How can one coax a model to generate novel hypotheses in materials science?

[PaisleyD]

I’ve been trying to use a large language model to help generate hypotheses for my materials science research, but I keep hitting a wall where its suggestions feel derivative of existing literature. It’s great at summarizing known relationships, but I’m struggling to prompt it toward genuinely novel, testable ideas. Has anyone else found a method to guide these models beyond pattern recognition into more speculative, yet scientifically plausible, territory?

[Eleanor_W]

I’ve tried prompts that push the model to justify a hypothesis with concrete, testable predictions and then deliberately probe those predictions with counterfactuals. It still leans toward adjusting known trends rather than spawning clearly new ideas.

[MatthewLD]

I once fed in a tiny, noisy synthetic dataset and asked for hypotheses that would improve robustness under perturbations; every time it offered a few outliers, but most were still copies of familiar relationships.

[ThomasT]

Two-stage prompts helped some: first a speculative phase where the model riffs on plausible ideas, then a constraint phase where I asked for feasibility and measurement details. It yielded a couple directions that sounded fresh, though still not slam-dunk novel.

[MichaelFL]

I worry the model is playing safe because it’s optimizing for plausible, literature-like outputs. The more you push for novelty, the more it seems to retreat behind standard physics explanations.

[Paisley.L]

We did a red-team style prompt: ask for hypotheses that would falsify a widely accepted trend under a specific set of experimental conditions. It forced sharper thinking and surfaced at least a few testable angles I hadn’t seen in the literature.

[AnthonyS]

Sometimes I drift into talking about data quality or experimental design instead of ideas, and the model keeps looping back to known correlations. It’s easy to lose the thread and end up with something similar to what we already have.

[Scarlett61]

Question: is the real bottleneck the model’s training data or is it our framing of novelty itself? If we define novelty more narrowly as falsifiable within a realistic budget, would that help?