# A random thesis idea I had

This is kind of connected to the high assurance digital twin
idea, but I am currently in the middle of reading and needed
to get this out of my head.

Here's the situation:

Manyu's work made a lot of progress to apply contract based
formal methods to nuclear power. To do this, an assumption
of a certain components output is fed into the input of the
next component. Math is done, and then the output of that
component becomes a guarantee, which is then the assumption
for the next component in line after that.

But here's a question: how do you know that your assumptions
and guarantee's are valid on a real system, in real time?
These contracts are based on having a model of the system
with which you can evaluate the assumptions/guarantee pairs.
But, real systems never will line up perfectly with a model,
and over time or different conditions, will absolutely have
different physical behaviors. Knowing if the contracts still
hold for the real system is a significant problem.

Here's where some online modeling in simulation can come in.
Perhaps, we can use a digital twin to estimate what the
critical model parameters for the contract methods are in
the real system. This is probably most easily accomplished
with either a physics informed neural network (PINN) or some
sort of particle filter bayesian nonsense. Once those
parameters are identified, we can reevaluate the contracts
to know a) if our system is safe, b) what our new
assumptions and safe operating range are, and c) make
strategic decisions about the plant control if necessary. 

This relates to the [autonomous framework paper](/Zettelkasten/Literature%20Notes/albertiAutomationLevelsNuclear2023.md) 
that talks about getting to higher levels of automation.
Level 3 is exactly this, the automated reactor operation
system being able to detect and diagnose what an error is.