imposing decisions that exclude fractional solutions. Column generation solves LP relaxations at those nodes. Together they form branch-and-price.
The challenge is that branching constraints can destroy the clean pricing structure. Good branching schemes preserve a pricing problem that remains solvable efficiently.
Exact algorithms therefore involve two levels of combinatorics: search over integer decisions and repeated generation of patterns inside each search node.
13. Branch-and-Cut-and-Price Adds Valid Inequalities Too
Cutting planes remove fractional LP solutions without excluding valid integer plans. Branch-and-cut-and-price combines branching, cutting planes and column generation.
Belov and Scheithauer studied such combinations for one-dimensional stock cutting and two-dimensional two-stage cutting. The method illustrates a wider operations-research pattern: strong relaxations, decomposition and valid inequalities cooperate rather than compete.
The algorithm name is less important than the architecture: tighten the lower bound, generate only useful patterns and branch only when integrality still requires it.
14. Bin Packing Optimises Objects; Cutting Stock Optimises Demands
In classical bin packing, each item is usually an individual object and the goal is minimum bins. In cutting stock, items come in types with demands and patterns can produce multiple copies.
When every demand equals one, the models become closely connected. Exact algorithm literature treats bin packing and binary cutting stock together because the combinatorial structure converges.
Industrial planning often lives between textbook categories: low quantities of many custom items behave like bin packing, while high quantities of standard sizes behave like cutting stock. A solver architecture should match that demand profile.
15. Trim Loss Is Material Leftover and Not All Leftover Is Waste
Traditional objectives often minimise stock count or trim loss, treating unused stock area as waste. In real factories, a sufficiently large remainder can be labelled, stored and reused on a later order.
The 2026 review of cutting stock with usable leftovers highlights this distinction across dimensions and planning settings. A remainder’s value depends on size, shape, material grade, storage cost and probability of future use.
Calling every remainder scrap can lead optimisation to destroy valuable future stock. Calling every remainder inventory can fill warehouses with unusable fragments. Mathematics needs a value model between those extremes.
16. Usable Leftovers Make the Problem Multi-Period
A leftover saved today changes tomorrow’s available stock. The optimisation is no longer one batch in isolation. Inventory state carries across periods.
Multi-period models decide whether to consume new stock, reuse leftovers, create new remnants intentionally and hold them for future demand. Storage capacity and handling cost enter. Uncertain future orders make the value of a remnant stochastic.
The problem begins to resemble inventory control: material is not merely cut; it is transformed among stock states over time.
17. A Leftover Needs a Future Probability of Use
Two remnants have equal area. One is a long narrow strip matching common future parts. The other is an awkward shape no order normally uses. Their economic value differs dramatically.
Historical demand can estimate reuse probability by dimension and material grade. A remnant value function can credit patterns for producing pieces likely to be reused and penalise storage of fragments with low expected value.
This is why “percentage utilisation” can be misleading. Material value is state- and demand-dependent. A good plan measures what remains useful, not merely what remains uncut.
18. Kerf Turns Geometry Into Accounting
A saw blade, laser, waterjet or slitting process removes material along a cut. Kerf width can be small relative to a stock sheet and large relative to tight nesting margins.
In one-dimensional cutting, total consumed length includes item lengths plus kerf contributions depending on sequence and whether edge trimming counts. In two-dimensional nesting, parts need spacing reflecting tool diameter, heat-affected zone or process clearance.
A geometric plan ignoring kerf may fit perfectly on screen and be physically impossible. Manufacturing optimisation must distinguish nominal part boundary from toolpath envelope.
19. Trim Distance Protects Parts From Stock Edges
Some materials have damaged edges, coatings, mill tolerances or handling marks. Parts must begin a minimum distance from the stock boundary.
The 2025 survey of industrial two-dimensional cutting explicitly identifies trim distance as a practical requirement. The usable stock rectangle is smaller than the nominal sheet, sometimes asymmetrically.
Edge margins demonstrate a recurring theme: plant rules change feasible geometry before optimisation. Feeding nominal dimensions to a solver and adding margins manually afterward can invalidate the pattern.
20. Defects Turn a Sheet Into a Map of Forbidden Regions
Wood contains knots. Steel plate can have surface defects. Glass may contain inclusions. Leather and natural stone have irregular quality regions.
A two-dimensional cutting model can represent defect rectangles, polygons or quality zones that items must avoid. Some item classes may tolerate certain defects while premium visible surfaces cannot.
Defect-aware nesting often produces less tidy patterns and much better material economics because the solver uses imperfect stock intelligently rather than discarding the entire sheet.
21. Grain and Direction Make Rotation Conditional
A rectangular part can geometrically fit after a 90-degree rotation and still be unusable because wood grain, fabric warp, brushed-metal direction or printed pattern must align.
Each item therefore has allowed orientations. Some allow 0° and 180° only, which are geometrically equivalent for rectangles but not necessarily for patterned material. Free-form nesting can include arbitrary rotations when the process and material allow them.
Orientation constraints are small in code and large in consequence. An optimisation that ignores them can report spectacular waste savings for parts the customer will reject.
22. Guillotine Cutting Restricts How Geometry Can Be Produced
A guillotine cut travels from one edge of the current rectangular piece to the opposite edge, splitting it into two rectangles. Panel saws and some glass or paper processes operate naturally this way.
Not every non-overlapping rectangle packing is guillotine-feasible. A visually dense nesting can require an internal cut that cannot be reached without cutting through another part.
Guillotine constraints therefore add hierarchical structure. The solver must design a cut tree, not just coordinates. Manufacturing sequence and geometry become one object.
23. Two-Stage Guillotine Cutting Mirrors Real Slitting Processes
In two-stage cutting, a stock sheet may first be divided into strips in one direction and then each strip cut in the perpendicular direction. Paper, glass and panel operations often resemble this structure.
The restriction reduces flexibility and simplifies machine handling. Exact branch-and-price research has solved large instances of two-stage guillotine cutting by extending pattern-based formulations.
The best mathematical pattern depends on the machine architecture. More geometric freedom is not always operationally desirable if it requires slower equipment or complicated handling.
