PhononIQ v2.0
Home
Study
Geometry
Physics
Mesh
Results
Export
PhononIQ
ML-powered phononic crystal bandgap prediction
with full fabrication specs in under 5 ms
3,000 COMSOL simulations
R² 0.9995 Neural network
<5 ms Prediction time
Band Structure
Predict dispersion, bandgaps & Dirac cones
Inverse Design
Find geometry for target frequency range
Design Atlas
Interactive parameter space heatmaps
Dispersion
Group velocity, DOS & waveguide
Compare
Side-by-side band overlay
Batch
Parametric sweeps
Quick Presets
Wide Bandgap
TR=0.36 FF=0.45
★★★ FBW >30%
Low Frequency
TR=1.50 FF=0.30 a₀=5mm
Sub-kHz range
Compact
TR=0.50 FF=0.40 a₀=0.5mm
MHz MEMS range
Max Attenuation
TR=0.80 FF=0.48
Deep gap, high IL
Unit Cell Geometry
Lattice a
Hole Radius
Plate Height
Tile this unit cell into a full plate using the current TR / FF / a₀.
Bandgap Analysis
From ML Neural Network Band Curves
Width
Lower
Upper
Center
Direct Predictions (from Grid/Nearest)
ML Regressor
COMSOL Nearest
Bandgap Summary
Gap Width (NN):
Center Freq (NN):
Gap Width (Regressor):
ML Predicted Band Structure
Bandgap Prediction Heatmap
Honeycomb Lattice & Unit Cell
Inverse Design Optimizer
Fixed material uses fast Mindlin solver. Custom materials use the 5-param CNN (slower).
kHz
kHz
⚙ Manufacturing Constraints
Feasibility Map
Shape Explorer — Fourier Descriptors
1.00
0.35
Area
Perimeter
Shape Factor
Eccentricity
Predicted Gap:
Crystal Plate Visualization
7
7
Unit Cell Geometry
Material Properties
Simulation Parameters
Honeycomb Plate — Topological Analysis
Honeycomb Phononic Crystal Plate — Topological analysis based on 3,000 COMSOL Multiphysics simulations of honeycomb lattice plates with cylindrical air holes (real 3D geometry, thickness h = TR × a). Neural network (R²=0.9995) predicts full 15-band structure in <5ms. Includes heuristic topology indicators, ML phase classification, and validated Dirac cone detection.
Current Design Prediction
Instant Topology Prediction for Current Design
Interpolates ML predictions to your current TR/FF values — updates automatically when you change parameters
Single Design — Heuristic Topology Analysis
Heuristic indicators from NN-predicted eigenfrequencies • True topological invariants require eigenvectors
Topological Indicators
Band Frequencies at High-Symmetry Points
Compares eigenfrequencies at Γ, K, M — band ordering changes indicate inversions
Zak Phase per Band
Estimated from band connectivity — quantized to 0 or π for inversion-symmetric systems (C₆v)
Berry Phase Indicators
Near-degeneracies with Dirac-like dispersion carry Berry phase ±π. Detected at high-symmetry points Γ, K, M.
Edge State Spectrum (Bulk-Boundary Correspondence)
Projected bulk bands (teal shading) with topological edge states (red/blue) in non-trivial gaps. Based on band inversion analysis.
Complete Bandgap Analysis
All bandgaps between adjacent bands (not just the largest)
Band Flatness Index
Flatness ratio = bandwidth / center frequency. Flat bands (F<0.05) indicate localization. Ref: Leykam et al. 2018
Van Hove Singularity Classification
Critical points where vg=0, classified by curvature: M₀ (minimum, DOS step-up), M₁ (saddle, log divergence), M₂ (maximum, DOS step-down). Ref: Van Hove, Phys. Rev. 89, 1189 (1953)
Spectral Flow Index
Counts net eigenvalue crossings through a reference frequency as a parameter sweeps — the only rigorous topological invariant computable from eigenfrequencies alone. Ref: Atiyah, Patodi & Singer (1975); Prodan & Schulz-Baldes (2016)
Band Evolution — Parameter Sweep
Uses current TR/FF from sidebar as fixed parameter
Full Design Space Maps (3,000 simulations)
Topological Phase
Dirac at K Probability (91.7% accuracy)
Dirac Cone Count (R²=0.841)
Number of validated Dirac cones per geometry — peak density marks topological transition boundary
Edge State Window
Frequency range supporting topologically-protected edge states — from bulk-boundary correspondence
Prediction Confidence
Model certainty — green = high confidence, red = uncertain (near phase boundaries)
Dirac Cone Count at K
Number of Dirac cones specifically at the K high-symmetry point — relevant for C₆ᵥ topological transitions
Design Space Exploration (full TR×FF grid)
Topological Phase Diagram
(scans entire TR×FF grid — may take ~15s)
Design Space Maps
(scans entire TR×FF grid — may take ~10s)
Band Touching Map (Dirac Cone Candidates)
Red regions = bands nearly touching along the k-path = potential Dirac points with Berry phase ±π.
2D Acoustic Model — Topological Phase Analysis
