Really? If you know about synths then it isn’t really that difficult. You have one or more sound generators, sine wave, saw tooth, etc. You have attack (how fast the sound reaches full potential), decay (how long that sound is played before it drops to sustain), sustain (how long the sound lingers), release (how long until the sound fades away to nothing).

You have filters for the sound, low pass, high pass, distortion, etc. that process the audio into output audio.

It’s basic sound engineering. However, using a neural net to find interesting combinations and be able to label them would be really fun to build a catalog of dreamscapes or tones.