Efficient Training of Graph Neural Networks for Dynamic Phenomena (Proposal): Unterschied zwischen den Versionen
Keine Bearbeitungszusammenfassung |
Keine Bearbeitungszusammenfassung |
||
| Zeile 7: | Zeile 7: | ||
|vortragsmodus=in Präsenz | |vortragsmodus=in Präsenz | ||
|kurzfassung=Graph Neural Networks (GNNs) have shown great potential for use cases that can be described as graphs. However, training GNNs presents unique challenges due to the characteristics 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 model the mold filling accurately, they are computationally expensive and require significant trial-and-error for parameter optimization. | |kurzfassung=Graph Neural Networks (GNNs) have shown great potential for use cases that can be described as graphs. However, training GNNs presents unique challenges due to the characteristics 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 model the mold filling accurately, they are computationally expensive and require significant trial-and-error for parameter optimization. | ||
We propose representing the mold geometry as a static graph and constructing additional node and edge features from domain knowledge. We plan to enhance our model with a self-attention mechanism, allowing dynamic weighting of a node's neighbors based on their current states. Further improvements may come from customizing the model’s message passing function and exploring node sampling methods to reduce computational complexity. We compare our approach to conventional machine learning models w.r.t. predictive performance, | We propose representing the mold geometry as a static graph and constructing additional node and edge features from domain knowledge. We plan to enhance our model with a self-attention mechanism, allowing dynamic weighting of a node's neighbors based on their current states. Further improvements may come from customizing the model’s message passing function and exploring node sampling methods to reduce computational complexity. We compare our approach to conventional machine learning models w.r.t. predictive performance, generalizability to arbitrary mold geometries and computational efficiency. | ||
This thesis is a follow-up work to a bachelor thesis written at our chair in 2022. | This thesis is a follow-up work to a bachelor thesis written at our chair in 2022. | ||
}} | }} | ||
Version vom 27. März 2023, 17:03 Uhr
| Vortragende(r) | Aaron Gätje | |
|---|---|---|
| Vortragstyp | Proposal | |
| Betreuer(in) | Daniel Ebi | |
| Termin | Fr 31. März 2023 | |
| Vortragsmodus | in Präsenz | |
| Kurzfassung | Graph Neural Networks (GNNs) have shown great potential for use cases that can be described as graphs. However, training GNNs presents unique challenges due to the characteristics 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 model the mold filling accurately, they are computationally expensive and require significant trial-and-error for parameter optimization.
We propose representing the mold geometry as a static graph and constructing additional node and edge features from domain knowledge. We plan to enhance our model with a self-attention mechanism, allowing dynamic weighting of a node's neighbors based on their current states. Further improvements may come from customizing the model’s message passing function and exploring node sampling methods to reduce computational complexity. We compare our approach to conventional machine learning models w.r.t. predictive performance, generalizability to arbitrary mold geometries and computational efficiency. This thesis is a follow-up work to a bachelor thesis written at our chair in 2022. | |