Inventory Costs Reduced Through Optimization
Manufacturing operations that cut parts from raw material stock face two compounding cost pressures. The first is inventory: material must be on hand when production needs it, but carrying more than necessary ties up working capital, occupies floor space, and creates exposure to obsolescence. The second is yield: even when the right material is available, how it is cut determines how much becomes finished part and how much becomes scrap.
The pattern is common in metal fabrication. A manufacturer cutting precision components from flat stock — sheet and plate across a range of grades, gauges, and dimensions — needs material available when production requires it. Purchasing decisions often evolve around informal rules: reorder when stock falls below a fixed level, order in standard quantities, maintain safety stock multiples set years earlier and never revisited. None of these rules account for current demand patterns, the variability in order volumes, or the actual behaviour of supplier lead times.
The result is a predictable imbalance. Months of inventory accumulate for some materials while production halts periodically to wait for stock that should have been ordered weeks earlier. Demand variability is the underlying driver — requirements for certain components are irregular and difficult to anticipate, making simple threshold-based rules an unreliable proxy for actual need. The cost runs in both directions: excess stock on one side, production stoppages on the other.
Material waste compounds the problem. Without a forward view of consolidated requirements, cuts are planned job by job rather than against the full demand picture. The nesting of parts onto sheet and plate is uncoordinated, offcut rates are high, and remnant material accumulates without a reliable path back into productive use.
We began with a full characterisation of the demand signal across the component range. Historical order data was decomposed by part family, customer, and order pattern to understand the structure of demand variability. Some components had stable, forecastable demand. Others were intermittent and volatile — a pattern common where orders arrive in project-driven batches rather than as continuous consumption. The forecasting approach had to account for both.
For the forecastable portion of the range, we trained a neural network demand forecasting model on the full order history, capturing seasonality, trend, and the interdependencies between related part families that simpler models cannot represent. Neural networks were the right tool here because demand across component families moved together in ways that reflected shared customers and shared production cycles — patterns a conventional univariate model would miss entirely. The model produces SKU-level demand forecasts across a rolling horizon, updated continuously as new orders arrive and production schedules are confirmed.
Lead time variability was modelled explicitly rather than approximated with a fixed assumption. We extracted the full distribution of historical delivery performance by supplier and material type, and built lead time models that captured how that variability shifted with order quantity, material grade, and time of year. Reorder points are now calculated dynamically — set not to a single expected lead time but to a percentile of the lead time distribution that delivers the business's target service level. As supplier performance shifts, reorder triggers adjust accordingly without manual intervention.
Cutting optimisation was applied to the material requirement plan produced by the forecasting system. Given the forward demand schedule and current stock positions, a nesting algorithm determines how to cut the required parts from available sheets and plates in the combination that minimises total offcut waste across the consolidated requirement. The algorithm respects material grade and thickness constraints, accounts for the reusability of remnant pieces for subsequent cuts, and sequences the cut plan to align with production priorities.
To validate the integrated system before any change was made to live purchasing, we built a discrete event simulation of the full plant. The simulation modelled inbound material deliveries, stock movements, machine utilisation, cutting schedules, and production demand — driven by the historical order data and the modelled lead time distributions. Running the simulation across thousands of scenarios allowed us to tune inventory policy parameters against a realistic representation of the plant's operating environment before committing to live deployment.
The result is a shift from static safety stock rules to a dynamic, model-driven inventory policy. By aligning reorder points to neural network demand forecasts and statistically modelled lead time distributions, the business eliminates the systematic over-purchasing that builds excess inventory across the range. Working capital tied up in raw material falls, and the floor space required to hold it contracts accordingly.
Material yield improves through the application of cutting optimisation. Planning cuts against a consolidated demand horizon rather than individual job requirements gives the nesting algorithm the scope to achieve higher sheet utilisation consistently. Offcut generation falls, remnant inventory decreases, and the material cost per finished component improves across the range.
The plant simulation continues to earn its place after deployment. As the product mix evolves and supplier conditions shift, it provides a testing environment for policy changes — giving the operations team a reliable way to validate inventory decisions before they are applied to the production floor.
Facing a similar problem? Start with a conversation.