A research thesis for scientific AI
I study how AI systems represent scientific worlds, act inside them, and discover principles that transfer beyond a single task.
Representation
It is a structured object in which data, dynamics, uncertainty, tools, and physical constraints remain distinct—and can still interact.
Neural operators, world models, simulators, PDE solvers, and databases become parts of one executable representation.
Agency
An AI scientist must decide what evidence is missing, choose an intervention, and change its belief when the world disagrees.
The scientific method becomes the agent architecture.
Transfer
A failure mode found in one basin should inform another. An instability found under one boundary condition should remain useful under the next.
Can an AI scientist discover reusable scientific principles rather than isolated answers?
This is the long-horizon thesis: learn what transfers, not merely what fits.