Modeling and decision-making in a semiconductor supply chain.

摘要

This work analyzes supply chain dynamics, specifically that of the semiconductor industry. Very high capital investments, uncertainties in demand and supply, short product life cycle, long lead times and short order to delivery times are characteristics of such supply networks.; Flow of material across the supply network is modeled as a linear programming problem (LP) with the decision variables controlling various release points of this network as output. Manufacturing starts well before the orders are placed for end products. Hence, to make up for the difference in supply and demand for end products, inventories are maintained at various key points of the network. An LP is executed with the latest available information on the inventory levels and demands for various end products. The decisions taken by the LP to control each day's supply are implemented. The LP is re-executed after accounting for a previous day's supply and demand and the new demand forecast for the following days. This process is repeated on a daily basis to adjust the supply to demand in the shortest possible time.; Such LP models are good for handling manufacturing facilities with constant throughput time (TPT). However, in a semiconductor supply network, the TPT of a lot starting on a given day depends on the total amount of that day's factory starts. A typical approximation of the nonlinear relationship between TPT and starts is given by a step function. The resulting mixed-integer programming problem becomes far too big to be solved by standard methods. This dissertation develops a hybrid method, combining the heuristics of a genetic algorithm (GA) with a linear programming approach. The GA determines a set of bounds for the allowable starts over the its time horizon. In doing so, the LP acts as a measure of fitness for the GA. This hybrid GA-LP algorithm was tested on several sample problems and its performance was compared with a best-fit LP algorithm. The hybrid algorithm captured the nonlinearity of the TPT much better than the LP algorithm and generated quantitatively correct schedule.