2025-08-08 13:56:35 -04:00

279 lines
11 KiB
Markdown

# Thesis Ideas 2025-07-30
Following our group meeting from Monday, July 28th, Dan
suggested I write down 6 ideas, and from them we shall
figure out a possible topic idea that I can really start
working on.
I used ChatGPT to do some of the heavy lifting based on the
papers I've been reading, and leveraged the 'deep research'
feature. Here are some of my favorite ideas, broken down
into goals, outcomes, impact, and related papers.
___________________________________________________________
## **Integrating Shielding into Nuclear Power Control**
### Goal:
The goal of this research is develop machine learning
enabled control algorithims for nuclear power applications
that incoporate shielding: a formal guarantee of adherence
to system specifications without augmenting the machine
learning process.
### Outcomes:
For this research to be successful, I will accomplish the
following:
1. Identify key controllers in a nuclear power context with
the most benefit from using an ML-based controller
2. Translate regulatory and system level requirements into a
formal specification to synthesize a controller 'shield'.
This shield monitors the ML controller and intervenes
whenever a requirement is predicted to be violated.
3. Evaluate performance of the ML controller with attached
shield, while assessing the amount of shield useage for
different operating scenarios (power up, shut down, regular
load following)
### Impact:
Machine learning controllers can outperform PID and
rule-based controllers by adapting to nonlinear dynamics,
optimizing over multi-objective cost functions, and changing
plant conditions. But, these ML controllers are often
*unexplainable*, meaning that their global behavior is not
easily understood.This unexplainability prevents ML based
controllers from being used in high-assurance usecases such
as nuclear power. Shielding can address this issue, by
providing a formal runtime assurance, allieviating the
burden of explainability away from the machine learning
algorithm. This work would further bring regulatory
requiremnts into the formal design of control systems and
help bridge the gap between high assurance systems and the
start of the art in control.
### Relevant Papers
[[safe-reinforcement-learning-via-shielding]]
[[evaluating-robustness-of-neural-networks-with-mixed-integer-programming]]
___________________________________________________________
## **Formally Verified Neural Network Control of Control Rod System**
### Goals:
The goal of this research is to use formal methods to ensure that
a neural network based control rod controller will never violate
safety guarantees of a reactor trip system. To do this, a
satisfiability modulo theory method will be applied to
exhaustively search the network for potential failure modes.
### Outcomes:
If this research is successful, I will have accomplished the
following:
- Build a neural network controller for real time control of a
control rod system.
- Formalize safety guarantees of shutdown margin in a
satisfiability modulo theory embedding
- Formally verify that the neural network based controller will
not violate any shutdown margin restrictions
### Impact:
SMT solvers and MILP formulations have been applied to neural
networks to ensure that the network is resilient to input
perturbations. I think we can expand this to more general
considerations of the state space, especially when there are a
relatively small number of states such as in power contexts. The
benefit of this system is that we would get closer to saying
neural network based systems can be high assurance for physical
systems.
### Related Papers:
[[reluplex-an-efficient-smt-solver-for-verifying-deep-neural-networks]]
[[evaluating-robustness-of-neural-networks-with-mixed-integer-programming]]
[[formal-verification-of-neural-network-controlled-autonomous-systems]]
___________________________________________________________
## **Temporal Logic Specifications for Autonomous Controller Synthesis**
### Goals:
The goal of this program is to use temporal logic
specifications to procedurally generate autonomous
supervisory controllers for a reactor system.
### Outcomes:
If this research is successful, I will have accomplished the
following:
- Captured high level safety and operating requirements in a
temporal logic language such as TLA+ or FRET
- Synthesize a supervisory controller from the temporal
logic specification that can be implemented on a real
control system with minimal user effort.
- Verify the supervisory controller generated adheres to
safety specifications using exhaustive model checking.
### Impact:
### Related Papers:
___________________________________________________________
## **Formally Verified Runtime Monitoring and Fallback**
### Goals:
If this research is successful, we will be able to generate
autonomous controller shields that provably adhere to specifications
written with temporal logic.
### Outcomes:
- Create an intermediary shield that mediates signals between an
optimal control system and the physical plant (MODBUS)?
- Translate specifications in a language like TLA+ into an
executable program
- Provide proof artifacts that automatically generated
shield components will not allow an arbitrary controller to
reach an unsafe state.
