Acoustic Metamaterial Platform

Acoustic metamaterials are engineered structures that control sound through geometry rather than mass. Where conventional soundproofing relies on thick, heavy materials, metamaterials use precisely shaped internal channels, chambers, and resonant cavities to absorb or redirect specific frequencies — achieving performance that would be physically impossible with uniform materials. These phenomena are often emergent and unpredictable, making them difficult to discover or optimize by traditional methods. So I'm experimenting with evolutionary algorithms and machine learning to generate them through mutation — the way nature arrives at her own cleverly optimized designs.

This platform is a side exploration into machine learning, evolutionary algorithms, generative design, and computational physics. The goal: automatically discover metamaterial geometries that maximize sound absorption in broadband and selected spectral bands, then fabricate them and validate in a physical testing setup.

The central challenge is representing a 3D geometry in a way that evolution can meaningfully search. Most generative design approaches use discrete primitives — boxes, cylinders, Boolean operations — but these can't express the thin-walled periodic channels that make metamaterials work. The search space is full of invalid geometries, dead ends, and structures that can't be manufactured.

I developed a continuous implicit representation that guarantees every point in the design space maps to a valid, manufacturable structure — thin walls, periodic boundaries, no mesh errors, no degenerate geometries. The encoding is smooth enough for gradient-based refinement but expressive enough to produce the complex internal channel networks that drive acoustic performance. 100% of the search space is viable, which means evolution never wastes a generation on garbage.

The population is seeded with diverse structural archetypes — channels, lattices, spirals, shells — then mutation and crossover explore freely, converging on geometries with the desired acoustic properties.

Commercial simulation tools like COMSOL can model acoustic material properties, but cost, licensing, and the need to evaluate thousands of geometries autonomously inside an evolutionary loop made them unfeasible. So I built a custom acoustic solver from scratch — a frequency-domain simulator that models wave propagation through each candidate geometry, accounts for viscothermal losses at wall boundaries, and extracts absorption coefficients following established measurement standards.

Building a physics simulator that produces trustworthy results was the hardest part of this project. Early versions produced artificially inflated absorption values — an entire 45,000-design evolution run had to be discarded when validation revealed multiple compounding errors in the boundary conditions, measurement model, and loss calculations. Seven commits rewrote the core physics. The corrected solver produces realistic 5–70% absorption values that converge to within 3–4% of analytical solutions.

The evolutionary loop runs on a distributed compute cluster with parallel workers. Each generation spawns offspring, mutates their genomes, and evaluates acoustic fitness through a multi-objective function designed to reward resonant structure over bulk porosity — penalizing the easy solutions (just make it porous) in favor of the interesting ones (geometry that manipulates wave behavior).

A neural surrogate model pre-screens the full population on GPU before committing to expensive physics simulation, passing only the most promising candidates to the full solver. This provides roughly a 50x speedup per generation. For elite designs, a gradient-based refinement step squeezes out additional absorption at targeted frequencies.

The best evolved design so far achieves 69% absorption at 2500 Hz, with broadband performance of 30–50% across 500–4000 Hz.

The platform includes a fabrication module that validates designs against the tolerances of 15 manufacturing processes — FDM, SLA, SLS, CNC milling, laser cutting, lost-wax casting, and more. Evolved geometries are exported as watertight meshes ready for 3D printing or toolpath generation. The geometry representation guarantees manifold output at any resolution, eliminating the mesh repair step that typically plagues generative design workflows.