SPLICE: Sparse Physics Library Identification and Compositional Expansion

Key points

• We ask whether a pretrained library can instead be adapted and grown from unlabeled trajectories in which multiple mechanisms act simultaneously.1
• We introduce SPLICE, a framework that discovers operators without mechanism labels on mixture data.1
• The contributions of the candidate operators to each trajectory are inferred by solving a convex optimization problem.1
• On Navier-Stokes, the discovered operators compose about as accurately as their supervised counterparts.1

Sources (1)

• [1]SPLICE: Sparse Physics Library Identification and Compositional Expansion
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 04:07 PM

  Neural operator splitting composes a library of neural operators, one per physical mechanism, into zero-shot simulators of new equations.

  We ask whether a pretrained library can instead be adapted and grown from unlabeled trajectories in which multiple mechanisms act simultaneously.

Extractive summary: sentences quoted from the sources.