Dr. Aarav Mehta, PhD

The thesis develops a verification framework establishing formal safety bounds for a class of learned control policies, with computational tractability sufficient for industrial-scale systems, and validates the framework on two production process-control deployments.

Learned components are increasingly used in industrial control, but the verification techniques available either do not scale to realistic system sizes or require assumptions that deployed systems violate. Regulatory approval consequently rests on testing evidence, which cannot establish the absence of unsafe behaviour.

The work combines abstract interpretation with a decomposition exploiting structural properties of the controller class, reducing verification cost from exponential to polynomial in the relevant dimension. The framework was implemented and evaluated on synthetic benchmarks and two industrial deployments under a research agreement.

Safety bounds were computed for controllers two orders of magnitude larger than previously tractable. On the industrial deployments the bounds obtained were tight enough to support regulatory documentation. A negative result established that the decomposition does not extend to controllers with unbounded recurrent state.

The thesis makes formal safety verification tractable for an industrially relevant class of learned controllers.

The method is in use on two production process-control systems; extension work addresses recurrent architectures.