24. Free Nesting Uses More Geometry and More Computational Difficulty
CNC laser, plasma, router and waterjet systems can trace complex toolpaths and are not limited to guillotine cuts. Parts can be arranged tightly with rotations, shared edges or irregular boundaries where process permits.
This increases material utilisation and turns the problem into two-dimensional or irregular nesting with collision detection and toolpath constraints. Exact optimisation becomes difficult; industrial systems use constructive heuristics, local search, metaheuristics and mixed-integer formulations selectively.
The geometry is continuous, yet manufacturing decisions remain discrete. A one-millimetre move can make one extra part fit and save an entire sheet.
25. No-Fit Polygons Convert Irregular Overlap Into Geometry
For irregular polygonal parts, the no-fit polygon describes relative placements where two parts touch without overlap. If the reference point of one part enters the forbidden interior region, the parts overlap.
This transforms repeated collision checking into precomputed geometric relationships and supports nesting search over candidate contact positions.
Computational geometry becomes an optimisation primitive. The solver reasons not only about area but about shape compatibility along boundaries.
26. Rectangle Packing Can Use Coordinate Variables or Patterns
One formulation assigns x and y coordinates to every item and binary variables indicating relative order: item A lies left of, right of, above or below B. Big-M constraints enforce non-overlap.
Another approach enumerates or generates feasible patterns or shelves. Dynamic programming builds guillotine structures. Arc-flow models represent placements through networks.
No formulation dominates every instance. Pattern methods can give strong relaxations and huge column sets. Coordinate models are direct and can have weak relaxations. Solver design is representation engineering.
27. Shelf and Level Heuristics Trade Flexibility for Speed
Sort items by height, create horizontal shelves and pack items left to right. The restriction ignores many interlocking arrangements and dramatically reduces search.
First-fit decreasing height and related heuristics can generate good starting solutions quickly. In online or high-volume settings, a fast solution available now can be more valuable than a marginally better solution arriving after production should have started.
Heuristics are not mathematical surrender. They are explicit choices about computation budget and solution quality.
28. Lower Bounds Tell Us Whether a Heuristic Is Actually Good
A plan uses 103 sheets. Is that excellent or terrible? If total demanded area divided by sheet area gives a lower bound of 100, the gap is at most three sheets before geometry strengthens the bound.
LP relaxations, area bounds, dual-feasible functions and combinatorial bounds provide benchmarks against which heuristic solutions can be judged.
Without a lower bound, “we saved 5% compared with the old heuristic” can hide large remaining opportunity. Optimisation needs both a feasible plan and evidence about how far that plan might be from the best possible.
29. Pattern Count Is a Hidden Factory Cost
Two solutions use the same number of stock rolls. Solution A uses four distinct cutting patterns. Solution B uses twenty-seven. B may require many knife changes, CNC programmes, operator instructions and quality checks.
Practical cutting-stock models often limit or penalise the number of patterns. Research on paper-industry cutting explicitly discusses trading trim loss against pattern count and machine knife restrictions.
The mathematical optimum in material can be economically inferior after setup cost. Multi-objective optimisation exposes the trade-off instead of hiding setup labour outside the model.
30. Knife Limits Turn One Feasible Pattern Into an Impossible Setup
A slitter may have only k movable knives. A mathematically valid roll pattern requiring twelve simultaneous slits is unusable on a machine with eight knives.
The pricing problem can include a cardinality constraint on distinct cut positions or produced strips. Other machines limit narrow strips, minimum spacing or edge-trim settings.
Adding these rules during pattern generation is far better than discovering after optimisation that half the patterns cannot be loaded onto the machine.
31. Sequencing Patterns Can Matter After Pattern Selection
A cutting plan may specify pattern counts and ignore order. In production, changing from one pattern to another takes setup time. Related patterns can be sequenced to minimise knife movement or tooling changes.
This creates a second optimisation problem resembling travelling-salesperson or scheduling structure over selected patterns. Integrated models can trade material loss against transition cost.
Optimisation often comes in layers: choose what to produce, then choose the sequence. When transition cost is large, solving those layers independently can be suboptimal.
32. Overproduction Is Not Always Waste and Not Always Free
A pattern may produce one extra part because exact demand matching creates much more trim. If the extra item can enter inventory and is likely to sell, overproduction may be rational.
If the part is custom and useless after the order, overproduction is another form of waste. Models therefore allow overproduction with item-specific inventory value or penalty.
Demand constraints Σaᵢⱼxⱼ≥dᵢ do not by themselves price the excess. The objective needs economics attached to quantities above demand.
33. Multiple Stock Sizes Change the Pattern Universe
A factory may hold 6-m and 12-m bars, several sheet sizes, virgin rolls and reusable remnants. Each stock type has cost, availability and its own feasible pattern set.
The master problem includes variables indexed by pattern and stock type. Pricing runs separately or jointly across stock options, searching for the most negative reduced-cost combination.
Having more stock sizes can reduce waste and increase inventory complexity. Optimisation reveals whether assortment flexibility pays for the warehouse burden.
34. Stock Cost and Waste Cost Are Different Objectives
A larger sheet may create more nominal waste and cost less per square metre. A remnant may have zero purchasing cost and high handling cost. Recycled material may be cheaper and limited in quality.
Minimising physical waste does not always minimise economic cost. Sustainability metrics can add embodied carbon or disposal cost. Multi-objective models show trade-offs among material consumption, money, emissions and setup effort.
The optimiser needs the objective the organisation actually cares about. “Waste” should not become a vague moral word when its operational meaning can be quantified several ways.
35. Carbon Accounting Can Change Which Pattern Is Best
Two raw materials with equal purchase cost can carry different embodied carbon. A slightly more expensive reused remnant can avoid opening a new high-carbon plate. Scrap streams can also have recycling credits.
Add carbon coefficients to stock use, cutting energy and scrap treatment. Optimise cost subject to an emissions cap, or optimise a weighted objective, or generate a Pareto frontier.
Sustainability becomes more credible when the trade-off is explicit. Mathematics does not decide the social price of carbon; it shows how a chosen value changes production plans.
