Background:

Previously, various types of TPMS structures were designed and 3D-printed as substrates for monolithic catalytic converters. Catalytic coating and material characterization were also performed on these additively manufactured substrates; resulting in a dataset, based on both numerical and experimental works. However, this dataset remains limited in size. The main reasons for this shortfall are the labor-intensive nature and high costs associated with conducting numerous experiments and designing complex geometries. Node.1 and Node.2 in Figure.1 are referring to this background.
Few-shot learning refers to the ability of a machine learning model to learn new tasks or categories with a very small amount of labeled data, often with just a few training examples per class. It contrasts with traditional deep learning models, which generally require large amounts of labeled data to perform well. Few-shot learning techniques are valuable in situations where data collection is costly, time-consuming, or where only limited data is available.

Objective of the work:

The aim of this master’s thesis is to identify the most suitable few-shot learning strategy based on the available datasets and to develop corresponding learning-models for conducting virtual experiments and predicting outcomes. Predicting the reaction surface of TPMS structures and thickness of the coated layers by varying the unit cell size, rotating the cells, or applying a non-uniform distribution of cells are potential outcomes of this study.

Tasks:
1. Understand the problem and choosing a suitable Few-Shot learning approach:
•    Possible approaches are transfer learning, meta learning, metric learning, generative approach, etc.
2. Data preparation including Base Dataset, Support Set and Query Set:
•    Base Datasets include TPMS designs and the corresponding reaction surfaces. These can be created within nTop software.
•    Support Sets can be also created within nTop software, by small changes in the previous TPMS designs.
•    Query Set: Calculation of the reaction surface without creating a CAD design.

3.    Building the Model:
•    Building a model capable of performing few-shot learning. Some common model architectures are Prototypical Networks, Siamese Networks and Model-Agnostic Meta-Learning (MAML).
4.    Training the Model
5.    Testing and Evaluating the Model on New Tasks

Requirements:
•    Studies in process engineering, chemical engineering, chemistry, materials science or related fields
•    Programming skills, especially in Python
•    Interest in articifial intresellince and machine learning
•    Using tools such as PyTorch or TensorFlow are recommended.
•    Knowledge of 3D printing and CAD-design are beneficial
•    Independent, creative and analytical way of working

Tasksetter: Prof. Dr. Christoph Klahn
Advisor: M.Sc. Sima Mehdipour