### Impact:
Shielding is one of the preeminent ways to do safe machine
learning controllers. Instead of putting the proof burden on
the machine learning component, shielding creates a safe
boundary in the state space where a safety controller will
step in if the machine learning controller endangers the
system. This technology solves a critical problem with high
assurance systems: high assurance systems have critical
safety requirements that make scrutiny on autonomous systems
safety intense. Shielding can provide a safety barrier for
the controller, allowing the architecture of the control
laws to be amenable to more efficient machine learning based
methods. Finally, utilizing an automatic translation from a
temporal logic formulation of a speculation will allow the
engineers of these systems to quickly and clearly implement
a shield, without all of the cumbersome derivation.
### Related Papers:
[[on-using-real-time-reachability-for-the-safety-assurance-of-machine-learning-controllers]]
[[enhancing-cyber-physical-system-dependability-via-synthesis-challenges-and-future-directions]]
[[safe-reinforcement-learning-via-shielding]]
___________________________________________________________
## **Data-Driven Fault Detection Using High-Assurance Digital Twins**
(8)
### Goals:
The goal of this research is to use machine learning to
identify system faults of a reactor control system during
runtime. A digital twin will be compared to measurements
from a real plant to identify issues such as coolant losses,
sensor and actuator failures, or component degredation so
that safety strategic decisions about the plant can be made
autonomously.
### Outcomes:
For this research to be successful, I will accomplish the
following:
- Create a simulation suite for the Small Modular Advanced
High Temperature Reactor (SmAHTR) to simulate fault
conditions of sensors, actuators, and component degradation.
- Develop a physics informed neural network (PINN) approach
to evaluate physics discrepancies in measured signals and
to estimate physically relevant parameters to determine
real system divergence from the nominal plant.
- Realize a proof of concept autonomous controller than can
react to fault conditions by switching to different
control modes rather than only responding with reactor
shutdown.
### Impact:
The nuclear energy industry's largest expense is operations
and maintenance (O&M). These costs include typical reactor repair
and refueling, the labor involved to complete such
maintenance, and finally the labor involved in operating the
reactor itself. Currently the largest of these O&M expenses
is the labor and part cost used in maintenance, while large
nuclear reactor facilities require a modest reactor operator
budget per megawatt of energy produced. The advent of small
modular reactors (SMRs) and microreactors (MRs) will change
these economics significantly.
As SMRs and MRs become more common, the cost of repair and
maintenance should reduce dramatically as nuclear power
components will become modular, replaceable parts instead of
the bespoke reactor designs currently operating. Operator
wages, however, can be expected to increase without
introducing greater controller autonomy. SMRs and MRs are
much smaller output designs per reactor core, and if they
are required to employ the same size reactor operator team
as a conventional large reactor, will suffer from much
larger operator expense per megawatt. Greater controller
autonomy can solve this problem by unloading some reactor
control responsibilities from the operator, and therein
reduce labor consumption.
<# TO DO #>
Finally reactor safety can be improved by greater autonomy
yada yada find some reasons to back this up.
### Related Papers:
___________________________________________________________
## **Verified Adaptive Control**
### Goals:
The goal of this research is to create an adaptive controller
that can adjust to system dynamics changes over time to maintain
an optimal control, while using formal methods to provide strong
safety guarantees about the malleable control law.
### Outcomes:
For this research to be successful, I will accomplish the
following:
- Create a simulation suite for the Small Modular Advanced High
Temperature Reactor (SmAHTR) to simulate component degradation
such as heat exchanger blockages and fuel concentration burn-up.*
- Create an adaptive control rod controller to maximize load following
precision for a simulated power grid demand.
- Use contract based verification at runtime to ensure that
learned parameters for the adaptive controller remain within
safety specification limits
*Is this actually even a problem for SmAHTR? Figuring the fuel is
suspended in the salt I'd assume chemistry is pretty strictly
controlled. I'm sure I can find other examples.
### Impact:
Certain reactor control systems are already automatic systems,
such as constant temperature or pressure controls for operating
at steady state. These simple controllers are able to follow load
changes from the power grid on their own, but over will lose
efficiency as the underlying plant mechanics become less
efficient, or maintenance is performed and components are
refreshed. For nuclear power contexts, fine control is ideal to
maximize profits and to minimize energy wasteage. This is not an
easy problem to solve, however, as the dynamics of the underlying
plant are constantly changing. Adaptive control can help address
this issue, but learnable controllers must come with guarantees
of safety in order to be attractive to the nuclear industry.
### Related Papers: