Autonomous closed-loop control of ultrafast laser processes remains an open challenge. Machine learning is actively being explored, but usually requires large datasets to be effective. From a practical point of view, measurements usually require taking the sample to a separate instrument.

This work presents simultaneous optimization of ablation depth and surface roughness Sa in only 8 iterations, using a physics-informed Bayesian algorithm with an integrated confocal profilometer performing measurements between processing steps.

Experiments were conducted on SS316L stainless steel using a <400 fs, 50 W femtosecond laser. Pulse energy and scan speed were varied over 1-50 µJ and 0.1-1.0 m/s respectively, with hatch spacing, beam waist, repetition rate and number of passes kept constant.

A two-temperature model of ultrafast laser ablation, calibrated on a small pre-existing dataset, generates seed experiments with predicted depths close to the target, reducing iterations before convergence. The Bayesian algorithm iteratively proposes the next parameter set, converging on dual-objective targets of 30 µm ablation depth within ±1 µm and Sa ≤ 0.35 µm.

The low-roughness regime was reached within 2 iterations, with full dual-objective convergence in fewer than 10 iterations. The approach is applicable to other materials and process targets with minimal recalibration, demonstrating a simple and practical architecture for the simultaneous optimization of multiple process targets in a closed-loop autonomous system.