About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Low-Rank Tensor Completion for Multi-Property Prediction of Polymers |
| Author(s) |
Andy Ramlatchan, Brandon McVade |
| On-Site Speaker (Planned) |
Andy Ramlatchan |
| Abstract Scope |
Accurate prediction of polymer properties across large, sparsely characterized chemical spaces remains a central challenge in polymer informatics. Existing machine learning approaches that adopt single-task frameworks ignore the cross-property correlations that arise from shared physicochemical mechanisms. We propose a tensor completion framework for simultaneous multi-property prediction of polymers. We construct a third-order tensor indexed by polymer identity, property type, and measurement conditions, and demonstrate that the inherent low-rank structure of polymer property space can be exploited to recover missing entries with high fidelity, also generalizable to higher dimensions. We then compare across CP and Tucker decomposition, Bayesian Probabilistic Tensor Factorization (BPTF), and a GNN-initialized approach. Our results show that tensor completion outperforms single-task baselines under data scarcity, and that the extracted latent factors correspond to interpretable physicochemical descriptors including chain stiffness, polarity, and free volume. This work establishes the first tensor completion benchmark for scalable and interpretable polymer property prediction. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Process Technology, Composites, Machine Learning |