Reflow profiling with the aid of machine learning models

Abstract

Purpose This paper aims to propose a method to quickly set the heating zone temperatures and conveyor speed of the reflow oven. This novel approach intensely eases the trial and error in reflow profiling and is especially helpful when reflowing thick printed circuit boards (PCBs) with bulky components. Machine learning (ML) models can reduce the time required for profiling from at least half a day of trial and error to just 1 h. Design/methodology/approach A highly compact computational fluid dynamics (CFD) model was used to simulate the reflow process, exhibiting an error rate of less than 1.5%. Validated models were used to generate data for training regression models. By leveraging a set of experiment results, the unknown input factors (i.e. the heat capacities of the bulkiest component and PCB) can be determined inversely. The trained Gaussian process regression models are then used to perform virtual reflow optimization while allowing a 4°C tolerance for peak temperatures. Upon ensuring that the profiles are inside the safe zone, the corresponding reflow recipes can be implemented to set up the reflow oven. Findings ML algorithms can be used to interpolate sparse data and provide speedy responses to simulate the reflow profile. This proposed approach can effectively address optimization problems involving multiple factors. Practical implications The methodology used in this study can considerably reduce labor costs and time consumption associated with reflow profiling, which presently relies heavily on individual experience and skill. With the user interface and regression models used in this approach, reflow profiles can be swiftly simulated, facilitating iterative experiments and numerical modeling with great effectiveness. Smart reflow profiling has the potential to enhance quality control and increase throughput. Originality/value In this study, the employment of the ultimate compact CFD model eliminates the constraint of components’ configuration, as effective heat capacities are able to determine the temperature profiles of the component and PCB. The temperature profiles generated by the regression models are time-sequenced and in the same format as the CFD results. This approach considerably reduces the cost associated with training data, which is often a major challenge in the development of ML models.