Institut Polytechnique de Paris

Machine-learning methods used on physical and engineering systems.

• TransformersAttention-based sequence models, and what they do to signals and time series rather than to text.
• Diffusion modelsGenerative modelling by learned denoising, and the score-based formulation behind it.
• Flow matchingContinuous-time generative modelling by regressing a velocity field, a cheaper route to the same transport.
• Scientific machine learningLearning where a physical model already exists: surrogates, operators and hybrid physical / learned models.
• Constrained learningTraining under constraints instead of unconstrained loss minimisation, which is what engineering problems actually look like.

Array calibration is an inverse problem, which is where generative priors, surrogates and constrained estimation apply:

• Calibration as inferenceRecovering the complex coupling matrix of an array from a limited number of noisy measurements is estimation under a measurement budget, not a training problem.
• Priors over driftA model of how an array drifts between calibrations is exactly the sort of temporal prior these methods produce.
• Surrogates for simulationA full-wave solve is expensive. A surrogate over geometry lets the search spend solver time where it changes the answer.
• Constraints are physicalPassivity, reciprocity and hardware limits on amplitude and phase are constraints, not regularisers chosen for convenience.

Where this is appliedAetherArray calibration trackRF computing and ML skills