2D Scalar Acoustic Model — Plane wave expansion (PWE) solver for pressure waves in a 2D honeycomb lattice. Trained on 10,000 PWE simulations with TopologyNet v2 (multi-task, symmetry-aware). Phase accuracy: 97.8% | Bandgap R²: 0.976 | Valley Chern R²: 0.812. Allows rA ≠ rB to break C6v → C3v symmetry and open topological valley bandgaps.
Interactive Design Explorer
Geometry Parameters
0.200
0.200
Unit Cell Preview
Symmetry breaking: When rA = rB, the lattice has C6v symmetry with Dirac cones at K. Breaking rA ≠ rB opens a topological valley bandgap with non-trivial Valley Chern number.
Valley A: rA > rB → CK > 0
Valley B: rB > rA → CK < 0
Instant Topology Prediction
Interpolates ML predictions to your current rA/rB values — updates in real-time
CK & Gap vs rB (slice at current rA)
Full Design Space — 120×120 Phase Diagram
Topological Phase (rA vs rB)
■ Trivial   ■ Valley A (CK>0)   ■ Valley B (CK<0)   --- rA=rB (symmetry line)
Valley Chern Number CK (R²=0.812)
Red (+) = Valley A topology  |  Blue (−) = Valley B topology  |  White = trivial (CK≈0)
Bandgap & Transport Properties
Bandgap Width (R²=0.976)
Topological valley bandgap — opens when C6v symmetry breaks. Maximum gap away from the symmetry line.
Group Velocity |vg| at K (Band 3)
Linear dispersion near K point — vanishes at Dirac point (rA=rB), finite in gapped phase
Band Curvature at K (Band 3)
Second derivative of band dispersion — sign indicates effective mass. Related to Berry curvature in valley models.
Prediction Confidence
Model certainty (entropy-based) — lowest near the rA=rB phase boundary where topology changes
Extended Properties
Total Gap Coverage (R²=0.841)
Fraction of frequency range covered by all bandgaps — higher coverage = better isolation performance
Generate Custom Material Phase Diagram
Run TopologyNet v2 in-browser with custom material parameters. The pre-loaded plots above use fixed ρ-ratio=10, c-ratio=3. Enter your own values and hit Generate to compute a fresh 80×80 phase diagram.
Training range: 2.0 – 20.0
Training range: 0.5 – 8.0
Z ratio = ρ × c = 30.0 (impedance contrast) — higher Z = stronger phononic bandgaps
ModelTopologyNet v2 (in-browser)
Training Data10,000 PWE simulations
Phase Acc97.8%
Gap R²0.976
Chern R²0.812
Ext R²0.841
Refs: He et al., Nat. Phys. 12 (2016) | Lu et al., Nat. Phys. 13 (2017) | Wu & Hu, PRL 114 (2015)
Mindlin Plate Model — Topological Analysis
Mindlin Plate Model (3D) — Solves coupled flexural equations for deflection w, rotations ψx, ψy in a honeycomb phononic crystal plate with thickness h. Captures Lamb wave modes (A0, S0), shear deformation, and rotary inertia. Bandgap is thickness-tunable.
Ref: Hsu & Wu, Phys. Rev. B 74 (2006); Vila et al., Phys. Rev. B 96 (2017)
Interactive Design Explorer — Plate Thickness Sweep
Geometry & Thickness
0.200
0.200
Plate Thickness h/a = 0.80
Unit Cell Preview
Mindlin plate: 3N×3N eigenvalue problem coupling deflection w and rotations ψx, ψy. Plate thickness h/a controls the balance between flexural and shear modes.
Thin plates (h/a<0.3): flexural-dominated, narrow gaps
Moderate (h/a~0.5-0.8): optimal topological gaps
Thick plates (h/a>1): shear-dominated, gaps close
Instant Prediction
Interpolated from MindlinTopologyNet at the selected h/a slice
Gap & CK vs Thickness (at current rA, rB)
Phase Diagram at Selected Thickness
Topological Phase
■ Trivial   ■ Valley A   ■ Valley B
Valley Chern CK
Red (+) = Valley A | Blue (−) = Valley B | White = trivial
Bandgap & Transport Properties
Bandgap Width (R²=0.905)
Flexural bandgap — thickness-tunable. Optimal gap at moderate h/a.
Group Velocity |vg| at K
Wave transport speed near the K valley. Vanishes at the Dirac point.
Band Curvature at K
Effective mass indicator — sign change across the symmetry line.
Prediction Confidence
Model certainty — lowest near phase boundaries and at extreme h/a.
Extended Properties
Total Gap Coverage (R²=0.759)
Fraction of frequency range covered by all bandgaps in the plate.
Model Information
ModelMindlinTopologyNet
Physics3N×3N eigenvalue (w, ψx, ψy)
Training10,000 Mindlin PWE simulations
Phase Acc98.5%
Gap R²0.905
Topo R²0.489
Ext R²0.759
Materialρ-ratio=8.0, E-ratio=10.0
h/a slices0.2, 0.5, 0.8, 1.0, 1.3
Refs: Hsu & Wu, Phys. Rev. B 74 (2006) | Vila et al., Phys. Rev. B 96 (2017) | Mindlin, J. Appl. Mech. 18 (1951)
Sensitivity Analysis
Group Velocity
Density of States
Transmission Spectrum
10
SHAP Feature Attribution
Feature Importance
Design Atlas — t-SNE Projection
30
Uncertainty Quantification Bootstrap NN

