Wafer-fab capacity planning is anchored by two hard physical facts: a typical VLSI process requires 100-500 operation steps over many weeks of cycle time, and reentry, lot-splitting, and batch operations make tool-level load anything but linear [S1]. The result is that strategic planning (years, fabs, technologies) and operational lot release (days, tools, shifts) are solved with two different mathematical toolkits, and the handover between them is where most planning systems either earn their keep or fall apart.
The strategic layer is dominated by multistage stochastic programming. A reference formulation for a major US manufacturer solves capacity allocation across 5 fab facilities and 29 technology categories simultaneously, minimizing the gap between product demand and the capacity assigned to each technology, with demand and capacity treated as separate stochastic processes [S2]. That same separation — demand uncertainty versus capacity-estimation uncertainty — is the structural feature every stochastic capacity model in the field carries, including scenario-tree variants that range from fully independent period-to-period states to arbitrary multi-period paths the planner specifies [S2].
Two-Model Stack: Strategic Stochastic Planning vs Operational Lot Release
Strategic planning runs on multi-year stochastic programs in which scenarios are paths from root to leaf in a scenario tree, with the objective stated as minimizing demand-capacity gaps across technology categories and fab facilities [S2]. The two canonical scenario constructs are an independent model with no prior information and an arbitrary model in which the planner injects sampled multi-period paths, and benchmarks show the independent formulation tolerates aggressive scenario aggregation without losing solution quality [S2].
Operational lot release, in contrast, runs on shorter horizons and must respect reentry, batch, and maximum-lot-wait constraints. A pull-philosophy Capacity Planning System (CPS) for multiple fabs decomposes the problem into three modules: a WIP-Pulling Module (WPM) that pulls work from the end of the process route to meet the master production schedule, a Workload Accumulation Module (WAM) that calculates expected equipment loading in time buckets, and a Wafer Release Module (WRM) that selects the lot release time, the start fab, and the equipment capability when WIP cannot satisfy the MPS [S1]. Simulation benchmarks within that study rank combinations of Adjusted Release Time (ART) with Path Load as the strongest performer across three workload-balance measures [S1].
Uncertainty Sources and Why Tool Capacity Is Treated Separately
Semiconductor demand is short-cycle and lumpy, while capital expansion is multi-year and irreversible, and that mismatch is the single largest reason capacity planning is modelled as a stochastic program rather than a deterministic LP [S3]. The same source frames the problem as a scenario-based stochastic programme in which market uncertainty, manufacturing-system uncertainty, short product life cycles, and expensive capital investment all hit at once [S3].
Tool-level capacity is a further, distinct layer because tool sets are special-purpose, capital-intensive, and acquired in batches that cannot be reversed inside a planning horizon [S4]. Research on tool capacity planning under demand uncertainty treats tool acquisition as a discrete allocation decision with multi-million-dollar annual budgets, where each tool type is special-purpose and substitution across types is limited [S4]. A separate stream studies the cross-functional coordination problem in which marketing managers reserve capacity from manufacturing and manufacturing managers allocate it back, with operating cost as the manufacturing objective and profit as the marketing objective [S6]. The coordination failure mode is well documented: without internal pricing or reservation signals, marketing over-reserves and manufacturing over-buys, and decentralization alone increases operating cost relative to the centralized optimum [S6].
Decision Criteria: Push vs Pull, and CPS Performance Metrics
The push-versus-pull comparison in multiproduct, multiline, multistage fab simulations is one of the few cases with a clean rule: machine utilization varies much less in the pull system than in the push system, and the pull system is more balanced overall [S1]. That balance is measured along three axes in the CPS benchmark — equipment workload among fabs, equipment workload across days, and equipment workload across equipment groups at different demand levels [S1]. The same three axes are exactly what an operational planner sees on a daily WAM output, which is why ART + Path Load is reported as the top combination on those three measures [S1].
Implementation: How the Math Lands in an Enterprise Planning System
The commercial instantiation of this stack in Oracle JD Edwards EnterpriseOne's Capacity Planning module follows the same decomposition: Resource Requirements Planning (RRP) and Rough Cut Capacity Planning (RCCP) feed into Capacity Requirements Planning (CRP), each producing a load-versus-available view against the master production schedule (MPS) or MRP output [S5]. When the plan shows insufficient capacity the documented choice is to either alter the plan or alter the capacity, which is the same trade-off the stochastic programme encodes at a higher resolution [S5].
For the operational layer the WPM-to-WAM-to-WRM pipeline mirrors the MPS-to-RRP-to-CRP flow: pull work backward from the customer due date, accumulate the load in time buckets at the equipment level, then re-allocate the gap by changing release time, start fab, or equipment capability [S1]. In practice this is where constraints such as reentry operations, batch tool minimums, and maximum lot waiting time are enforced, and any stochastic-programming output that ignores them will not survive contact with the shop floor [S1].
Selection Criteria: Who Uses Which Layer
A fab planning team should pick its modelling layer by horizon and decision type, not by software availability. If the question is "how much capacity do we need, and where, over the next 1-3 years," use multistage stochastic programming with explicit scenario trees over 29-or-fewer technology categories and 2-5 fab facilities, and benchmark the independent-scenario formulation against an arbitrary-path version to confirm aggregation tolerance [S2]. If the question is "which tool type and how many, under demand uncertainty," use a discrete tool-allocation model with explicit tool-type substitution limits, since tool budgets are multi-million-dollar commitments that cannot be unwound inside the planning horizon [S4].
If the question is "should marketing or manufacturing own the capacity decision," the published evidence is that a centralized stochastic programme achieves a lower operating cost than a decentralized scheme without internal coordination, so for a single-firm scope the recommended approach is centralized planning with shared scenario inputs from both sides [S6]. If the question is "how do I release tomorrow's lots across three fabs," the pull-philosophy CPS with ART + Path Load is the documented best combination across three workload-balance measures, and the WPM/WAM/WRM decomposition maps directly to the enterprise RRP/RCCP/CRP output [S1].
Limitations and Failure Modes
Three failure modes recur across the literature. First, stochastic programmes are only as good as their scenario generation: an independent-scenario model that aggregates aggressively can lose dependency information between periods, and only the arbitrary-path construct can replay specific multi-period scenarios with full prior information [S2]. Second, the push system is more sensitive to demand-level changes than the pull system, and at high utilization levels no single policy combination dominates on all performance measures, which is why pair-wise comparisons like ART vs Path Load have to be reported per metric rather than as a single number [S1]. Third, decentralized coordination without an internal market mechanism systematically allocates more capacity than the centralized optimum, because marketing's profit objective and manufacturing's operating-cost objective are not aligned by default [S6].
For fabs that operate as part of a larger industrial process chain — including the upstream memory and DRAM tool flow that feeds packaging and test — these planning layers also have to interface with adjacent capacity problems in the supply chain, which is why a HBM and DRAM capacity chain spec map and an HBM memory structural supply squeeze become relevant context for any wafer-allocation decision in 2026. For the data-plane side of the same fab, edge computing gateways and the industrial Ethernet protocol mix determine how WAM-equivalent load data actually reaches the planning system in real time.
Trackable signals for the next planning cycle: (1) any vendor release note from Oracle EnterpriseOne that revises the CRP time-bucket granularity or the RRP-to-RCCP handoff logic, since that is the operational surface most fab planners touch daily [S5]; (2) new stochastic-programming benchmarks that report solution quality versus scenario-aggregation level on the same 29-technology, 5-fab reference case, because that case is the closest thing the field has to a shared benchmark [S2].
Spec-level background on the components involved: pressure transmitter, flow meter, and industrial valve.