Veröffentlichungen und Präsentationen

Green14 is developing a new, lower-impact technology for producing metal powders for additive manufacturing. From the start, the process is paired with a Physics AI loop rather than the traditional empirical FEM approach: the physical model itself is learned through a reinforcement-learning loop between experiments on a pilot reactor and high-fidelity simulations, before being used to optimise the process and scale it up. COMSOL Multiphysics® is the high-fidelity reference in this loop, producing the simulations from which a fast AI surrogate is trained. Our foundation model is being built on Ti-6Al-4V (Ti-64) production, before extension to other critical raw materials. This work presents the COMSOL side of the loop: a full digital twin of the DC plasma reactor and a custom phase-change particle-tracing layer that fills a gap in the Particle Tracing for Fluid Flow Module, which does not provide solid–liquid–vapor phase change out of the box.

The reactor twin provides the Eulerian background fields into which the Lagrangian particles are released. Built with the CFD Module and the Heat Transfer Module, it solves compressible turbulent Navier–Stokes with the SST closure, multicomponent species transport, conjugate heat transfer through cooled walls, water cooling, and gas-phase radiation through the full chamber and converging–diverging nozzle. The arc-heated Ar/H₂ plasma jet enters as an inlet boundary condition, leaving the model centred on the chamber where the particles actually live. The geometry is fully parametrized and version-controlled in COMSOL Model Manager Server, and every prediction is directly checked against measurements from the Green14 pilot reactor.

Building on top of the Particle Tracing for Fluid Flow Module, we have developed an extended particle model that captures the full solid → liquid → vapor phase change of the injected powder together with droplet breakup in the supersonic environment of the converging–diverging nozzle. Particles are injected following the same particle-size distribution used in the real process, so the simulated population mirrors the feedstock seen by the reactor. The same framework is reused for SiO₂, WO₃, TiO₂, and Ti-64 by swapping a single material card.

The model produces, for every injected particle, the full time history of temperature, melting, and evaporation as it travels through the reactor, together with the vapor distribution it leaves in the gas phase. SiO₂ and Ti-64 cases have been run end-to-end, and the same framework is now being calibrated against on-reactor Ti-64 atomization campaigns. Together with the continuously validated reactor twin, this consistent phase-change particle layer gives the COMSOL stack what it needs to serve as ground truth for Green14's downstream AI surrogate, closing the experiment–simulation–AI loop on which the technology is being developed.