6.0 KiB
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 to develop machine learning control algorithms for nuclear power applications with strict safety guarantees.
Outcomes:
If this research is successful, I will have accomplished the following:
-
Develop controller shielding methods for nuclear power contexts
-
Provide concrete safety guarantees for autonomous control of a nuclear asset.
-
???
Impact:
Machine learning based systems have been shown to be more efficient than typical PID based controllers, and are able to learn more complex objective functions than a typical controller can. The problem with these controllers though is that they are often unexplainable. This is not acceptable for high assurance applications, where slight perturbations on inputs can yield wildly different outputs. Shielding can solve this problem, helping ensure safety of ML based controllers while not limiting their development or construction.
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 automatically.
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 Model Verification for High-Assurance Digital Twins
(8)
Goals:
Outcomes:
Impact:
Related Papers:
Verified Adaptive Control
(10)