In current postdoctoral research, I am exploring graph-based diffusion models and differentiable optimization layers for control and planning problems where predictions must satisfy hard physical and operational constraints. This work is motivated by stochastic optimal load shedding, resilient power-system planning under climate and uncertainty, and broader decision problems in cyber-physical systems where fast inference is valuable only if the resulting actions are feasible.

Key contributions

• Investigating discrete diffusion models for high-dimensional binary decisions such as load-shedding actions on power grids
• Using optimization-aware projection and differentiable physics layers so generated actions remain feasible under hard operational constraints
• Connecting this line of work to broader postdoctoral themes in uncertainty-aware control, flood operations, and decision-making under partial observability

Impact

The broader goal is to connect modern generative modeling with the guarantees, feasibility checks, and risk-awareness required in high-stakes engineering systems, where speed alone is not enough and physically valid decisions are critical.