A Data-Driven Approach to Determine the Single Droplet Post-Impingement Pattern on a Dry Wall Using Statistical Machine Learning Classification Methods 2021-01-0552

The study of spray-wall interaction is of great importance to understand the dynamics during fuel-surface impingement process in modern internal combustion engines. The identification of droplet post-impingement pattern (contact, transition, non-contact) and droplet characteristics can quantitatively provide an estimation of energy transfer for spray-wall interaction, thus further influencing air-fuel mixing and emissions under combusting conditions. Theoretical criteria of single droplet post-impingement pattern on a dry wall have been experimentally and numerically studied by many researchers to quantify the hydrodynamic droplet behaviors. However, apart from model fidelity, another issue is the scalability. A theoretical criterion developed from one case might not be well suited to another scenario.

In this paper, a data-driven approach for single droplet-dry wall post-impingement pattern utilizing arithmetical machine learning classification methods is proposed and demonstrated. The droplet-wall post-impingement patterns are formulated as a classification problem. An experimental data library of for single droplet impinging on a dry wall (442 datasets from MTU inhouse experiments and 229 datasets from existing literature) is established for training and validating the classifications models. Typical parameters such as viscosity and density of the liquid droplet, temperature, Weber number, etc. that describe liquid properties and wall characteristics are discriminated against one another. Six well-known classification methods are applied to the database, and their performance is evaluated and compared. The performance of each classification method for individual post-impingement region is compared and characterized by four statistical measures (accuracy, precision, recall and F1 score) to obtain the best classifier. A high accuracy of classification methods reveals the potential of data-driven approach in determining different post impingement regions of the single droplet-wall interaction.