Kopitca, Latifi and Zhou take the Aalto plate work a step further. A centrally actuated silicon plate, 600-micrometre solder balls, a neural network trained on measured displacements, and machine vision in the loop: the algorithm repeatedly picks the excitation that shrinks the gap between where the particles are and where you want them.
They assembled up to 100 particles into user-specified two-dimensional shapes, including Latin letters. Runs took between 38 seconds and 38 minutes depending on the complexity of the target, with 63.2 to 88.5 per cent of particles ending up inside the shape boundary. The authors point towards cell and particle sorting and new fabrication routes, and note the method should extend to smaller particles and to fluid media.
What is worth stealing is the control structure rather than the hardware: predict, actuate, look, correct — the same loop that makes a plate rig repeatable instead of lucky.