Termin in Kalender importieren: iCal (Download)
Vorträge
| Vortragende(r)
|
Aaron Gätje
|
| Titel
|
Graph Attention Network for Injection Molding Process Simulation
|
| Vortragstyp
|
Masterarbeit
|
| Betreuer(in)
|
Daniel Ebi
|
| Vortragsmodus
|
in Präsenz
|
| Kurzfassung
|
Graph Neural Networks (GNNs) have demonstrated great potential for simulating physical systems that can be represented as graphs. However, training GNNs presents unique challenges due to the complex nature of graph data. The focus of this thesis is to examine their learning abilities by developing a GNN-based surrogate model for the injection molding process from materials science. While numerical simulations can accurately model the mold filling with molten plastic, they are computationally expensive and require significant trial-and-error for parameter optimization.
We propose a GNN-based model that can predict the fill times and physical properties of the mold filling process. We model the mold geometry as a static graph and encode the process information into node, edge, and global features. We employ a self-attention mechanism to enhance the learning of the direction and magnitude of the fluid flow. To further enforce the physical constraints and behaviors of the process, we leverage domain knowledge to construct features and loss functions. We train our model on simulation data, using a multi-step loss to capture the temporal dependencies and enable it to iteratively predict the filling for unseen molds. Thereby, we compare our models with different distance-based heuristics and conventional machine learning models as baselines in terms of predictive performance, computational efficiency, and generalization ability. We evaluate our architectural and training choices, and discuss both the potential applications and challenges of using GNNs for surrogate modeling of injection molding.
|
| Vortragende(r)
|
Christoph Batke
|
| Titel
|
Improving SAP Document Information Extraction via Pretraining and Fine-Tuning
|
| Vortragstyp
|
Masterarbeit
|
| Betreuer(in)
|
Edouard Fouché
|
| Vortragsmodus
|
in Präsenz
|
| Kurzfassung
|
Techniques for extracting relevant information from documents have made significant progress in recent years and became a key task in the digital transformation. With deep neural networks, it became possible to process documents without specifying hard-coded extraction rules or templates for each layout. However, such models typically have a very large number of parameters. As a result, they require many annotated samples and long training times. One solution is to create a basic pretrained model using self-supervised objectives and then to fine-tune it using a smaller document-specific annotated dataset. However, implementing and controlling the pretraining and fine-tuning procedures in a multi-modal setting is challenging. In this thesis, we propose a systematic method that consists in pretraining the model on large unlabeled data and then to fine-tune it with a virtual adversarial training procedure. For the pretraining stage, we implement an unsupervised informative masking method, which improves upon standard Masked-Language Modelling (MLM). In contrast to randomly masking tokens like in MLM, our method exploits Point-Wise Mutual Information (PMI) to calculate individual masking rates based on statistical properties of the data corpus, e.g., how often certain tokens appear together on a document page. We test our algorithm in a typical business context at SAP and report an overall improvement of 1.4% on the F1-score for extracted document entities. Additionally, we show that the implemented methods improve the training speed, robustness and data-efficiency of the algorithm.
|
- Neuen Vortrag erstellen
Hinweise
