Science-based audio plugins: a field guide

Direct answer

Two questions sort the whole field.

A growing shelf of music software claims some kind of scientific ancestry. Physics-based synths, chaos oscillators, plugins named after chemists. Some of these run real equations inside the audio path. Others use science as an interface metaphor over conventional DSP. Both can be excellent instruments; they are different things, and it helps to know which one you're buying.

Before going further: I run Chiral Audio, one of the studios on this list. Eight of the entries below are mine. I've tried to be as plain about everyone else's work as I am about ours, and where a competitor does something better than we do, I say so. Judge for yourself whether the map survives the bias.

Two questions place any tool on the map.

First: where does the science live? In some tools the equation is the engine. Delete it and there is no sound. In others the science shapes the controls, the visuals, the workflow, while the audio is produced by ordinary oscillators and filters. The second kind isn't lesser, but it is different, and vendors rarely volunteer which kind they built.

Second: does the difference reach your ears? Suppose you swapped the physical model for an LFO and a random number generator tuned to imitate it. If no listener could tell, the physics is decoration no matter how faithfully it's computed. The honest test of a science-based instrument is that the cheap fake fails.

Region one

Simulation as instrument.

The largest and most mature cluster. The physics here is mechanical: masses, springs, strings, plates, resonant bodies. This is physical modeling in the term-of-art sense, and the tools in it are some of the best-sounding software ever made.

Anukari is the current frontier: a 3D physical-modeling synthesizer and effect where you assemble instruments from simulated masses, springs, mallets, bows, and virtual microphones, and watch the assembly vibrate. The simulation is the whole product, computed brute-force in real time. Nobody else is this literal, and the results sound like objects, not patches.

Baby Audio Atoms takes a related mass-spring network and bows it, trading Anukari's open construction set for a curated, playable instrument. It finds eerie, breathing, almost-acoustic territory fast.

AAS Chromaphone 3 is the veteran. Modal synthesis of acoustic objects, drumheads, bars, plates, membranes, tubes, coupled and struck. Twenty years of refinement show; if you want to know what disciplined physical modeling sounds like at maturity, start here.

Madrona Labs Kaivo crosses granular synthesis with physical models, so a sample cloud can excite a modeled string or body. Wood, water, and metal timbres that neither technique reaches alone.

Newfangled Pendulate earns its place on a technicality that matters: its oscillator is a double pendulum, the motion of the simulation sonified directly and tamed through a wavefolder. That's a chaotic system as the sound source itself, which is why it moves from sine-like order into snarling nonlinearity the way real chaos does.

Sound Particles simulates too, but the simulation drives space rather than timbre: particle systems, thousands of point sources with positions and velocities, rendered to any speaker format. Film sound has quietly depended on it for years.

Region two

Automata as sequencer.

A smaller cluster where the science generates events rather than waveforms: rule systems deciding when and what, with conventional synthesis downstream.

Lectric Panda Automata runs cellular automata as a MIDI sequencer. Evolving bit patterns drive triggers, pitch, velocity, and controller data. Whether a given rule produces music or noise is genuinely emergent, which is the appeal.

Audio Damage Automaton did the same thing to a buffer effect years earlier, a Conway-style grid triggering stutters and bitcrush. It's discontinued, and it's still the reference for the idea.

Region three

Metaphor as interface.

Here the science organizes how you work, not how the sound is computed. Done well, this is not a consolation prize.

Sonic Charge Synplant is the proof. You plant seeds, grow branches, splice genes; underneath is a conventional (and lovely) synth engine. The genetics is a metaphor, and it produces a way of exploring sound space that knob-per-parameter design never found. Synplant is the best argument in the field that a brilliant metaphor beats a timid simulation.

Unusable Engineering works the border. Their effects take concepts from fluid dynamics, cavitation, granular discharge, and translate them into unusual signal processors. How much is simulation and how much is inspiration varies by product, which is itself worth knowing before you buy.

Region four

Stochastic engines.

sonicLAB Cosmosf descends from the Xenakis tradition: probability distributions composing sonic events across nested time scales. The mathematics of chance is the instrument, and the results have a character no envelope will give you. Demanding, and deep.

Region five

Biosonification.

PlantWave and MIDI Sprout read electrical variation from living plants and translate it into MIDI. The measurement is real biology. The music is a mapping choice: the same signal could be a drone, a melody, or a drum pattern depending on decisions a designer made. That doesn't make it a trick, but the plant is not composing, and the honest vendors say so.

Disclosure, continued

How we test ours.

Chiral Audio's eight plugins live in the simulation region, with a bias toward sciences the mechanical-physics cluster doesn't touch: statistical mechanics, reaction kinetics, relativity, condensed matter, glass physics. Our internal bar is the two questions above, enforced adversarially. The equation must run in the audio path, and a memoryless fake must fail a listening test against it. This is our QA standard, not a grading rubric for anyone else's work.

One worked example of the bar in practice. Anneal is a synthesizer whose partials carry a thermal history: cool a patch slowly and it lands as a tuned crystal, quench it and disorder freezes in as glass, at the same knob position. Before shipping it we built the cheap fake, a static detune map scaled by the same controls, and measured both against every path-dependence metric we could define: hysteresis loop area, cooling-rate response, aging. The fake scored zero on all of them. The full numbers are in the path-dependence note. When one of our products can't beat its own fake, it doesn't ship; that test has killed prototypes.

The other seven, each with its governing science: Foxfire (Kuramoto synchronization as a chorus), Autocatalysis (logistic population growth as resonators), Soliton (dispersion-saturation balance as a delay that won't decay), Boltzmann (an energy landscape as a MIDI sequencer), Anderson Freeze (localization as a living spectral freeze), Geodesic (Schwarzschild time dilation as a delay field), and Arrhenius (reaction kinetics as a multi-band gate). Mechanism derivations, with the original papers, are in the research notes.

Reference

The table.

"Audio path" means the science computes or directly shapes the signal; "event generation" means it decides what happens and when; "interface" means it organizes the workflow over conventional DSP. Border cases are marked as such.

The open map

The thin regions.

Read the table by science rather than by tool and the imbalance is obvious. Mechanical physics is crowded, and for good reason: strings and plates map onto sound with almost no translation. The rest of science barely appears. Ecology, fluid dynamics beyond flavor, geology, electrochemistry, phase transitions, developmental biology. Most of these have dynamics that are audibly distinctive: hysteresis, avalanches, fronts, competition, memory. The mapping work is harder because the source system isn't already acoustic, and that is exactly why the territory is still open.

If you build in this space, or know a tool this map missed, tell me. The guide gets corrected, not defended.

Keep reading.