36. Demand Uncertainty Makes Perfect Patterns Fragile
A plan optimised for exact orders can perform poorly when customers change quantities tomorrow. Cutting all stock into today’s exact parts can leave no flexible material for late demand.
The 2025 industrial survey and 2026 leftover review both highlight uncertainty as a practical challenge. Stochastic programming can optimise expected cost across demand scenarios. Robust optimisation protects performance across a set of plausible demands.
A robust plan may leave more material uncut today because optionality has value. Maximum immediate utilisation is not always maximum long-term efficiency.
37. Two-Stage Stochastic Cutting Separates Decisions by When Information Arrives
First-stage decisions are made before uncertain demand is known: perhaps purchase stock or pre-cut standard intermediate sizes. Second-stage recourse decisions respond after demand is revealed.
Optimisation minimises first-stage cost plus expected recourse cost across scenarios. Scenario trees become large; decomposition methods such as Benders variants can help.
The modelling principle is general: do not let an optimiser use tomorrow’s information in today’s decision. Stochastic programming keeps the information timeline honest.
38. Robust Optimisation Protects Against a Set Rather Than a Probability Distribution
Sometimes managers know demand can vary within ranges and do not trust a precise distribution. Robust optimisation chooses a plan that remains feasible or performs acceptably across an uncertainty set.
Worst-case protection can be conservative. Budgeted uncertainty and adjustable decisions moderate that conservatism. The appropriate model depends on what is known about future orders and how expensive shortages are.
Probability models and robust sets answer different epistemic situations. Mathematics should reflect the quality of knowledge rather than manufacture a distribution because the solver prefers one.
39. Forecasting and Cutting Should Talk to Each Other
A forecast predicts likely part demand. A cutting optimiser decides how much material to commit. If forecasts include uncertainty intervals or scenario probabilities, the optimiser can value leftovers and overproduction intelligently.
Repeated rolling-horizon optimisation updates the cutting plan as new orders arrive. Material not yet cut remains flexible. Saved remnants become state variables.
The manufacturing system becomes a feedback loop: forecast, optimise, cut, observe demand, reconcile inventory, optimise again.
40. Rolling Horizons Prevent One Giant Forecast From Controlling the Year
A year-long exact optimisation assumes a year-long demand forecast deserves equal trust. In practice, near-term orders are known better than distant ones.
Rolling-horizon planning optimises a finite future window, commits the near-term cuts, then moves the horizon forward with updated information.
This balances global coordination with flexibility. The main risk is myopia: decisions near the end of each window can harm future periods. Terminal values for leftover inventory help preserve future value.
41. Common-Line Cutting Can Save Kerf and Time
Two adjacent rectangular parts can sometimes share one cut instead of requiring separate parallel toolpaths with spacing. Common-line cutting reduces kerf and machine time.
It is only feasible when material, tool and quality rules permit shared boundaries. Thermal processes may require spacing to manage heat distortion. Some parts need edge finishing that prevents sharing.
Once allowed, the objective includes not only placement but adjacency. The value of putting two parts together depends on their shared edge length.
42. Toolpath Optimisation Begins After Nesting and Can Feed Back Into It
A nesting plan tells where parts lie. A CNC machine still needs an ordered path of cuts, pierce points and rapid traverses. Shorter toolpaths reduce cycle time and energy.
Travel among cut contours resembles routing. Thermal constraints can force alternation to avoid heat accumulation. Internal holes must be cut before outer contours so parts do not move prematurely.
An extremely tight nest can create a terrible toolpath. Integrated optimisation trades material utilisation against cutting time and process stability.
43. Heat-Affected Zones Turn Distance Into Quality
Laser and plasma cutting heat the material around the toolpath. Place cuts too close and accumulated heat can warp thin parts or alter metallurgy.
Minimum spacing and sequence constraints can model heat management approximately. More advanced systems use thermal simulations or empirical rules.
This is another reason geometric density is not the only objective. A nesting percentage can be excellent and the production quality unacceptable.
44. Cutting Direction Can Change Edge Quality
Machining, routing and some composite cutting processes have preferred directions. Climb versus conventional milling changes forces and finish. Fibres can tear differently depending on cut direction.
Part orientation and toolpath direction therefore become manufacturing constraints beyond pure shape. One edge may require a finishing pass.
Operations research works best when process engineering supplies these rules explicitly instead of asking an optimiser to discover them accidentally from historical scrap.
45. Stack Cutting Multiplies Output and Adds Thickness Constraints
Textile, paper and some sheet processes cut several layers simultaneously. A pattern applied to a stack produces multiple copies per cut.
Stack height is limited by tool capability and material behaviour. Demand quantities may determine how many layers make sense. Colour or material lots can restrict which sheets may be stacked.
The optimisation now chooses pattern, stack composition and layer count. Production rate and material utilisation become coupled.
46. Batch and Lot Constraints Create Combinatorial Coupling
Parts for one customer may need the same material heat, dye lot or grain pattern. Mixing lots inside a pattern can complicate traceability.
Add lot compatibility constraints or require certain item groups to come from designated stock batches. This can increase waste while protecting quality and auditability.
Traceability is a real objective even when it does not appear in the geometry. Manufacturing optimisation lives inside a larger quality system.
47. Pattern Similarity Can Be a Production Objective
The 2025 industrial survey identifies cutting patterns based on item similarity among practical aspects. Similar patterns can reduce setup complexity, handling and sorting.
A solver can cluster orders by dimensions, due dates or downstream process. It can penalise patterns mixing many item families.
The result may use slightly more material and reduce hidden operational friction enough to lower total cost. This is why an optimisation model should evolve after observing the factory rather than remain a clean textbook objective forever.
48. Due Dates Turn Cutting Into Scheduling
Two orders need the same material. One ships today, one next week. A pure cutting-stock model can combine them to reduce waste and delay the urgent order.
Integrated models add due dates, tardiness penalties and machine availability. Patterns acquire time as well as material meaning.
Production planning then chooses not only how to cut, but when. A globally efficient material plan can be commercially wrong if it misses customer commitments.
