AI Adoption GuideManufacturingPack
3D Bin and Pallet Configuration Optimization
Combinatorial ML finds the carton and pallet configuration per order that minimizes void fill, carton count, and outbound freight weight.
Manufacturing processPlanSourceMakeInspectPackShipServiceReturn
By Don, DoneThat’s AI coach · updated
Overview
Pack-out is where carton mix, pallet pattern, and billed freight are locked. A 3D configuration optimizer treats that lock as a search problem: given the lines on an order, the cartons and pallets you actually stock, and carrier cube and weight rules, it returns a build plan. The plan is a recommendation. A packer or pack-station lead still confirms orientation, dunnage, and label placement before the carton closes or the wrap cycle starts.
The same search that cuts air in a shipper also changes carton count and stacked density, which is why freight weight and DIM weight move together. If length, width, height, or unit weight is missing for any line that must ship, the system issues an empty plan and the order stays on the exception lane. Guessed dimensions produce a plan you cannot defend at audit or at the dock.
What the solver is allowed to change
The optimizer proposes three coupled decisions: which shipper or on-demand carton each subset of lines goes into, how those units orient and nest inside the carton, and how finished cartons (or loose cases) stack on a pallet or in a gaylord. It does not change the sales order, the pick path, or the carrier account. It also does not override hazardous, orientation, or crush constraints that packaging engineering already coded as hard rules.
A useful plan is specific enough to execute without a second meeting. Each carton line should name SKU, quantity, rotation (allowed axes only), layer, and nest offset. Each pallet step should name cartons or cases, layer pattern, column versus interlocking stack, max height, and max weight. Void fill is an output of that geometry, not a separate art project. If the solver cannot place a line without violating a hard rule, that line belongs on the exception list, not in a “best effort” carton.
Related work on board grade, void material, and corner protection lives with Packaging Material Specification Optimizer. Geometry and material spec should share the same SKU master. If they diverge, the 3D plan will look dense on screen and still fail drop or clamp tests.
Inputs that must be complete before a plan is issued
Issue a plan only when every shippable line has a trusted outer carton or each dimension, unit weight, and pack factor. Inner packs and multi-packs need their own records. A each that cubes as a case will overstate density and understate carton count. Capture or refresh those values with a cubing station (Cubiscan is the usual plant-floor pattern) and treat stale dims the same as missing dims: empty plan, not a stale plan.
The carton catalog is part of the model, not a suggestion list. Include inner dimensions, flute or board that sets usable height, max fill weight, opening style, and whether Packsize or another on-demand converter may cut a custom footprint. If on-demand is in scope, the search space is continuous within machine limits. If you only stock a finite RSC set, the search is discrete. Mixing both without telling the solver which lanes can convert will print sizes the pack station cannot make.
Pallet and carrier constraints close the model: pallet footprint, max height including pallet, max weight, overhang policy, and whether the lane is parcel DIM, LTL, or FTL cube. Orientation flags (this-way-up, do not rotate, liquids upright) and nesting rules (bags in voids, no nesting for rigid cosmetics) are hard constraints. Weights missing on a kit component that ships assembled should fail the kit, not only the component.
When any required field is null, zero, or older than the plant’s refresh rule, return no carton list, no pallet layers, and no freight estimate. Downstream WMS pack screens should show a blocking reason, not a blank grid that looks like “pack however you want.”
How carton and pallet search actually runs
Practitioners do not need the textbook name of the heuristic. They do need to know the objective and the stop rule. Typical objectives, in the order most outbound plants actually care about: (1) feasible placement of every line, (2) fewest cartons, (3) lowest residual void that still meets protection rules, (4) lowest rated freight using the carrier’s cube or DIM formula, (5) stable pallet (low center of gravity, interlocking where required). Weight those in engineering, not at the pack station. If the packer can pick “tighter” versus “fewer boxes” per order, you will never get a clean before/after.
Search is combinatorial because SKU assignment to cartons and 3D orientation explode together. Production systems therefore bound the tree: pre-group by family or temperature, fix large items first, try the current carton set before opening a custom cut, and stop when a time budget or a “good enough” void threshold is met. Manhattan Associates and other WMS cartonization engines already do a version of this for discrete cartons. A 3D overlay adds rotation, nesting, and pallet layers on top of that assignment. Do not run two independent cartonizers on the same order. One owner, one plan ID.
On-demand converters (Packsize and peers) change the last step: after assignment, the machine cuts length and width to the packed cloud plus flap rules. The optimizer must output a cut file the converter accepts, and a human-readable layer map the packer can follow if the machine is down. If the converter is offline, fall back to the discrete carton catalog with the same constraint set, or hold the order. Silent substitution of a larger stock carton without a new plan ID breaks freight audit.
Palletization is a second 3D bin, not a pretty picture of the cartons. Column stacks are faster to build and worse for clamp and transit. Interlocking patterns cost seconds and save claims. Encode that trade in the objective by lane: parcel pallets that break down at the carrier hub can stay column-stacked; customer-ready retail pallets usually cannot.
Pack-out confirmation, exceptions, and damage feedback
The operator confirms the pack-out. Confirmation means the station records plan ID, actual carton or cut size used, dunnage type, catch weight if the lane requires it, and pallet ID. If the packer rotates a carton 90 degrees to close the flaps, that is a plan deviation. Capture it. Unlogged deviation is how “the model said 12% void” and “the trailer was half air” coexist.
Empty plan and hard-block when dims or weights are missing. Soft-block when the plan is feasible but void, height, or weight sits outside the band packaging engineering set for that lane. Soft-block still requires a named override (lead login) and a reason code. Never auto-complete pack on a 3D plan. Vision can watch the result. Vision-Based Label and Contents Verification is the right partner for “did the right SKUs go into the carton the plan named,” not a substitute for geometry.
When a shipment is damaged, feed the claim back to the configuration that shipped, not only to the SKU. Shipment Damage Risk Scoring needs the same plan ID, stack pattern, and void residual. Without that join, you will blame board grade for a column stack that should have been interlocking, or blame the solver for a packer who added a carton the plan never issued.
Systems, vendors, and the metrics that prove it
Cubiscan (or equivalent) owns the dim and weight master and the recube cadence. Packsize (or equivalent) owns custom footprint when the lane is converted. Manhattan Associates (or the WMS you already run) owns order, inventory, pack confirmation, and freight rating inputs. The 3D optimizer can sit as a service those systems call at pack allocation. It should not become a third carton master. Publish one carton catalog, one pallet profile set, and one constraint table.
Instrument the loop with operational counts, not marketing percentages. Track share of orders that received a non-empty plan, share blocked for missing dims, pack-station cycle time versus the previous standard pack, cartons per order, residual void estimated from the confirmed build (inner cube minus item cube), billed weight versus physical weight, and damage claims tagged with plan ID. Compare lanes with the solver on against matched lanes still on tribal pack, same SKU mix and carrier. If cycle time rises more than void and carton count fall, the layer map is too hard to build. Simplify patterns before you tighten the objective.
A configuration optimizer earns its keep when the pack station can execute the drawing, the WMS can rate the true cube, and missing master data cannot invent a box. Keep the operator in the loop, keep the plan empty until the tape measure and scale have spoken, and keep material spec, damage scoring, and contents verification on the same order and plan identifiers.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first