Five design patterns for AI-for-Science systems: decide what the model shouldn't learn
bravo_abad · x · 2026-09-25
This week's Discovery at Scale distills five recurring patterns from AI-for-Science papers, centered on the idea that the best scientific AI systems improve by deciding what the model should not have to learn.
- Split responsibilities: LLMs reason; deterministic tools handle operations that must be exact
- Encode known structure: build physics/chemistry/geometry constraints in rather than rediscovering them
- Choose the right target: mechanistically meaningful variables beat easy proxies
- Include the right context: scale, environment, neighbors, and conditions can be part of the sample
- Inspect residuals: what the model fails to explain reveals rare states or missing structure
The broader lesson: model design in science is often a division of labour between what must be learned and what is already known.
Related event: Five Design Patterns for AI for Science: Learn Only What Matters(2 posts)→
More from Research
- TorchCurves: Differentiable Learnable Parametric Curves for PyTorch — Alex Shtoff · 2026-11-11
- AI-assisted proof shows no voting rule beats Borda on strategyproofness — conitzer · 2026-11-11
- Markowitz mean-variance portfolio theory applied to mixing MCMC samplers — michaelchchoi · 2026-11-11
- AI agents run 24/7 for a month to rewrite Météo-France's AROME weather model — IgorCarron · 2026-11-11
- NYU's Steertuning Lets a Robot Learn a Chore From 5 Minutes of Human Demo — CyberRobooo · 2026-11-11
- New preprint maps human embryonic patterning in early organogenesis at spatial single-cell resolution — anshulkundaje · 2026-11-11