49. Multiple Machines Create Assignment Decisions
A factory can have saws, lasers or slitters with different bed sizes, speeds, kerfs and capabilities. A pattern feasible on one machine may be impossible or expensive on another.
Decision variables assign patterns or jobs to machines. Machine-specific pricing problems generate patterns respecting each capability. Scheduling constraints limit capacity by shift.
Material optimisation and equipment loading interact. One large machine can cut efficiently and become a bottleneck while smaller machines sit idle.
50. A Digital Twin Can Test Cutting Plans Before Material Is Touched
A digital production model can simulate pattern loading, toolpath, cycle time, machine queues and remnant inventory. Optimisation proposes plans; simulation checks whether shop-floor effects create unexpected congestion.
This is especially useful when setup and handling effects are difficult to express analytically. Simulation can estimate the downstream cost of a material-efficient but complex plan.
The important boundary is that a visually realistic digital twin does not validate a bad process model. Measured cycle times and scrap data must return to calibrate the twin.
51. Heuristics Give a Fast Incumbent While Exact Methods Prove How Good It Is
Industrial solvers often combine approaches. A constructive heuristic quickly builds a feasible plan. Local search improves it. LP or mixed-integer optimisation provides lower bounds and explores difficult choices. Exact methods can certify optimality when time allows.
A strong incumbent makes branch-and-bound faster because poor nodes can be pruned. A strong lower bound tells managers whether further compute is worth waiting for.
“Heuristic versus exact” is therefore a false binary. The most effective systems often let them help one another.
52. Large Neighbourhood Search Rebuilds Part of a Plan
Take an existing cutting solution. Remove several badly utilised stock objects and repack their items together. Or destroy patterns associated with one bottleneck and regenerate them.
Large Neighbourhood Search modifies enough decisions to escape local minima while retaining most of a good plan. Adaptive variants learn which destroy-and-repair operators work well on the current instance family.
The method fits industry because existing production plans often need improvement or repair rather than complete reinvention after every order change.
53. Genetic Algorithms Search Populations of Patterns or Layouts
Evolutionary algorithms encode a candidate plan, recombine parent solutions and mutate decisions. Fitness measures waste, cost and constraints.
For geometric nesting, specialised crossover operators are needed because naive recombination destroys feasibility. Hybrid genetic algorithms often use deterministic placement heuristics inside the chromosome evaluation.
Metaheuristics are attractive for irregular and highly constrained problems where exact formulations scale poorly. They provide good solutions without general optimality certificates.
54. Reinforcement Learning Can Learn Placement Rules and Still Needs Hard Constraints
Recent research applies learning to packing and cutting heuristics: choose the next item, orientation or placement based on a learned policy.
Learning can exploit recurring order distributions and produce fast decisions. It can also produce infeasible placements unless geometry and machine rules are enforced externally.
A strong architecture lets learning rank or propose actions while exact collision and constraint checks remain authoritative. Creativity can be learned; feasibility should not be optional.
55. Real Waste Data Should Return to the Model
The optimiser predicted 96.5% utilisation. Actual measured yield was 93%. Where did the missing material go? Kerf larger than assumed? Damaged stock? Operator overrides? Scrapped parts? Remnants not returned to inventory?
Production receipts can compare predicted and actual material flows. Persistent discrepancies reveal model ownership problems rather than optimiser weakness alone.
A cutting optimiser should become a learning system: plan, cut, measure, reconcile, update constraints and costs, then plan again.
Advanced Field Manual | How Industrial Cutting Optimisers Actually Work
56. Column Generation Needs Stabilisation in Difficult Instances
Pure column generation can oscillate because dual prices change sharply as new patterns enter. A pricing problem generates one extreme pattern, the master pivots, duals swing and the next iteration generates a very different pattern.
Dual stabilisation methods keep prices near a centre or penalise large moves, producing smoother convergence. Interior-point information, box constraints around duals or bundle-like penalties can reduce the number of unproductive iterations.
The mathematical optimum is unchanged when stabilisation is designed correctly. What changes is the path taken to reach it. Algorithm engineering matters even after the model is theoretically solved.
57. Multiple Pricing Can Add Several Good Patterns Per Iteration
Classical column generation adds the single best negative-reduced-cost pattern. Solving the master after every one new pattern can be expensive when many similar useful patterns exist.
Multiple pricing returns several attractive patterns at once, perhaps through k-best knapsack solutions or diversity rules. This enlarges the restricted master faster and can reduce master/pricing alternations.
Too many columns create their own burden. The useful balance depends on pricing speed, master size and how redundant the generated patterns are.
58. Heuristic Pricing Finds Useful Patterns Before Exact Pricing Proves None Remain
Early column-generation iterations often contain obvious improving patterns. A fast heuristic can find them without solving the knapsack or geometric pricing problem to optimality.
Once heuristic pricing fails, an exact pricing algorithm is required before concluding no negative-reduced-cost column exists. This two-level strategy saves time and preserves correctness of the LP bound.
The pattern repeats across optimisation: approximate methods generate progress; exact methods certify when the approximation has run out of useful ideas.
59. Dominated Patterns Should Not Waste Master-Problem Space
If pattern A produces at least as many of every demanded item as pattern B using the same stock and no more setup cost, B is dominated under a simple objective.
Preprocessing can remove dominated patterns in enumerated models. Pricing algorithms can avoid generating patterns with obvious unused capacity when another feasible item still fits, subject to practical constraints.
Dominance becomes more complex when leftovers have value or pattern count matters. A pattern with greater unused width may intentionally create a valuable reusable remnant and cease to be dominated.
60. Reduced Cost Is a Local Economic Signal, Not the Final Integer Value
A pattern with very negative reduced cost can dramatically improve the LP relaxation and never appear in the final integer solution. Another pattern with weak reduced cost can become essential after branching or rounding.
Dual prices describe the current fractional master. They are guidance, not a guarantee of integer importance.
This explains why a good column pool often retains more than the patterns active in the final LP basis. Integer repair needs options the fractional optimum did not use heavily.
