ScramOpt

Overview

Expanding from my work on gas turbines, I am also working with my little brother on a scramjet optimization project based on JAXA’s M12-02 engine. We use Bayesian optimization with Gaussian process regression and Latin hypercube sampling to explore geometry and flow conditions, optimizing for thrust while considering thermal limits and avoiding inlet unstart conditions. The workflow I designed is fully automated and headless.

Using FreeCAD macros, the pipeline parametrically generates new scramjet geometries, varying key features such as inlet contraction, combustor length, fuel injection location and rate, and nozzle expansion. Each geometry is then meshed and simulated in SU2, where we run compressible Navier–Stokes solvers with turbulence and combustion models tuned to Mach 5.5 freestream conditions.

The CFD results feed into a surrogate model built with Gaussian Process Regression, seeded with Latin Hypercube Sampling for initial coverage. A Bayesian Optimization loop then guides the exploration, training the surrogate to predict flow outcomes and only sending the most promising candidates back to SU2 for verification.

What I Built / Did

• To accelerate convergence, I also incorporate adjoint-based solvers to extract gradient information, and the pipeline automatically recovers from failed simulations so it can run continuously on HPC resources.
• We have scheduled large-scale runs on the Perlmutter supercomputer at Lawrence Berkeley National Laboratory, where I will deploy the full optimization framework.

Missing asset: ga_phi_psi_Pib_best_mean.png