A Machine Learning Approach in Pendant Drop Tensiometry Using Image Moments

One of the most recent emerging technique for Surface Tension caluclation is called Pendant Drop Tensiometry which is a Axisymmetric Drop Shape Analysis. Machine Learning and Neural Networks are another fields which are emerging and could be commonly used for any kind of topic. Although Pendant Drop Tensiometry and Machine Learning are inevitably improving fields separately, there are not many works previously in which both fields are merged to solve a problem. We have followed to main works and first main work helped this thesis to extract and understand the usage of differential equations of Pendant Drop Tensiometry. Other recent work could be interpreted as the only recent work which could be useful to create the hypothesis of this master thesis since it combined a Machine Learning approach with Pendant Drop Tensiometry. This Machine Learning approach consist of a Neural Network approach to predict indirectly Surface Tension via measuring non-dimensional gravitational control parameter and non-dimensional apex pressure. These two non-dimensional parameters are the key parameters and predicted parameters by the Neural Network which leads to the computation of Surface Tension. This was an indirect approach that is also the approach in this thesis too. Before applying a Machine Learning approach to Pendant Drop Tensiometry, Young-Laplace equation was written in terms of differential equations and then, they are solved iteratively together with image analysis techniques using digital images captured from a camera. This thesis builds a bridge between Pendant Drop Tensiometry and Machine Learning once but the aim is to present an innovative and much simpler solution since the previous Machine Learning approach was quite complex and computationally demanding. The main objective of this thesis is to provide a Machine Learning approach in Pendant Drop Tensiometry using Image Moments in order to predict Surface Tension of pendant drops. This Machine Learning approach based on a Neural Network architecture to predict non-dimensional parameters according to regions which are defined according to validity of non-dimensional parameters. This validity of non-dimensional parameters are considered according to shape factor of a pendant drop since very most likely pendant drop shapes have a shape factor equal to 2 or 3. These valid regions are defined with calculated curves in a graph. After extracting non-dimensional parameter values from valid regions, these values are used to create synthetic pendant drop shapes. Image Moments takes the role after this point and image moments of synthetically generated drop images are calculated to feed Neural Network architecture. Image Moments are the innovation because they are scale, translation and rotation invariant to create a robust networks to make the algorithm work for also rotated, scaled and translated pendant drop image versions. Non-dimensional parameters are calculated individually in separate Neural Networks and corresponding Mean Square errors were approximately 0.021. Previous Machine Learning approach was trained for 3 weeks but this thesis only took hours for measuring Surface Tension which is the proof that, the aim, which is to create simpler model using Image Moments, was achieved consequently.