failed_experiment.db
← Back to browse

Materials Science› Materials Chemistry› Machine Learning in Materials Science

Polymer informatics — Fine-tuned general-purpose LLMs (Llama-3-8B, GPT-3.5) for polymer thermal property prediction

Submitted 2026-10-22

Field: Materials Science › Materials ChemistryType: A method hit a wallArea: polymer informaticsTarget: Fine-tuned general-purpose LLMs (Llama-3-8B, GPT-3.5) for polymer thermal property predictionStage: full study

Hypothesis

Fine-tuning general-purpose language models on text descriptions of polymers could match fingerprint-based methods for predicting glass transition, melting and decomposition temperatures, while skipping hand-built representations.

What failed

On a curated set of 11,740 entries, both fine-tuned LLMs generally fell short of Polymer Genome, polyGNN and polyBERT in accuracy and were less computationally efficient; multi-task training did not help the LLMs, which failed to exploit correlations between properties.

Broken assumption

Natural-language transfer learning can stand in for domain-specific chemical representations in polymer property prediction.

Why

Analysis of the learned embeddings suggests general-purpose LLMs do not capture fine chemical-structure information as well as handcrafted features or domain-specific embeddings.

From the published paper: Gupta et al. Macromolecular Rapid Communications, 2025. ↗CC BY

Research ideas from this failure

Unvetted lead — read the papers

New hypotheses this failure suggests, each with a crucial experiment, a rival explanation of the original result, and the papers it rests on. Written by AI from paper titles and abstracts found on OpenAlex (and Europe PMC, for life and health sciences).

Related failures

Nothing else logged looks close to this one yet.

Links in published research

Where this experiment's genes are known to matter outside its own research area, and the papers behind those links. Sources: Open Targets Platform and Europe PMC.