Yield Maximization In Injection Molding by the Virtual Search Method Dongzhe Yang, Graduate Research Assistant David Hatch, Graduate Research Assistant David Kazmer, Assistant Professor Kourosh Danai, Professor Abstract The Virtual Search Method (VSM) is an efficient method of tuning for injection molding. The salient feature of this method is its utilization of an input-output (I-O) model as a virtual process to search for the process inputs. VSM uses learning to update the I-O model after each tuning iteration so as to improve its representation of the process. The VSM has already been tested experimentally for regulation of part dimensional and qualitative attributes. This paper focuses on extension of VSM to improving the quality of the part, where it can be used for maximization of production yield and molded part consistency. VSM's performance with two I-O models is investigated using production of optical media with six input parameters and four quality attributes. Introduction The need for minimal assembly times, reduced tooling costs, and ease of recycling demands mechanical systems comprised of fewer and more complex components with increased functional requirements. This has necessitated injection molded parts with tighter tolerances and superior finish, which can only be achieved by more accurate control of the process [1, 2]. Towards this objective, significant progress has been made in improving the stability of the process on-line [3-6], but relatively little attention has been paid to efficient specification of setpoints for various machine inputs such as melt and mold temperatures, injection pressure, and injection speed. A typical commercial component is molded with the goal of minimizing material and processing costs subject to ten critical dimensional tolerances plus five additional surface and structural requirements. The molding process is typically over-constrained, with selection of the operating setpoints affecting the resulting part attributes in a non- deterministic manner. The traditional approach to machine input selection (tuning) in the plastics industry has been ‘trial and error.’ For this, shots are taken during start-up and part quality attributes are measured after each shot to evaluate the acceptability of produced parts. The process engineer then uses his/her knowledge of the process to select the machine inputs in such a way as to improve the quality of the part from shot to shot. This tuning exercise is repeated until the specifications for part quality are satisfied. The main drawback of the traditional tuning approach is its inefficiency due to its ‘ad hoc’ nature. Humans usually use linear relationships to relate machine inputs to quality attributes, so they often have difficulty adjusting the inputs over large ranges [7]. They also tend to treat the various attributes as independent, thus, ignore the couplings among the attributes. These human conceptual models and the added difficulty of coping with noise often lead to time- consuming tuning sessions and considerable waste Commercial molders have realized the economic impact of limitations in their operational ability, and are purchasing auxiliary quality systems. One alternative to the traditional trial and error approach has been the use of expert systems. Trouble-shooting expert systems, which have attracted considerable attention in recent years, represent corrective guidelines in the form of ‘if-then’ rules [8-10] so they have the appeal of replacing the human expert in providing trouble-shooting knowledge. However, expert systems have not yet proven successful for injection molding since a generalized set of rules may not be appropriate across a broad range of part geometries, material properties, and machine dynamics. Furthermore, expert systems are limited by their non-quantitative nature and their inability to cope with quality issues not addressed by the instantiated rules. The current practice for tuning injection molding machines in large production operations is to develop an empirical model based on data obtained from a set of designed experiments [11]. Based on this model, the objective function of an unconstrained optimization problem is defined as a function of the part quality attributes, and the set of inputs that produce the best quality attributes are obtained as the ‘optimal’ point of this optimization problem. While Design of Experiments (DOE)-based methods offer a systematic approach to tuning that can also be used for mold qualification [12-15], they require significant investment in training and technology while consuming valuable production time in developing global quality models. Recently, the Virtual Search Method (VSM) of tuning has been introduced that incorporates aspects of both the trial and error approach and DOE [16]. The block diagram of VSM is shown in Figure 1. It consists of a ‘search algorithm’ that determines prospective changes to the