Bootstrap perturbation of neural network weights to estimate prediction confidence. Adds Gaussian noise (σ=2% of weight std) to NN weights for N=30 iterations. Ref: Lakshminarayanan et al. 2017, NeurIPS

Uncertainty Heatmap (Mean σ across bands)
Manufacturing Tolerance / Robustness Analysis Fabrication-Ready

Perturb TR and FF within manufacturing tolerance bounds and assess bandgap survival probability. Ref: Sigmund et al. 2009, JASA; Wang et al. 2011

Multi-Objective Pareto Optimizer (NSGA-II) Evolutionary

Simultaneously optimize bandgap width, center frequency, and robustness using NSGA-II. Ref: Deb et al. 2002, IEEE TEC; Bilal et al. 2011, PRE

800 kHz
50
100
Pareto-Optimal Designs
Batch Prediction & Design of Experiments

Upload CSV

CSV with columns: TR, FF (one design per row). Optionally include a for custom lattice constant.


Generate DOE Samples

50
Gap Width Distribution
Results Scatter
Results Table
Dispersion Engineering vg Analysis

Group velocity, slow-wave regions, negative refraction detection, and effective mass analysis. Ref: Kushwaha 1993, PRB; Sukhovich et al. 2008, PRL

Group Velocity Map
Effective Mass (Band Curvature)
Effective Medium Properties Metamaterial

Retrieve effective mass density ρeff(ω) and bulk modulus Keff(ω) from band structure. Negative values indicate metamaterial behavior. Ref: Fokin et al. 2007, PRB; Li & Chan 2004, PRB

Effective Refractive Index
Attenuation / Insertion Loss Engineering dB

Engineering insertion loss in dB for N-layer phononic crystal. Ref: Hussein et al. 2014, Applied Mechanics Reviews

10
Layer Comparison (N=1,3,5,10,20)
Stopband Metrics
Waveguide / Defect Mode Preview

Supercell band folding approximation for line-defect waveguide modes. Ref: Khelif et al. 2004, PRB; Pennec et al. 2010, SSC

Waveguide Geometry
Waveguide Transmission
Finite Structure Mode Shapes

Approximate vibration mode shapes for a finite N×N phononic crystal plate using Bloch mode superposition.

6
6
800 kHz
Mode Animation (Harmonic Oscillation)
Frequency Response (Center Displacement vs. Frequency)
Frequency-Domain Wave Animation

Visualize wave propagation through the phononic crystal at selected frequency. Inside bandgap: evanescent decay. Outside: dispersive propagation. Ref: Laude 2015, Phononic Crystals (Textbook)

800 kHz
Wave Snapshot (Static)
Comparison Dashboard
Overlaid Band Structures
Side-by-Side Band Structures
Radar Comparison
Metrics Comparison
Active Learning — Next Simulation Suggestions

Identifies regions of highest model uncertainty to suggest optimal next FEA simulation points. Ref: Settles 2009, Active Learning Literature Survey

Suggested Simulation Points
Upload 3D Model (STL) → Auto-extract Height & Geometry NEW

Upload an STL file of one phononic-crystal unit cell (binary or ASCII). PhononIQ parses the mesh, computes the bounding box, and extracts plate height h, lattice constant a, thickness ratio TR = h/a, and fill fraction FF from a top-down projection. The 64×64 / 128×128 binary mask is staged on the CNN preview below — click Predict band structure to run inference. STL units are assumed millimetres.

…or drag & drop an STL here
Mesh preview (isometric)
Upload an STL to see bounding box, lattice constant, plate height, thickness ratio, and fill fraction.
What happens next: the extracted FF, TR, and lattice constant a auto-fill the geometry inputs in the panel below. The top-down rasterised mask appears in CNN-input panel ③. Material parameters (DR, EMR, PRR) still come from the matrix-material picker below — STLs don't carry that information.
Upload Image → Band Structure (PICNN)

Upload a binary unit-cell mask (dark = solid, light = hole/outside). The image is resampled to 64×64, thresholded, and run through the trained PhononCNNHybrid in your browser. The CNN learned on rhombic honeycomb cells with two circular holes — off-distribution images will give unreliable predictions. The image encodes FF only; the material parameters DR, EMR, PRR, TR enter through the parallel MLP branch below.

Model resolution:
① Original upload
(as you see it, full colour)
② Luminance binarised
(solid = dark pixels, pre-mask)
③ CNN input (64×64)
(after rhombus mask)
The CNN expresses material as ratios against BOROFLOAT 33 (ρ₀=2500 kg/m³, E₀=72 GPa, ν₀=0.23). Holes are always air. Pick a preset or enter your own ρ/E/ν below.
Ratios vs BOROFLOAT 33 (what the CNN actually sees): DR =  ·  EMR =  ·  PRR =
Training used a = √3 ≈ 1.732 mm. Frequencies scale as 1/a: halving a doubles the predicted bandgap centre frequency.
Predicted Band Structure & Group Velocity
Dispersion (ω vs k)
Group velocity vg = dω/dk (m/s)
Group velocity is the per-band slope of the dispersion curve, smoothed and computed segment-wise (Γ→K, K→M, M→Γ). Flat plateaus → slow modes; sign flips → backward / negative-refraction regimes.
3D PhononCNN (v5 — 64×64×64 voxel input · capacity-doubled) NEW · 808K params · R² 0.944

This is the v5 deployment of the 3D voxel model — capacity-doubled architecture (channel widths 24/32/48/64/96, FC head 384) trained for 100 epochs with PCA-30 target reparameterisation, PICNN pseudo-labels (1500 samples), mixup, EMA, and safe-symmetry augmentation. The cell geometry is built as a 643 binary voxel volume (1 = solid, 0 = void) with three additional CoordConv channels (x, y, z normalised to [−0.5, 0.5]) and fed through a 3D-conv encoder with squeeze-and-excitation blocks. The 28-feature material MLP runs in parallel; the fusion head predicts 30 PCA coefficients which are reconstructed inside the model graph to the full 15-band × 61-k-point dispersion. ONNX export 3.4 MB total. Best validation R² = 0.9442 (epoch 88 of 100), RMSE 175.4 kHz. v4 (lean, 280K params, R² 0.9429) and v1 (baseline, R² 0.9407) checkpoints retained on disk for ablation reference.

