Multi-objective Bayesian optimization reveals HiPIMS deposition rates exceeding direct current sputtering

Pulsed sputter processes are essential for a wide range of applications from hard coatings to microelectronics. High power impulse magnetron sputtering (HiPIMS) offers unique possibilities for process design due to its high ionized flux fraction. Because of the high peak-current densities involved, correspondingly high plasma densities generate many ions. This enables tuning of kinetic energies by applying accelerating bias potentials. However, some HiPIMS processes have low deposition rates due to the back-attraction these ions experience with the negatively charged sputter target. Recent work in HiPIMS has targeted this problem, exploring complex pulsing schemes such as bipolar, multi-pulse, or chopped HiPIMS. These pulsing schemes introduce another problem: process complexity. As such, the process space, even with coarse intervals quickly blows up to 1E5-1E20 possible process combinations. This calls for efficient and intelligent sampling, which can be enabled by machine-learning algorithms such as Bayesian optimization. In this work we use multi-objective Bayesian optimization (MOBO) to efficiently sample the process space of pulse-packets, targeting high ionization flux fractions and deposition rates. Over 2,000 distinct processes were sampled during this exploration. Surprisingly, we found that many HiPIMS processes, including single pulse, can achieve deposition rates higher than DC references. In fact, at a working pressure of 1.0 Pa and moderate ionization rates of 5-20% we found that ms-scale pulses were ideal. We hypothesize these processes maximize gas rarefaction, creating a low-collision channel between the target and the substrate. These results showcase previously unreported behavior of pulsed sputter processes and challenge a long-held trade-off between film quality and speed in thin-film manufacturing.