A soft-computing framework for automated optimization of multiple product quality criteria with application to micro-fluidic chip production

Abstract

We describe a general strategy for optimizing the quality of products of industrial batch processes that comprise multiple production stages. We focus on the particularities of applying this strategy in the field of micro-fluidic chip production. Our approach is based on three interacting components: (i) a new hybrid design of experiments (DoE) strategy that combines expert-and distribution-based space exploration with model-based uncertainty criteria to obtain a representative set of initial samples (i.e., settings of essential machining process parameters), (ii) construction of linear and non-linear predictive mappings from these samples to describe the relation between machining process parameters and resulting quality control (QC) values and (iii) incorporation of these mappings as surrogate fitness estimators into a multi-objective optimization process to discover settings that outperform those routinely used by operators. These optimized settings lead to final products with better quality and/or higher functionality for the clients. The optimization module employs a co-evolutionary strategy we developed that is able to deliver better Pareto non-dominated solutions than the renowned NSGA-II multi-objective solver. We applied our proposed highlevel surrogate-based multi-objective strategy both in a single/late-stage optimization scenario and in a more challenging multi-stage scenario, yielding final optimization results that improved parameter settings and thus product quality compared to standard expert-based production process parameterizations.

Lightweight Interpolation-Based Surrogate Modelling for Multi-objective Continuous Optimisation

“…Multi-objective evolutionary algorithms (MOEAs) have emerged as very popular MOOP solvers due to their ability to discover high-quality PS approximations called Pareto non-dominated sets (PNs) after single optimisation runs [1,19]. The successful application of MOEAs to increasingly complex industrial MOOPs ranging from product design [10] to calibration [6] and quality assurance [16] has also helped to highlight that when the process of evaluating F (x) is computationally-intensive 1 , the effectiveness of the solver can be severely impacted as far fewer candidate solutions/individuals x ∈ V d can be evaluated during the optimisation run. One of the most promising approaches for alleviating the effect of expensive F (x) formulations is to replace the original fitness functions with an easy to evaluate surrogate formulation [9,11].…”