…or drag & drop an STL here. TR + FF auto-fill, mask drives inference.
Top-down mask (64×64) fed to the voxel encoder
No STL uploaded — using analytical (TR, FF) volume.
How this changes inference: when an STL is loaded, the 643 voxel volume is built by stacking the top-down mask along z up to z = TR·a. Without STL, the volume is built analytically from (TR, FF). The 28-feature material MLP branch is the same in both modes. Clear the STL to switch back to analytical mode.
Lattice constant a sets the physical scale. Training reference: a = √3 ≈ 1.732 mm. Frequencies scale as 1/a — halving a doubles the predicted band frequencies. STL upload auto-fills this from the XY bounding box; override here if STL units differ.
How this differs from the 2D PICNN page: the 2D model rasterises the cell to a single 64×64 image and lets the parallel MLP carry the plate height TR. The 3D model voxelises both — TR is encoded directly by how many z-layers are filled. Same dataset, different input encoding. R² is slightly lower (0.94 vs 0.98) because cylindrical-only training data has no extra 3D signal for the conv to learn — the architecture is built for the upcoming non-cylindrical (tapered / blind / stepped-hole) dataset.
Dispersion (ω vs k)
Bands coloured by vg = dω/dk (m/s) — blue = +, red = −
Group velocity computed segment-wise (Γ→K, K→M, M→Γ) via 5-point central differences, same routine the Dispersion Engineering page uses. Symmetric RdBu colour scale: white = 0 m/s, blue = forward modes, red = backward / negative-refraction regimes.
Embeddable Widget & API

Embed PhononIQ in your website or use the programmatic API for automated predictions.

Widget Configurator

Embed Code

Preview

JavaScript API

// Predict band structure
const result = PhononIQ.apiPredict({ tr: 1.0, ff: 0.35 });
console.log(result.bandgap.width); // kHz

// Batch prediction
const results = PhononIQ.apiBatchPredict([
  { tr: 0.5, ff: 0.30 },
  { tr: 1.0, ff: 0.35 },
  { tr: 1.5, ff: 0.40 }
]);

// Model info
const info = PhononIQ.apiGetModelInfo();
Bayesian Optimization GP + EI

Gaussian Process surrogate with Expected Improvement acquisition function. Intelligently suggests the next best simulation point to maximize bandgap discovery efficiency.

GP Mean (Bandgap Width)
GP Uncertainty (σ)
Expected Improvement
BO Convergence
Suggested Next Simulations
Physics-Informed Neural Network Analysis PINN

Quantifies how well the trained NN satisfies known physics constraints: Bloch–Floquet boundary conditions, band positivity, band ordering, and effective medium consistency.

Effective Medium Consistency
Multi-Fidelity Surrogate AR-1

Auto-regressive model combining cheap effective medium predictions (Bragg condition) with expensive NN predictions: yHF = ρ · yLF + δ(x). Demonstrates multi-fidelity acceleration.

Low-Fidelity (Bragg Estimate)
High-Fidelity (NN)
Multi-Fidelity Prediction
Residual (HF − MF)
Correlation Analysis
Graph Neural Network for Arbitrary Geometries GNN

Message-passing GNN that predicts band structure from mesh graph representation. Enables arbitrary unit cell geometries beyond parameterized cylinders.

Mesh Graph
Node Embeddings
GNN vs Standard NN Band Structure
GNN vs Nearest COMSOL Simulation
Architecture & Metrics
Multi-Lattice Explorer 3 Lattices

Compare phononic bandgaps across hexagonal, square, and triangular lattice symmetries. Uses physics-based scaling from the hexagonal NN model.

Unit Cell
Band Structure
Lattice Comparison (Bandgap vs FF)
2D Acoustic FDTD Sandbox FDTD

Full 2D finite-difference time-domain wave propagation through phononic crystal slabs. Visualize transmission, reflection, and evanescent decay in real time.

Generative Design (VAE) VAE

Variational Autoencoder learns the phononic crystal design space and generates novel designs meeting target bandgap specifications.

Latent Space (colored by gap width)
Training Loss
Generated Designs
Material Variation Explorer 5-Parameter
Material Properties
Density ρ (kg/m³)
kg/m³ (ratio: 1.00)
Young's Modulus E (GPa)
GPa (ratio: 1.00)
Poisson's Ratio ν
(ratio: 1.00)
Geometry
Thickness Ratio (h/a)
0.10 – 2.00
Filling Fraction (f)
0.25 – 0.50
Lattice Constant a (mm)
√3 ≈ 1.732 mm (scale: 1.00×)
Surrogate Model
PICNN runs a 64×64 deterministic raster through the CNN branch.
Dispersion (ω vs k)
Bands coloured by vg = dω/dk (m/s) — blue = +, red = −
Bandgap Landscape (TR vs FF slice, other params fixed)
Material Selection
PropertyValueUnit
Density2,230kg/m³
Young's Modulus64GPa
Poisson's Ratio0.200
Knoop Hardness480
Bending Strength150MPa
TypeBorosilicate glass

Currently supports BOROFLOAT® 33 borosilicate glass with air inclusions. Additional materials will be available in future versions.

Lattice Settings
ParameterValueDescription
a₀1.000 mmBase lattice constant
a = a₀√31.732 mmHoneycomb lattice constant

Frequency scale factor: 1.000× (default)

Data was computed with a = √3 mm. When using a custom lattice constant acustom, frequencies are scaled by √3 / acustom. Larger lattice → lower frequencies.

ML Model Information

Geometric Prediction (2-Parameter)


Material Variation (5-Parameter)

User Guide

Getting Started

PhononIQ is an ML-powered phononic bandgap prediction platform for honeycomb lattice phononic crystal plates. It uses neural networks trained on thousands of COMSOL Multiphysics simulations to predict band structures in real time.

1
Set Parameters

Use the sidebar inputs to set Filling Fraction (f) and Thickness Ratio (h/a) to any value within range.

2
Click Analyze

Press the ▶ Analyze button in the toolbar. The ML model predicts the full band structure instantly.

