1. Start with a non-crazy shape that sort of does the same thing but worse (ie non-linearly or with big power fluctuations for different frequencies). Almost all shapes will interact with a sound wave in a frequency dependent manner so you can quickly scan through many initial options.
2. Define a goal function describing the desired outcome (like the "rainbow" shape of the linked article)
3. Permute the shape you started with iteratively with algorithms like simulated annealing, using the goal function from 2 as a means of defining the quality of the current solution.
The actual scientific paper (https://www.science.org/doi/10.1126/sciadv.ads7497) is worth reading. The goal function is the "Figure Of Merit" (FOM), described in equation 1. They also make a two-prong version they call the "lambda emitter" that takes in a mix of sound frequencies and directs low and high-frequency sound waves in different directions.
2. Define a goal function describing the desired outcome (like the "rainbow" shape of the linked article)
3. Permute the shape you started with iteratively with algorithms like simulated annealing, using the goal function from 2 as a means of defining the quality of the current solution.
The actual scientific paper (https://www.science.org/doi/10.1126/sciadv.ads7497) is worth reading. The goal function is the "Figure Of Merit" (FOM), described in equation 1. They also make a two-prong version they call the "lambda emitter" that takes in a mix of sound frequencies and directs low and high-frequency sound waves in different directions.