61. Column Pools Preserve Useful Manufacturing Knowledge
Factories repeatedly cut similar dimensions. Patterns discovered today can become excellent starting columns tomorrow. A pattern pool stores historically useful recipes by material, machine and stock size.
Warm-starting with that pool reduces pricing iterations and gives planners familiar setups. Old patterns still need feasibility checks because knife counts, trim rules or order dimensions may change.
Optimisation can therefore accumulate institutional memory. The solver does not need to rediscover every good combination from zero each morning.
62. Demand Aggregation Reduces Model Size and Can Hide Important Detail
Orders with identical dimensions and material can be aggregated into one item type with total demand. This makes cutting-stock models dramatically smaller.
Aggregation can be wrong when the orders have different due dates, customers, quality grades or traceability requirements. Two 500-mm strips are not interchangeable if one must ship today from a specific batch.
Good modelling aggregates physical equivalence and keeps business distinctions that actually constrain production.
63. Symmetry Can Waste Enormous Search Effort
Ten identical stock sheets are indistinguishable. A coordinate MIP assigning items to sheet 1, 2, …, 10 can explore many solutions that differ only by sheet labels.
Symmetry-breaking constraints order stock usage or impose canonical assignments. Pattern formulations avoid much of this symmetry by describing stock recipes rather than named individual sheets.
Representation affects computation because a solver can spend hours proving that relabelled copies of the same physical plan are equivalent unless the model tells it not to.
64. Arc-Flow Models Compress Many Patterns Into a Network
An arc-flow formulation represents ways to fill a stock length as paths through a directed graph whose nodes correspond to used capacity. Arcs represent placing items or leaving loss.
Many complete cutting patterns share subpaths and are represented compactly in one network. Flow constraints select combinations without listing every pattern explicitly.
Graph compression can make general-purpose integer programming surprisingly competitive with column generation on some instances. Again, the central question is how to represent a giant combinatorial family without enumerating every member.
65. Dynamic Programming Is Both a Solver and a Pattern Generator
The one-dimensional pricing knapsack can be solved with dynamic programming over capacity. Let V(c) be the best dual value obtainable using capacity c. Recurrences consider adding each item length.
The table yields the best reduced-cost pattern and can recover the chosen item counts by backtracking. Machine restrictions can add state dimensions such as knife count or number of narrow items.
Dynamic programming is effective because capacity is one-dimensional and discrete. Two-dimensional pattern generation loses that simplicity and requires richer states or other techniques.
66. Two-Dimensional Pricing Is an Optimisation Problem Inside an Optimisation Problem
A pattern-based 2D master can still use column generation, but the pricing problem must itself find a high-dual-value non-overlapping arrangement of rectangles under cutting constraints.
That pricing subproblem may be NP-hard and solved using dynamic programming for guillotine structures, MIP, constraint programming or heuristics.
Decomposition does not make complexity disappear. It moves complexity into a subproblem whose structure is better aligned with specialised algorithms.
67. Constraint Programming Is Natural for Geometric Rules
Constraint programming represents item coordinates, orientation and non-overlap relations with global constraints and propagates infeasible domains before explicit search.
It can express logical machine rules—if this orientation is chosen, this edge must stay away from a defect; if two parts share a cut, their lengths must match—more directly than some MIP formulations.
Hybrid CP/MIP systems exploit the strengths of each: CP for combinatorial geometry, linear optimisation for costs and aggregate resource use.
68. CP-SAT Blurs the Boundary Between Logical and Linear Search
Modern SAT-based optimisation engines combine Boolean propagation, integer constraints and linear relaxations. They can be effective on scheduling and placement variants with many discrete rules.
For moderate 2D instances, optional interval variables can represent item placement and no-overlap constraints. Search heuristics exploit logical structure that big-M formulations may hide.
The best industrial solver architecture is empirical. Model the true rules, benchmark formulations on real order distributions and keep the one that produces reliable solutions inside the planning deadline.
69. Big-M Constants Can Make a Correct MIP Numerically Weak
To express “A is left of B or above B”, a MIP often uses binary variables and big-M inequalities. If M is excessively large, the LP relaxation becomes weak and numerical conditioning suffers.
Tight M values derived from sheet dimensions strengthen the formulation. Indicator constraints can express logic more cleanly when supported by the solver.
This is a general optimisation lesson: a mathematically valid formulation can be computationally terrible. Bounding constants with physical knowledge is part of modelling quality.
70. Valid Inequalities Teach the Solver Facts Humans Already Know
Area alone gives a weak bound. If five large items cannot pair on one sheet because of dimensions, at least five sheets may be needed even though total area suggests four.
Valid inequalities encode such combinatorial truths, tightening relaxations without excluding feasible integer plans. Cover inequalities, clique inequalities and problem-specific cuts can reduce branch-and-bound search.
Good cuts are compressed reasoning. They tell the solver conclusions that would otherwise require exploring many branches to rediscover.
71. Lagrangian Relaxation Prices Difficult Constraints Instead of Enforcing Them Immediately
Suppose a complicating demand or machine constraint couples otherwise separable subproblems. Move it into the objective with a multiplier and solve easier relaxed problems.
Multipliers are adjusted to penalise violation, generating lower bounds and sometimes good primal heuristics. Lagrangian ideas connect closely to dual prices and decomposition.
Relaxation is not ignoring the rule. It is temporarily converting the rule into a price so the problem structure can be exploited computationally.
72. Benders Decomposition Separates Strategic Stock Decisions From Operational Cutting
A planning model may choose how much material to buy or which stock sizes to hold while operational subproblems decide whether each demand scenario can be cut efficiently.
Benders decomposition lets a master choose strategic variables and subproblems return feasibility or optimality cuts describing the downstream consequence.
The method becomes attractive in multi-period and stochastic planning where cutting detail would otherwise overwhelm a large scenario model.
73. Remnant Inventory Needs Identity, Not Just Total Area
A warehouse containing 100 m² of leftovers can be nearly useless if every piece is too narrow. Area aggregation destroys the dimensional information that determines future cut feasibility.
