Conventional deflagration based propulsion systems often require converging-diverging nozzles to accelerate the flow to a supersonic Mach number to maximize performance. In rotating detonation engines (RDEs), however, the high pressures and temperatures generated by the detonation wave may enable supersonic exhaust without a conventional converging diverging nozzle. To investigate this, we performed high resolution numerical simulations using our in house HMM solver. Detailed chemical kinetics and adaptive mesh refinement were employed to resolve the complex flow and detonation wave dynamics. We found that modifying the combustion chamber geometry from a uniform cross section to purely converging or diverging sections substantially alters the detonation wave structure and propagation behavior. Despite these significant changes in wave dynamics, the resulting effects on overall thrust and specific impulse are comparatively small.

Scientific discovery is fundamentally driven by the triad of abduction, induction, and deduction. These three modes of reasoning also form the basis of modern AI systems. Our current AI agents work alongside human researchers to reduce cognitive load, particularly across the inductive and deductive components of scientific inquiry. Researchers remain central to the abductive process: forming plausible explanations and hypotheses from sparse, incomplete, or uncertain observations.  Most scientific campaigns also depend on a functional numerical codebase for testing and exploration. To support this, we develop Domain-Specific AI Harnesses (DSHs) tailored to individual scientific campaigns. Within this framework, the Chain-of-Unit Physics and AUTO systems work jointly to develop hardware-optimized, verified physics codes that provide the computational foundation for large-scale numerical experimentation and scientific exploration.

A key differentiator of the DSH is its treatment of agent knowledge and uncertainty. Agents are designed to use their internal knowledge while explicitly identifying and reporting knowledge gaps within the same reasoning sequence, thereby establishing provenance for agent errors and failures. The harness also preserves the causal structure of the scientific process. Agent states, actions, reasoning transitions, and interactions are recorded as a graph, producing a persistent trace of how a scientific result was generated. This graph becomes an additional artifact of the scientific campaign, enabling subsequent analysis, verification, interpretation, and reconstruction of the discovery process.  We are developing the underlying methods required for these persistent scientific AI systems, including verification of AI-generated scientific workflows, interpretability, emergent behavior in agent swarms, and novel architectures for operationalizing scientific abstraction and hypothesis generation.

Simulating a rotating detonation combustor (RDC) presents a formidable computational challenge due to its inherent multiphysics, multiscale nature. Capturing these physics accurately requires resolving reaction zones on a micron scale with chemical time scales orders of magnitude faster than the fluid flow, while simultaneously tracking wave dynamics across a chamber spanning tens of centimeters over millisecond durations. This vast disparity which encompasses roughly five orders of magnitude in length and an even broader temporal domain makes uniform fine-resolution modeling computationally unfeasible across entire domains. To address this we have developed a compressible flow solver for turbulent reacting flows that utilizes  adaptive mesh refinement and hardware acceleration provided by GPUs. A significant effort has been made to optimize the solver for GPUs details of which are reported in our recent work (Carreon et al., 2026). 

Reducing the dimension of a physical model can substantially lower computational cost, but the resulting simplifications introduce errors that affect its predictions. We are developing  a principled method for deriving these model-form errors and expressing them as explicit, interpretable correction terms tied to their underlying sources. By distinguishing contributions from simplified physics, geometry, and boundary conditions, we can trace errors in a reduced model to specific modeling assumptions.