Scientific ML · surrogate modeling · UQ
My scientific-ML projects focus on moving from empirical model-building toward mathematical structure: Gaussian processes, surrogate approximation, model reduction, operator learning, and uncertainty-aware validation.
A mesh-free NURBS-based boundary element solver for solenoid magnetic-field computation, written from the integral formulation through to code. It generates the multifidelity electromagnetic datasets the surrogate work below is trained on, and carries a closed-form verification suite that runs in CI.
A systematic study of surrogate methods from first principles, including Gaussian process regression, response surface methods, radial basis functions, uncertainty behavior, and failure modes of approximation.
A clearer account of why surrogate models generalise, how their uncertainty can be quantified, and where their assumptions break down — stated precisely enough to be tested rather than asserted.
A surrogate is only useful if you know where it stops being accurate. These projects are about characterising that boundary rather than reporting a single error metric — which is the same question as the rest of my work, asked about approximations instead of pipelines.