Exploration of generative models for the generation and validation of synthetic data in the context of material testing (Masterarbeit)

Kurzfassung

There are significant interests in aerospace industry to accelerate material development using machine learning models by identifying patterns and relationships from the material testing. To achieve this, the testing pyramid (i.e. building block approach) needs to be reshaped to correlate the behaviour at larger scaled structures to a small number of fundamental material properties by analysing data on existing materials. However, creating the input dataset through material testing is costly and time consuming, which results in a data scarcity. This shortage of data poses significant challenges in training machine learning models, as limited data may lead to overfitting and poor generalization. Data augmentation and synthetic data generation techniques are commonly employed to alleviate these issues by creating synthetic data. The data generation methods need to meet the criteria that the generated data needs to be similar to the available existing data, i.e., it needs to be realistic. The material test data are time series in nature and generated data should capture the time series nature along with the crucial patterns reflecting the material properties. An evaluation method need to be proposed to evaluate the model and the generated synthetic data, which can capture the criterions of the generated data to be realistic by means of a metric. The evaluation metric should judge the performance of the model. Also, an acceptability criteria to judge the acceptability of the generated data has to be designed. The synthetic data satisfying the acceptance criteria would be used for the input dataset for machine learning. The aim of the thesis is to explore models to generate synthetic material test data and propose an evaluation criteria for assessment of the generative models and the generated data.