3
Explore Results

View band structure plots, bandgap analysis, sensitivity, topology, and more using the tree navigation on the left.

Module Overview

ModuleDescription
Analysis Main workspace: 3D unit cell visualization, ML band structure prediction, bandgap analysis, and prediction heatmap across the full parameter space.
Inverse Design Specify a target bandgap width and center frequency. The optimizer searches the parameter space using a genetic algorithm to find designs that match your target.
Shape Explorer Interactive 3D visualization of inclusion shapes using Fourier descriptors. Explore how different shapes affect the band structure.
Crystal Plate Full 3D phononic crystal plate visualization with honeycomb lattice arrangement. Supports isometric and top views with STL export.
Topology Band structure analysis: band inversion detection, band connectivity at high-symmetry points, band inversion map across parameter space, and isofrequency contour visualization.
Sensitivity SHAP-based feature attribution, sensitivity analysis of bandgap to parameter changes, density of states (DOS), transmission spectrum, and group velocity analysis.
Design Atlas t-SNE dimensionality reduction to map the full 3,000-design space into a 2D atlas. Color by bandgap width, center frequency, or parameters.
Material Explorer 5-parameter material variation explorer using a separate NN. Vary density, Young's modulus, Poisson's ratio, thickness, and filling fraction independently.

Detailed Module Guides

Analysis

The main workspace displays two side-by-side band structure plots:

  • Left plot — ML Neural Network: The neural network predicts band frequencies at any (TR, FF) value instantly. Bands are shown as solid colored lines. Bandgap regions are highlighted in green.
  • Right plot — COMSOL Nearest: Shows stored band data from the nearest COMSOL simulation in the dataset. Bands appear as solid orange lines. The nearest sample index and distance are displayed below.

The Bandgap Summary panel reports two independent predictions: the gap extracted from the NN band curves, and the gap from the standalone classifier+regressor ML models.

The Prediction Heatmap shows bandgap width across the entire TR×FF design space. Click any cell to jump to that design.

Inverse Design

Specify a target bandgap by entering desired width and center frequency. The optimizer uses a genetic algorithm (GA) to search the (TR, FF) parameter space:

  • Set target gap width (e.g. 100 kHz) and center frequency (e.g. 400 kHz)
  • Click Run Optimizer — the GA evolves a population of candidate designs over multiple generations
  • Results show the best-fit design(s) with their predicted band structures
  • A convergence plot tracks the fitness improvement over generations

Shape Explorer

Visualizes how the inclusion shape (the cross-section of the air hole) affects phononic behaviour using Fourier descriptors:

  • Adjust Fourier harmonic coefficients to morph the inclusion from a circle to more complex shapes (ellipses, polygons, stars)
  • The 3D viewer shows the deformed unit cell in real time
  • Explore how breaking circular symmetry modifies the band structure and bandgap

Crystal Plate

A full 3D phononic crystal plate assembled from multiple unit cells:

  • The plate is built by tiling rhombic unit cells along the honeycomb lattice vectors
  • Each unit cell contains two cylindrical air holes (honeycomb arrangement)
  • Toggle between 3D Isometric and Top-Down views
  • The adjacent panel shows a single isolated unit cell for reference
  • Hole size and plate thickness update automatically when you change TR/FF

Topology

Analyzes band ordering and connectivity of the phononic band structure:

  • Band Inversion Map: Shows where band ordering swaps between high-symmetry points across the (TR, FF) parameter space. Inversions are a necessary (but not sufficient) condition for topological transitions
  • Band Connectivity: Visualizes band frequencies at Γ, K, and M points. Bands involved in inversions are highlighted with dotted lines
  • Inversion Count: Heatmap of how many band pairs swap ordering at each (TR, FF) point
  • Isofrequency Contours: Band proximity map in (TR, FF) parameter space at a chosen frequency, showing which designs have bands near that frequency
  • Dirac Cones: Detected automatically where two bands nearly touch with linear dispersion (marked with circle indicators)

Sensitivity

Explores how the bandgap responds to parameter changes, with four sub-panels:

  • SHAP Waterfall: Shows how each input feature (TR, FF, and their engineered terms) contributes to the predicted bandgap width, based on SHapley Additive exPlanations
  • Group Velocity: The slope of each band (dω/dk) plotted as a function of k-point. Zero group velocity indicates standing waves; high values indicate fast propagation
  • Density of States (DOS): Histogram of available modes at each frequency. Bandgaps appear as regions with zero DOS
  • Transmission Spectrum: Estimated transmission through N layers of the crystal. Shows deep attenuation dips inside bandgaps

Design Atlas

A 2D map of the entire design space created using t-SNE dimensionality reduction:

  • Each dot represents one COMSOL simulation design (up to 500 subsampled from the 3,000 total)
  • Designs with similar properties (TR, FF, bandgap) cluster together on the map
  • Designs with different properties are placed far apart
  • Color the dots by: bandgap width, center frequency, TR, FF, or bandgap presence
  • Click any dot to load that design into the main Analysis page
Why is this useful? The atlas reveals the global structure of the design space: whether designs with large bandgaps form tight clusters or are scattered, which parameter regions are well-explored, and where phase boundaries or transitions might occur. It provides an intuitive overview that parameter sweeps alone cannot show.

Material Explorer (5-Parameter)

Extends the prediction to 5 independent parameters including material properties:

  • Adjust density (ρ), Young’s modulus (E), and Poisson’s ratio (ν) using real engineering units
  • Select from 16 material presets (steel, titanium, copper, PMMA, etc.) or enter custom values
  • Click Predict Band Structure to run the 5-parameter neural network
  • The Bandgap Landscape heatmap below shows bandgap width vs TR and FF for your chosen material
  • Click any cell in the heatmap to auto-fill sliders and predict that design
Note: This model uses a separate neural network (128→64→915) trained on 784 COMSOL simulations with BOROFLOAT® 33 glass as the base material. It explores how material substitution affects bandgap behaviour.

