Skip to content
AI for Science

Discovery Engines

The Recipe & Its Limits

8 min+85 XPStep 1/10

One recipe, many sciences

Step back and every story in this track rhymes:

  • Generate labels with an expensive oracle — crystallography for AlphaFold, assimilated reanalysis for GraphCast, DFT for GNoME, a symbolic engine for AlphaGeometry
  • Train a cheap surrogate on the oracle's output — a model that approximates the expensive answer in milliseconds
  • Search at surrogate speed — predict, roll out, or screen millions of candidates no oracle could afford
  • Verify with the oracle, feed results back — active learning closes the loop

The breakthroughs were never 'AI replacing science.' They were AI amortizing science's most expensive step across a million candidates.