NSF AI award $4M: an AI institute for dynamic systems that obey physical law (University of Washington, $20M estimated total)

1What "dynamic systems" refers to

Dynamic systems is the collective term for subjects whose state changes over time. Fluid flow, the motion of an aircraft or a vehicle, the power grid, the climate, reactions inside a living body — in each case the next state follows from the present state and the laws at work. The field carries a long tradition of describing these with differential equations and treating stability and controllability mathematically.

2How this differs from pattern recognition

AI for images and text grew by finding statistical regularities in large numbers of examples. Dynamic systems, by contrast, come with constraints that must not be broken — conservation laws, thermodynamics. A prediction that fits the data well but is physically impossible is unusable here.

Research therefore concentrates on how to learn while respecting those laws, how to produce trustworthy predictions from limited data, and how to carry predictions through into control. An institute is the chosen form because these questions cut across application areas.

3Where it sits in the AI Research Institutes program

NSF's AI Research Institutes program builds large centers with several participating organizations. Here the record shows $4M obligated against a $20M estimated total across five years. The obligated figure is what has been committed so far; the scale of the whole plan is what the estimated total expresses.

4The focus rests on three things

The problem for AI facing dynamical systems is not improving goodness of fit. Against subjects carrying constraints that must not be broken, such as conservation laws and thermodynamics, the question is what to verify while learning.

1. 1Learn while obeying the lawsA prediction that fits the data well but is physically impossible is unusable in this field
2. 2Predict reliably from little dataFor subjects where experiments and observations are few, the spread of the error has to be knowable
3. 3Carry prediction through to controlKnowing the next state and being able to drive the system to a desired one are separate problems

Dynamical systems is the general term for subjects whose state changes over time — fluid flow, the motion of aircraft and vehicles, power grids, climate, reactions inside an organism. In each, the next state follows from the present state and the laws at work.

The field carries a long record of describing these with differential equations and treating stability and controllability mathematically, and an institute is the form chosen because these questions cross application areas.