Parameter Definitions

ParameterSymbolDescriptionRange
Filling Fraction f Ratio of inclusion area to unit cell area. Controls the size of the cylindrical air holes within the plate. Higher values mean larger holes. 0.25 – 0.50
Thickness Ratio h/a Ratio of plate thickness (h) to lattice constant (a). Controls the out-of-plane dimension of the phononic crystal plate. 0.10 – 2.00
Lattice Constant a = a₀√3 Distance between nearest lattice points in the honeycomb arrangement. Default a₀ = 1 mm giving a = √3 mm. Changing this scales all predicted frequencies. > 0

Material Variation Parameters (5-Parameter Model)

ParameterSymbolDescriptionRange
Density ρ ρ (kg/m³) Inclusion material density. Base: BOROFLOAT 33 (ρ₀ = 2500 kg/m³). Covers soft polymers (~876) to dense metals (~5500). 876 – 5,500 kg/m³
Young's Modulus E E (GPa) Inclusion elastic modulus. Base: E₀ = 72 GPa. Covers flexible plastics (~22 GPa) to stiff ceramics (~144 GPa). 21.6 – 144 GPa
Poisson's Ratio ν ν Inclusion Poisson's ratio. Base: ν₀ = 0.23. Covers nearly incompressible (~0.46) to auxetic-like (~0.09). 0.09 – 0.46
Thickness Ratio h/a Same as geometric model: ratio of plate thickness to lattice constant. 0.10 – 2.00
Filling Fraction f Same as geometric model: ratio of inclusion area to unit cell area. 0.25 – 0.50

ML Predictions vs. COMSOL Reference

By default, PhononIQ displays ML-predicted band structures as solid colored lines. These predictions come from a neural network trained on COMSOL simulation data.

To compare with the nearest COMSOL simulation, enable "Show COMSOL Reference" in the sidebar. COMSOL data appears as dashed gray lines.

Why ML-primary? The neural network interpolates continuously across the parameter space, providing predictions at any (TR, FF) combination — not just the discrete points simulated in COMSOL. This demonstrates the value of the ML model for rapid exploration and design optimization.

Exporting Results

Use the File toolbar buttons to export:

  • SVG — Vector graphics of the current band structure plot (publication-ready)
  • PNG — Raster image of the current plot
  • CSV — Comma-separated band structure data for external analysis
  • LaTeX — Auto-generated LaTeX table snippet for papers

Data & Training

  • Geometric Model: Trained on 3,000 COMSOL simulations spanning TR [0.10, 2.00] × FF [0.25, 0.50], sampled via optimized Latin Hypercube Design (maximin criterion, seed 42). 15 frequency bands × 61 k-points each.
  • Material Model: Trained on 784 COMSOL simulations varying 5 material/geometric parameters (Density Ratio, Young’s Modulus Ratio, Poisson’s Ratio Ratio, TR, FF), using BOROFLOAT® 33 glass as base material with air inclusions. A 5,000-sample expanded dataset (LHS, maximin) is in preparation.
  • Band NN Architecture: Multi-layer perceptron (MLP) with 128→64 hidden units, feature engineering including quadratic cross-terms.
  • Brillouin Zone Path: Γ → M → K → Γ (61 k-points along the irreducible path for honeycomb lattice)
Theory & Governing Equations

Phononic Crystals

Phononic crystals (PnCs) are periodic structures that control the propagation of elastic and acoustic waves. When waves interact with the periodic arrangement, Bragg scattering can create bandgaps — frequency ranges in which wave propagation is forbidden.

PhononIQ focuses on honeycomb lattice phononic crystal plates: thin plates with a periodic array of cylindrical air holes arranged in a honeycomb pattern.

Governing Equations

Elastic Wave Equation

The propagation of elastic waves in a solid is governed by the Navier-Cauchy equation:

ρ ∂²u/∂t² = ∇ · σ

where u is the displacement vector, ρ is the mass density, and σ is the stress tensor.

Constitutive Relation (Hooke's Law)

For a linear elastic isotropic material:

σij = λ δij εkk + 2μ εij

where λ and μ are the Lamé parameters:

λ = Eν / [(1+ν)(1−2ν)]     μ = E / [2(1+ν)]

Strain-Displacement Relation

εij = ½(∂ui/∂xj + ∂uj/∂xi)

Bloch-Floquet Periodicity

For a periodic structure with lattice vectors a1 and a2, the displacement field satisfies:

u(r + R) = u(r) eik·R

where k is the Bloch wave vector and R = n1a1 + n2a2 is a lattice translation vector. This reduces the eigenvalue problem to a single unit cell.

Brillouin Zone & Band Structure

The hexagonal Brillouin zone for a honeycomb lattice has three high-symmetry points:

PointCoordinatesDescription
Γ(0, 0)Zone center
M(π/a, 0)Edge midpoint
K(4π/3a, 0)Zone corner (Dirac point)

The band structure is computed along the irreducible Brillouin zone path Γ → M → K → Γ, sampling 61 k-points. Each eigenfrequency ωn(k) forms a band, and frequency ranges with no bands constitute bandgaps.

Bandgap Definition

A complete bandgap exists between bands n and n+1 when:

mink ωn+1(k) > maxk ωn(k)

The bandgap is characterized by:

  • Lower edge: maxk ωn(k) — top of band n across all k-points
  • Upper edge: mink ωn+1(k) — bottom of band n+1 across all k-points
  • Width: Δω = upper edge − lower edge
  • Center frequency: ωc = (upper + lower) / 2
  • Relative width: Δω / ωc × 100%

Topological Acoustics — Overview

Topological phononic crystals exhibit symmetry-protected edge states that propagate along interfaces without backscattering, even in the presence of defects. In a 2D honeycomb lattice (C6v symmetry), topological behaviour arises from band inversions at high-symmetry points of the Brillouin zone.

