What is physical modeling synthesis?

Hit a real cymbal and a lot happens at once. The metal flexes, waves race outward, reflect off the edge, collide, and exchange energy between hundreds of modes of vibration. Hit it softer, or a centimetre to the left, and all of that changes. The sound is not a thing the cymbal has; it is a thing the cymbal does.

Physical modeling synthesis takes that seriously. Instead of storing what an instrument sounded like once (a sample), or stacking oscillators until something similar comes out (subtractive synthesis), it builds the instrument as equations: mass, stiffness, damping, geometry. Play a note and the computer solves those equations in real time. Whatever pressure wave falls out of the maths is the sound you hear.

That one design decision changes everything downstream. The model responds to how hard you hit it and where. It rings sympathetically. It can be pushed into territory the real object would never survive, because nothing breaks when you double the stiffness of a simulated plate.

Sampling solved realism early: record a piano note, play the recording back, and it is exactly a piano. The catch is that it is exactly one piano note, frozen. Velocity layers and round-robins soften the repetition, but a sample library is a photo album, not an instrument. Nothing in it responds; it replays.

Subtractive and FM synthesis sit at the other pole. They are gloriously playable and endlessly shapeable, but their building blocks (oscillators, filters, envelopes) have no physics in them. A sawtooth does not care how it was struck. When a subtractive patch imitates a bowed string, it is doing an impression.

Physical modeling is the third pole: the responsiveness of a real instrument with the freedom of a synthesizer. You give up the photographic accuracy of a sample and get behaviour in return.

The field has a neat origin story. In 1983 Kevin Karplus and Alex Strong published an algorithm so cheap it could run on the home computers of the day: fill a short buffer with noise, loop it, and average adjacent samples each pass. The result sounds uncannily like a plucked string, because the buffer behaves like one: the loop is the string's length, the averaging is its energy loss.

Julius Smith and colleagues at Stanford generalised that trick into digital waveguides, which model a string or an air column as waves travelling in both directions and reflecting at the ends. Waveguides became the first commercial face of physical modeling: Yamaha's VL1, in 1994, modeled saxophones and violins convincingly enough to unsettle people.

Alongside waveguides came modal synthesis, which describes an object as its resonant modes (each mode a frequency, an amplitude, a decay) and rings them like a bank of tuned filters. It is efficient and lovely for bells and bars, though the modes are fixed: the object cannot change shape while it rings.

The most direct approach is also the most expensive: finite difference time domain (FDTD) simulation, which chops the object into a grid and steps its equation of motion forward thousands of times per second, no shortcuts. For decades that was research-lab material. Modern CPUs changed the economics; a full plate simulation now runs happily inside a plugin. That is the approach CymaticPlate takes.

Under every physical model is a differential equation: a rule that says how each point of the object accelerates, given how the object is bent right now. For a stiff plate the rule is the Kirchhoff plate equation; for a string it is the wave equation with a stiffness term; for a fluid surface something wilder.

A computer cannot solve those rules exactly, so it chops them into pieces it can. Space becomes a grid of points; time becomes steps, tens of thousands per second. At each step, each grid point looks at its neighbours, applies the rule, and updates. Strike the object by adding energy to a few points; listen by reading one point's motion out as audio.

The craft is in the details. Grids must be fine enough to carry the frequencies you care about. The numerical scheme must be stable, or the simulation explodes into digital noise. Damping has to be frequency-dependent, because real objects lose treble faster than bass. Get all of it right and something remarkable happens: the model stops sounding like a synthesizer patch and starts sounding like a thing.

Sound is not the only output. Because the whole object is simulated, you can also watch it: CymaticPlate renders the plate's actual displacement, and the sand in its visualiser drifts to the real nodal lines, exactly as it would on a physical plate.

Real instruments never repeat themselves. Two strikes are never identical, because the object is still moving from the first when the second lands, and the new energy interacts with the old. That interaction is the character of an instrument, and it is exactly what playback can never give you.

A physical model gets it for free. State persists between notes. Hit a ringing plate and the strike scatters off waves already in flight. Systems like fluids and chemical patterns go further: they are nonlinear, meaning tiny differences in conditions grow into different outcomes. The result is music's favourite ratio, order with a controlled amount of surprise.

It also changes how you play. With samples you audition; with a model you provoke. You learn its tendencies like you learn an acoustic instrument, and it pays you back by never being quite the same twice.

rumcake's collection applies the same philosophy to six different pieces of physics. None of them fake it; in each one, the simulation is the sound engine, and each has a guide or manual that goes deeper.

• CymaticPlate: a clamped Kirchhoff plate solved by FDTD. Strike it, bow it with audio, and watch the sand find the nodal lines. The science of cymatics.
• Mercury: 88 mercury tilt switches, one per piano key, each a damped pendulum on a shared rail. Gravity is the sequencer.
• Lichtenberg: a dielectric breakdown model. Play a note and a forked discharge crawls across the field, arpeggiating every column it crosses.
• Bloom: a Gray-Scott reaction-diffusion pattern whose moving wave fronts fire self-playing string pads. The science of reaction-diffusion.
• Ferro: a ferrofluid pushed past the Rosensweig instability on a tempo-synced pulse. The spike lattice writes a looping techno riff. The science of ferrofluids.
• LoopDove: the odd one out, an effect rather than an instrument. A murmuration looper that catches a few beats and flies them, scattering your audio the way starlings scatter.

Every plugin ships as an Audio Unit and VST3 for macOS, each has a free trial, and the whole collection is available as one bundle.