This thesis is in cooperation with one of our industry partners, where you will dive deeper into material development for high-performance elastomers, which is time-intensive and costly, relying on iterative trial-and-error approaches. Foundation Models — AI systems pre-trained on large-scale datasets — have demonstrated breakthrough potential in materials science for property prediction, inverse design, and accelerated discovery. Recent advances include transformer-based models for polymer property prediction (TransPolymer, PolyBERT), multimodal architectures integrating molecular structure with material properties, and graph-based representations for geometric molecular data. Despite this progress, industrial elastomer development remains underexplored, facing challenges such as limited proprietary data, multimodal integration complexity, and insufficient experimental validation.

Research Objectives
This thesis investigates the application of foundation models to automate and accelerate elastomer material development and selection. The research aims to:

• Systematically evaluate foundation model architectures for elastomer property prediction
• Develop a framework integrating multimodal data (molecular structure, process parameters, material properties)
• Implement and benchmark models on elastomer-relevant datasets
• Design transfer learning strategies for industrial low-data scenarios
• Validate prediction quality and out-of-distribution robustness
• Evaluation and Interpretability:
  • Benchmarking against traditional ML baselines and domain expert knowledge
  • Out-of-distribution robustness analysis
  • Interpretability analysis to extract chemical and physical insights from learned representations

Expected Contributions
Scientific Contributions:

• Systematic comparison of foundation model architectures for elastomer property prediction
• Novel multimodal integration framework combining molecular structure, filler, and process data
• Analysis of transfer learning strategies for low-data industrial materials scenarios
• Interpretability study linking learned latent representations to physicochemical properties

Practical Contributions:

• Open Python/PyTorch framework for multimodal elastomer property prediction
• Benchmark study guiding model selection for materials informatics applications
• Transfer learning pipeline ready for industrial deployment
• Guidelines for integrating foundation models into elastomer development workflows

Requirements

• Strong background in machine learning and deep learning
• Programming proficiency in Python
• Experience with transformer architectures and/or graph neural networks
•  

Literature

1. Pyzer-Knapp, E. O., Manica, M., Staar, P., Morin, L., Ruch, P., Laino, T., Smith, J. R., & Curioni, A. (2025). Foundation models for materials discovery – current state and future directions. npj Computational Materials, 11, 61. https://doi.org/10.1038/s41524-025-01538-0
2. Moro, V., Loh, C., Dangovski, R., Ghorashi, A., Ma, A., Chen, Z., Kim, S., Lu, P. Y., Christensen, T., & Soljačić, M. (2025). Multimodal foundation models for material property prediction and discovery. Newton (Cell Press). https://doi.org/10.1016/j.newton.2025.100016
3. Xu, C., Wang, Y., & Barati Farimani, A. (2023). TransPolymer: a Transformer-based language model for polymer property predictions. npj Computational Materials, 9, 64. https://doi.org/10.1038/s41524-023-01016-5
4. Soares, E. A., et al. (2024). MOL-MOE: Multi-View Mixture-of-Experts for predicting molecular properties using SMILES, SELFIES, and graph-based representations. NeurIPS 2024 Workshop on AI for Accelerated Materials Design (AI4Mat). IBM Research / FM4M.
5. Zhang, T., & Yang, D.-B. (2025). Multimodal machine learning with large language embedding model for polymer property prediction. Chemistry of Materials, 37(18), 7002–7013. https://doi.org/10.1021/acs.chemmater.5c00940