PhononIQ computes surrogate topological indicators from the ML-predicted band structure. These are physically motivated proxies that use band ordering and connectivity analysis at high-symmetry points (Γ, K, M). True topological invariants (Chern number, Z2) require eigenvectors and full Brillouin zone integration, which are unavailable from our frequency-only NN model.

Physical basis: For the honeycomb lattice with C6v symmetry, the relevant topological framework is the quantum valley Hall effect (QVHE). When C6v is broken down to C3v (e.g., by detuning the two sublattice holes), valleys at K and K′ acquire opposite Berry curvature, enabling valley-polarized edge states.

1. Band Inversion Detection

A band inversion occurs when the ordering of two adjacent bands swaps between high-symmetry points. This is the primary indicator of a topological phase transition.

Method

For each adjacent band pair (n, n+1), we extract frequencies at high-symmetry points and compute the gap:

ΔfΓ = fn+1(Γ) − fn(Γ),    ΔfK = fn+1(K) − fn(K),    ΔfM = fn+1(M) − fn(M)

Detection criteria:

  • Gap closure: mink |fn+1(k) − fn(k)| < 5 kHz along the k-path → bands nearly touch (potential Dirac point)
  • Slope convergence: Adjacent bands approach each other with converging slopes and gap < 20 kHz → near-crossing detected
  • Sign change: If Δf changes sign between two symmetry points, the band ordering has swapped → band inversion confirmed

Z2 Index (Surrogate)

The surrogate Z2 index is computed from the parity of band inversions at Γ:

Z2 = (number of inversions at Γ) mod 2

Z2 = 1 (odd inversions) → topologically non-trivial phase
Z2 = 0 (even inversions) → topologically trivial phase

Limitation: The true Z2 invariant requires computing the Pfaffian of the time-reversal operator acting on Bloch eigenstates at TRIM points (Fu & Kane, 2006). Our surrogate uses band-ordering parity as a proxy, which is valid when band inversion directly drives the topological transition (as in the QVHE honeycomb case).

2. Zak Phase Estimation

The Zak phase is a 1D topological invariant that measures the geometric phase acquired by a Bloch state transported across the entire Brillouin zone:

θnZak = ∮BZ An(k) dk = ∮BZ i⟨unk | ∇k | unk⟩ dk

where An(k) is the Berry connection and |unk⟩ is the periodic part of the Bloch function. For systems with inversion symmetry, the Zak phase is quantized to 0 or π.

Surrogate Method (Connectivity Analysis)

Without eigenvectors, we estimate the Zak phase from band connectivity along the k-path:

  1. Scan each band n along the k-path for avoided crossings with adjacent bands (n−1, n+1)
  2. An avoided crossing is detected when: |fn+1(k) − fn(k)| < 8 kHz and the gap doesn’t exceed 2.5× this threshold in the neighborhood
  3. Count the number of such crossings for each band
  4. Quantize: θZak = π if crossing count is odd, θZak = 0 if even
Physical basis: In systems with inversion symmetry, the Zak phase relates to the Wannier center position. A Zak phase of π indicates the Wannier center is at the unit cell boundary, while 0 indicates it is at the center. Each avoided crossing along the k-path changes the connectivity, flipping the Zak phase by π (Xiao, Chang & Niu, Rev. Mod. Phys. 2010).

3. Berry Phase at Dirac Points

The Berry phase is the geometric phase acquired by a state adiabatically transported around a closed loop in k-space. Near a Dirac cone, the Berry phase is quantized to ±π:

γC = ∮C A(k) · dk = ±π    (around a Dirac point)

Detection Method

We identify Berry phase ±π at locations exhibiting Dirac-like conical dispersion:

  1. Scan high-symmetry points (Γ, K, M) for inter-band gaps < 20 kHz
  2. Test the neighborhood (dk = 4 k-points) for V-shaped or Λ-shaped dispersion:
    • Upper band minimum: slopes on both sides are negative (band dips toward the point)
    • Lower band maximum: slopes on both sides are positive (band rises toward the point)
  3. If both conditions hold → Dirac cone detected, Berry phase = ±π
  4. If the gap < 3 kHz (near-degeneracy) at a symmetry point, it is automatically flagged
Physical significance: The ±π Berry phase at K and K′ points is the hallmark of the quantum valley Hall effect in honeycomb lattices. When inversion symmetry is broken, the Dirac cone opens into a gap with opposite Berry curvature at K vs K′, enabling topologically protected valley-polarized edge states (Lu et al., Nature Physics 2017).

4. Edge State Prediction (Bulk-Boundary Correspondence)

The bulk-boundary correspondence principle states that a topologically non-trivial bulk gap must host protected edge states at interfaces between topologically distinct domains.

Method

  1. Find bandgaps: Identify all gaps with width > 2 kHz from the bulk band structure
  2. Classify each gap: Check for band inversions in the bands bounding the gap. If the upper and lower bands show slope convergence at Γ (consistent with a band inversion), the gap is classified as topological
  3. Generate edge dispersion: For topological gaps, construct a synthetic edge state that crosses the gap linearly:
    fedge(k) = fcenter + vedge · (k − k0)
    where fcenter is the gap midpoint, vedge is the edge state velocity, and k0 is the crossing point
  4. Penetration depth: Estimated as:
    ξ ≈ 2 fcenter / (Δf + 1)
    This gives the e-folding decay length of the edge state into the bulk. The minimum number of unit cells for confinement is Nmin ≈ 3ξ.
Limitation: The edge spectrum shown is a synthetic visualization based on bulk-boundary correspondence, not a supercell ribbon calculation. A rigorous edge state computation requires solving the eigenvalue problem for a semi-infinite strip (supercell with N unit cells in one direction), which is beyond the scope of the ML model. The synthetic edge states are qualitatively correct for QVHE-type systems but do not capture details like edge state dispersion curvature or localization profiles.