Remnants need attributes: dimensions, thickness, grade, lot, age, storage location and perhaps defect map. Inventory systems can classify them into equivalence classes or store individual piece identities.
Optimisation quality is capped by remnant data quality. An unlabelled offcut on a rack is mathematically invisible even if it is physically valuable.
74. Barcodes and Vision Can Close the Remnant Data Loop
After cutting, a camera or machine controller can measure remnant geometry. Print a barcode or RFID tag with material and dimensions. The warehouse system records location.
The next optimiser run can treat that remnant as a stock object with known cost and constraints. If the piece is consumed, inventory state updates automatically.
This turns “reuse leftovers” from an operator memory task into a closed information loop. Mathematics needs logistics infrastructure before sustainable optimisation can become routine.
75. Remnant Age Can Reduce Future Value
Some materials corrode, warp, oxidise or become obsolete as product designs change. A leftover’s value can decay through time.
Inventory objectives can include holding cost or age-dependent salvage value, encouraging old remnants to be consumed before new stock where feasible.
Without such a penalty, an optimiser may continually save “useful” remnants and never use them, creating a mathematically tidy but physically growing graveyard of offcuts.
76. The Warehouse Can Become the Bottleneck of Sustainable Cutting
Saving every reusable piece increases storage, search and handling. A forklift may spend more time finding remnants than the material is worth.
Set minimum remnant dimensions, maximum inventory count or storage-capacity constraints. Add retrieval cost by location or awkwardness.
Sustainability is a system objective. Material reuse that doubles warehouse labour and damages remnants through poor handling can reduce overall benefit.
77. Due-Date Batching Can Create Material Synergy
Orders due within nearby windows can be pooled so complementary dimensions share stock. Waiting one hour for another order can reduce scrap dramatically.
Delay too long and service suffers. Real-time order batching becomes an online optimisation problem balancing material synergy against customer waiting.
This mirrors ride pooling and freight consolidation: efficiency comes from matching compatible demand, and matching requires enough time for compatible demand to arrive.
78. Online Cutting Decisions Face an Information Constraint
An order arrives now. Future orders are unknown. Cutting a full sheet immediately may create scrap; waiting may combine it with a future order that never arrives.
Online algorithms make irrevocable or partially reversible decisions without seeing the future. Competitive analysis compares their performance with an omniscient offline optimum.
Forecasts improve decisions and do not remove the fundamental asymmetry. Real factories live online; historical benchmark instances often assume all orders are known.
79. Scenario Simulation Tests Whether a Policy Survives Demand Volatility
Generate future order streams from historical distributions or stress scenarios. Run the remnant and cutting policy across months. Measure virgin stock use, service level, remnant inventory and setup effort.
A policy optimised for one static batch can look wonderful and accumulate useless remnants over repeated operation. Simulation reveals the long-run consequence of local rules.
Policy evaluation should match the time horizon over which management claims the savings.
80. A Solver Must Produce a Plan Operators Can Read
An optimal solution represented as anonymous column 17 used 43 times is not a production instruction. Operators need pattern drawings, cut dimensions, stock IDs, setup sequence and part labels.
Explainable output includes why a remnant was selected, which customer parts share a sheet and what alternative plan exists if material is damaged.
The last mile of optimisation is translation. A mathematical solution creates value only when the shop floor can execute it accurately enough for the predicted savings to survive.
Advanced Field Manual | Measuring Whether the Optimiser Helped the Factory
81. Utilisation Percentage Is Necessary and Not Sufficient
Material utilisation is typically useful area divided by stock area or useful length divided by stock length. A rise from 90% to 95% sounds compelling.
Now count setup hours, remnant handling, production delays, overproduction and machine idle time. A high-utilisation plan can lose money if it fragments production into too many patterns.
Dashboards should therefore pair utilisation with total stock cost, trim-loss value, pattern count, setup time, throughput and service performance. One ratio cannot carry the whole factory.
82. Trim Loss Should Be Reported in Physical and Economic Units
Ten kilograms of titanium and ten kilograms of mild steel are not the same loss. Nor are ten square metres of decorative veneer and commodity board.
Report scrap quantity, scrap percentage and monetary value. Add carbon or scarcity metrics where relevant. Separate reusable leftovers from irreversible scrap.
This prevents a low-cost high-volume material from dominating a percentage metric while expensive rare material remains badly optimised.
83. Yield Variance Shows Whether the Process Matches the Plan
The optimiser predicts 94% yield across twenty batches. Actual yield varies from 88% to 95%. The average gap matters and the variance matters.
High variance can indicate inconsistent machine calibration, operator changes, material defects or unrecorded rework. The optimisation model may be fine while execution is unstable.
Statistical process control on yield creates an early warning that the planning model and physical process are drifting apart.
84. Pattern Adoption Rate Reveals Whether Operators Trust the Solver
If planners override 40% of generated patterns, something important is missing. Track override frequency and reason codes.
Operators may know a machine rule the model lacks, dislike unstable plans, or prefer standard setups. Some overrides may be habitual and costly; others reveal genuine requirements.
An optimiser should learn from overrides instead of treating them only as noncompliance. Human disagreement is data about model ownership.
85. Solve Time Is a Production KPI
A mathematically superior solution arriving two hours after the cutting shift starts is operationally inferior to a near-optimal plan produced in two minutes.
Measure time to first feasible solution, time to target optimality gap and time to certified optimum. Set service-level objectives for planning latency.
This allows an anytime architecture: start with a feasible heuristic, improve while time remains, stop when production needs the plan.
86. Optimality Gap Makes Waiting a Business Decision
Current plan costs $100,000. The solver’s lower bound says no plan can cost less than $99,500. The gap is about 0.5%. Waiting another hour to prove the last few hundred dollars may not be worthwhile.
Another instance costs $100,000 with a lower bound of $90,000. A 10% gap signals major uncertainty about solution quality.
Bounds turn solver progress into managerial information. They answer not merely “is this plan good?” but “how much improvement can possibly remain?”
87. A/B Testing Optimisers Requires Comparable Order Mixes
Week A uses the old planner and easy repeat orders. Week B uses the new optimiser and custom irregular demand. Comparing raw utilisation is unfair.
