PolUQBench: A Benchmark Study on Uncertainty Quantification of Polymer Property Prediction

Alif Bin Abdul Qayyum, Byung-Jun Yoon

Keywords: Uncertainty quantification, polymer property prediction, polymer language model

Abstract: Large Language Model(LLM)s have demonstrated remarkable capabilities to tackle multidomain challenges, a capability often lacking in conventional machine learning methods. This makes them particularly promising for understanding the complex relationship between a material's composition and its properties, which can significantly accelerate materials design, especially for polymers. Leveraging the hidden states of domain-specific pretrained LLMs for downstream tasks like property prediction has gained significant traction. This approach is now widely used for small molecules and proteins, along with recent efforts also extending to polymers. In addition to achieving superior predictive performance, Uncertainty Quantification (UQ) is another crucial aspect for enhancing the reliability of machine learning models used as property predictors. This is particularly important for high-stakes applications like the discovery of new functional polymers. We introduce **Pol**ymer Property Predictor **U**ncertainty **Q**uantification **Bench**mark, a pioneering study that evaluates the effectiveness of embeddings extracted from a Polymer Language Model for representing polymer data and assesses the performance of several different UQ methods for reliable polymer property prediction.

Submission Track: Benchmarking in AI for Materials Design - Short Paper

Submission Category: Automated Material Characterization

Institution Location: College Station, TX, USA

AI4Mat Journal Track: Yes

AI4Mat RLSF: Yes

Submission Number: 90