Replay identical historical order batches through both methods, or randomise comparable jobs when operationally possible. Control for material, stock sizes, machine availability and demand complexity.
Optimisation claims deserve experimental design too. Better algorithms should beat the baseline on matched problems, not merely during a favourable month.
88. Worked Example: The Long Strip That Saves a Roll Tomorrow
Today’s order leaves either one 900-mm strip or three 300-mm strips of equal total width. Immediate waste is identical if all leftovers are credited equally.
Historical orders frequently request parts 850 mm wide. The 900-mm strip has high reuse probability; the 300-mm strips almost none. A leftover-value model credits the first pattern.
Tomorrow, the saved strip avoids opening a new stock roll. What looked like equal utilisation today becomes unequal total material use over two periods. Time changes the value of geometry.
89. Worked Example: Fewer Patterns Beat Lower Trim
Plan A wastes 1.5% and uses eighteen patterns. Plan B wastes 2.0% and uses five. Each pattern change costs twelve minutes of machine setup.
Thirteen avoided changes save 156 minutes. If material value of the extra 0.5% is lower than labour, machine capacity and delay cost of those setups, Plan B is economically superior.
The example explains why “minimum trim” is not automatically the right objective. Total cost needs to include the factory’s real conversion process.
90. Worked Example: A Defect-Aware Sheet Beats a Perfect Sheet
A discounted steel plate has one small defect near a corner. A clean plate costs more. Without defect modelling, the damaged plate is rejected for a high-quality order.
Defect-aware nesting places internal cutouts or tolerant low-grade parts around the damaged region and premium surfaces elsewhere. The plate becomes fully usable for a mixed order.
The optimisation has created value not by changing material, but by matching quality constraints to spatial defects precisely.
91. Worked Example: Why the Best Local Pattern Can Be Globally Bad
Pattern P fills 99% of a roll with common 200-mm strips. Pattern Q fills 96% and includes one awkward 370-mm strip whose demand is small.
If all awkward 370-mm demand is ignored until the end, the factory may open several nearly empty rolls just for those pieces. Using Q earlier can reduce total rolls even though each Q roll has lower utilisation.
Column-generation dual prices capture this global scarcity. The awkward item receives a high dual value and makes Q attractive. The method optimises the order set, not one roll in isolation.
92. Worked Example: Kerf Changes the Last Item
A 3,000-mm bar must produce three 1,000-mm pieces. On paper, perfect fit. The saw kerf is 3 mm per separating cut.
Two internal cuts consume 6 mm, so three nominal 1,000-mm pieces cannot all come from a 3,000-mm bar unless part tolerances or edge conventions change the arithmetic.
The example is trivial and expensive in production if forgotten. Manufacturing optimisation begins by modelling dimensions the machine actually consumes.
93. Worked Example: Why Area Bounds Can Lie About Two-Dimensional Feasibility
A sheet has area 100. Two requested rectangles each have area 45, so total area 90 suggests both might fit.
If each rectangle is 9×5 and the stock is 10×10, they can fit stacked. Change each to 6×7: total area 84, yet two may fail to fit under restricted orientation depending on placement.
Area is a necessary resource and not sufficient geometry. Two-dimensional cutting is hard because shape and dimension distribution matter beyond scalar utilisation.
94. Failure Mode: The Solver Optimises the Wrong Unit of Measure
One system stores millimetres; another sends inches. A stock width is interpreted as 96 mm instead of 96 inches. Or kerf is entered in centimetres against millimetre parts.
Geometry can become infeasible or absurd while remaining numerically valid. Unit validation, ranges and dimensional metadata belong in the data contract.
High-end optimisation cannot rescue elementary semantic corruption. Data meaning is part of Mathematics because equations assume quantities are commensurable.
95. Failure Mode: Nominal Dimensions Ignore Tolerance
A part nominally 100 mm wide allows ±0.5 mm; stock width varies by mill tolerance; kerf changes with tool wear. Deterministic exact dimensions can produce patterns with no process margin.
Robust models reserve tolerance bands or use worst-case dimensions where quality demands. Statistical process capability can justify tighter margins when measurements support them.
The optimisation should not consume safety margin that process engineering never agreed to sell.
96. Failure Mode: The CAD Shape and Toolpath Shape Are Different
A CAD contour describes finished geometry. The tool centreline may need offset by tool radius. Tabs, lead-ins, lead-outs and corner treatment need extra space.
Nesting directly on finished contours can create collisions between toolpaths or insufficient skeleton material holding parts in place.
Production nesting should use the manufacturing envelope, not only the design envelope.
97. Failure Mode: The Optimiser Creates an Unstable Skeleton
After many parts are cut from a sheet, the remaining web can become weak. Thin bridges vibrate, tip or distort under heat and machine motion.
Rules can preserve minimum web widths, cut sequence and tabs. Simulation can estimate structural stability for demanding processes.
Maximum packing density can reduce process stability. Geometry must survive the path by which it is manufactured, not only the final static layout.
98. Failure Mode: Reusable Leftovers Become Untraceable Scrap
The optimiser saves a remnant and credits its future value. The operator drops it in an unlabelled rack. Inventory still says it exists. The next plan assigns an order to material no one can find.
Then planners stop trusting remnant inventory and disable the feature. The mathematical model was reasonable; the information system failed.
Every credited future resource needs a custody process. Optimisation cannot assume inventory truth that operations do not maintain.
99. Failure Mode: A Near-Optimal Plan Changes Too Much
Production is already staged. A new order arrives. Reoptimising globally reduces material use by one sheet and changes thirty patterns already prepared.
Change cost should enter the repair objective. Freeze committed decisions, penalise deviations or optimise only a neighbourhood around affected orders.
The best plan before execution and the best plan during execution are different problems. Stability becomes part of optimality after humans and machines have begun acting.
100. Primary Mathematics: Cutting Optimisation Begins With Remainders
Primary students already solve sharing and remainder problems. How many 23-cm strips fit in a 100-cm length? What remains? Which combination of 30-cm and 20-cm pieces wastes less?
Rectangles introduce area and orientation. Tiling exercises show that equal area does not guarantee fit. Fractions and percentages describe utilisation.
The advanced industrial problem grows directly from these foundations. The difference is scale: thousands of combinations, machine rules and future value.
101. Secondary Mathematics: Cutting Becomes Integer Optimisation
Secondary students add simultaneous equations, inequalities, vectors, graphs, probability and algorithms. Capacity becomes Σlᵢaᵢ≤L. Demand becomes linear constraints. Integer variables represent counts.
Geometry explains rectangle placement. Probability models uncertain demand. Graphs and dynamic programming represent guillotine cut structures and knapsack pricing.
This is a powerful bridge from school Mathematics to operations research: a remainder question becomes a mixed-integer decision system once real production scale is introduced.
102. Advanced Mathematics: Cutting Stock Connects Whole Branches of Optimisation
Linear programming provides dual prices and lower bounds. Integer programming enforces discrete stock counts. Column generation handles huge variable sets. Dynamic programming solves pricing. Branch-and-price enforces integrality. Computational geometry handles nesting. Stochastic and robust optimisation handle uncertain demand.
Graph algorithms, constraint programming, metaheuristics and machine learning contribute where exact formulations become expensive or geometry is irregular.
The cutting problem is therefore an excellent study of mathematical architecture: different methods solve different layers because no one representation is efficient for every constraint.
103. Why This Improves the World
- It reduces virgin material consumption. Better patterns satisfy the same demand using fewer stock objects.
- It reduces irreversible scrap. Kerf, trim and awkward offcuts are accounted for explicitly.
- It makes remnants economically reusable. Leftovers can become tracked future stock instead of automatic waste.
- It improves factory throughput. Pattern count, machine assignment and toolpath time can be optimised beside material.
- It supports sustainability with measurement. Material, cost and carbon trade-offs can be made explicit.
- It turns hidden plant rules into formal constraints. Defects, grain, knife counts and due dates become part of the plan rather than after-the-fact exceptions.
- It provides evidence about solution quality. Lower bounds and optimality gaps show how much improvement might remain.
104. What Mathematics Does Not Do
Cutting optimisation does not make kerf disappear. It does not turn defective material into perfect material. It cannot reuse a remnant that operations failed to label. It cannot know tomorrow’s demand exactly. It cannot make a geometrically dense pattern executable on a machine whose cutting sequence forbids it.
An exact optimum is exact only for the model supplied. Wrong stock dimensions, missing setup costs, ignored tolerances or outdated machine rules can make an optimal plan operationally wrong.
And lower waste is not automatically lower total environmental impact if extra setups, energy, storage or logistics overwhelm the material saving. System boundaries remain part of the decision.
Frequently Asked Questions
What is the cutting stock problem?
It is an optimisation problem that chooses how to cut demanded smaller items from available stock objects so demand is satisfied while minimising stock use, trim loss, cost or related objectives.
What is column generation?
Column generation solves a restricted master problem containing only some variables, then uses dual prices in a pricing subproblem to generate new variables—cutting patterns—that can improve the solution.
Why is a knapsack problem used in cutting stock?
In one-dimensional pricing, the solver chooses item counts whose total length fits the stock while maximising dual value. That is a knapsack-style optimisation.
What is branch-and-price?
Branch-and-price combines branch-and-bound for integer decisions with column generation for solving the LP relaxation at search nodes.
What is guillotine cutting?
Guillotine cutting requires each cut to run completely across the current rectangular piece from one boundary to the opposite boundary, creating a hierarchical sequence of rectangular subdivisions.
What is kerf?
Kerf is the material width removed by the cutting process. It must be included in feasible pattern geometry and can materially affect tight layouts.
Why keep usable leftovers?
Large or useful remnants can satisfy future demand and reduce virgin stock consumption. Their value depends on dimensions, material, storage cost and probability of reuse.
Is maximum material utilisation always the best objective?
No. Setup count, machine time, due dates, overproduction, remnant value, carbon and handling cost can make a slightly lower-utilisation plan better overall.
Sources and Further Reading
- Senergues, Brahimi, Cherri, Klein and Péton, Cutting Stock Problem with Usable Leftovers: A Review, European Journal of Operational Research, 2026.
- Two-Dimensional Bin Packing, Cutting Stock, and Open Dimension Problems: A Survey of Practical Requirements, Computers & Operations Research, 2025.
- Gilmore and Gomory, Multistage Cutting Stock Problems of Two and More Dimensions, Operations Research, 1965, building on their foundational 1961 and 1963 column-generation work.
- Solving Real-World Cutting Stock Problems in the Paper Industry: Mathematical Approaches, Experience and Challenges, European Journal of Operational Research.
- A Branch-and-Cut-and-Price Algorithm for One-Dimensional Stock Cutting and Two-Dimensional Two-Stage Cutting, European Journal of Operational Research.
- Bin Packing and Cutting Stock Problems: Mathematical Models and Exact Algorithms, European Journal of Operational Research.
Continue Through eduKateSG
Continue with How Mathematics Works. Compare this article with Fitting More Cargo Into the Same Container Without Crushing Anything: both are packing-and-allocation problems, but cargo loading manages movable finished objects while cutting stock creates demanded parts by irreversibly transforming raw material.
Also read Keeping Supermarket Shelves Full Without Filling Warehouses With Waste. Cutting with leftovers becomes an inventory problem when saved remnants carry value into future periods.
Final Thought: Waste Often Begins Before the Blade Moves
A customer asks for parts.
The warehouse contains sheets, bars, rolls and yesterday’s remnants.
Millions of feasible patterns exist.
Most will never be used.
Dual prices reveal what demand is difficult.
A pricing problem invents a better pattern.
Kerf removes a little.
A defect removes another corner.
A leftover survives because tomorrow may need it.
The operator sees one practical cutting plan.
Behind it, Mathematics has searched a space too large to write down.
The material was always finite.
The useful combinations were not.
Mathematics improves the world here by finding better combinations before irreversible cuts turn possibility into scrap.