5. Topological Phase Diagram

The phase diagram maps the surrogate Z2 index and inversion count across the full (TR, FF) design space:

Construction

  1. For each (TR, FF) grid point, run the ML model to predict the full 15-band structure
  2. Apply band inversion detection at each point
  3. Record: Z2 index, total inversion count, and minimum gap at Γ
  4. Plot as heatmaps: Z2 = 1 regions are non-trivial phases, phase boundaries occur where gaps close at Γ

Band Touching Map

A complementary map shows the minimum inter-band gap at each (TR, FF) point. Regions where bands nearly touch (< 5 kHz) are potential Dirac points where topological transitions occur. These gap-closing lines trace the topological phase boundaries in the design space.

Design guidance: To engineer a topological phononic crystal, choose (TR, FF) in a non-trivial (Z2 = 1) region, then create an interface with a design from the trivial (Z2 = 0) region. The phase diagram helps identify which parameter changes drive topological transitions.

6. Accuracy Summary & References

Indicator Method Accuracy
Band Inversion Gap comparison at Γ, K, M Reliable — direct from band structure
Z2 Index Parity of inversions at Γ Surrogate — proxy for true Z2
Zak Phase Avoided crossing count (quantized to 0/π) Heuristic — based on connectivity, not integration
Berry Phase Dirac cone shape detection at K, M Qualitative — identifies Dirac-like features
Edge States Synthetic from bulk-boundary correspondence Illustrative — not from supercell calculation
Phase Diagram Full (TR, FF) sweep of inversion analysis Reliable — maps where bands close/invert

Key References

  • Lu, J. et al., “Observation of topological valley transport of sound in sonic crystals,” Nature Physics 13, 369 (2017)
  • Xiao, D., Chang, M.-C. & Niu, Q., “Berry phase effects on electronic properties,” Rev. Mod. Phys. 82, 1959 (2010)
  • Miniaci, M. et al., “Experimental observation of topologically protected helical edge modes in patterned elastic plates,” Phys. Rev. X 8, 031074 (2018)
  • He, C. et al., “Acoustic topological insulator and robust one-way sound transport,” Nature Physics 12, 1124 (2016)
  • Fu, L. & Kane, C.L., “Topological insulators with inversion symmetry,” Phys. Rev. B 76, 045302 (2007)

ML Model Architecture

Feature Engineering

For the 2-parameter geometric model, input features are derived from (TR, FF):

  • Raw features: TR, FF
  • Quadratic: TR², FF², TR×FF
  • Transforms: √TR, √FF, log(TR), log(FF)

For the 5-parameter material model, 28 features are engineered from (DR, EMR, PRR, TR, FF) including all pairwise cross-terms and nonlinear transforms.

Neural Network (Band NN)

Architecture: MLP with 2 hidden layers (128 → 64 neurons), ReLU activation, outputting 15×61 = 915 frequency values (15 bands, 61 k-points each).

Input (features) → Dense(128, ReLU) → Dense(64, ReLU) → Dense(915, Linear) → Reshape(15, 61)

Bandgap Classification & Regression

A two-stage ML pipeline is used for bandgap prediction:

  1. Stage 1 — Classifier: Random Forest or Gradient Boosting classifier predicts whether a bandgap exists (binary).
  2. Stage 2 — Regressor: For samples classified as having a bandgap, a Gradient Boosting regressor predicts the bandgap width and center frequency.

Group Velocity

The group velocity describes how fast energy (a wave packet) propagates through the crystal at each k-point:

vg,n(k) = ∇k ωn(k)

Along the 1D k-path Γ→K→M→Γ, this reduces to the slope dω/dk.

Numerical Method

PhononIQ computes group velocity using 5-point central finite differences (O(h&sup4;) accuracy) within each Brillouin zone segment independently:

vg(k) = [−f(k+2h) + 8f(k+h) − 8f(k−h) + f(k−2h)] / (12h)

At segment edges (near K and M symmetry points), a 3-point central difference is used:

vg(k) = [f(k+h) − f(k−h)] / (2h)

Values at exact symmetry points (Γ, K, M) are set to NaN because the path direction changes discontinuously there.

Physical Conversion to m/s

vg [m/s] = (df [kHz] × a0) / (dknorm × (1 + 1/√3))

where a0 is the lattice constant. This factor converts from normalized k-space derivatives to physical velocities. Pre- and post-smoothing (moving average) suppresses NN prediction noise.

Physical Insights

  • Zero group velocity at band edges and flat bands — waves are standing, no energy transport
  • High group velocity — fast propagation; linear dispersion (vg = const) implies non-dispersive behaviour
  • Inside bandgaps, no propagating modes exist — group velocity is undefined
  • Near Dirac cones, vg is approximately constant (linear dispersion), analogous to massless Dirac fermions
  • Slow-wave regions (|vg| < 5% of max) are identified automatically — relevant for enhanced acoustic energy density and sensing (Baba, Nature Photonics 2, 465, 2008)
Note: Group velocities are computed from NN-predicted band frequencies, not directly from COMSOL eigenvectors. The numerical derivatives amplify any small prediction noise, so a smoothing kernel is applied. Results are qualitatively accurate but may differ from exact COMSOL eigenvalue derivatives by a few percent.

Effective Mass (Band Curvature)

The effective mass of a phononic mode is inversely proportional to the curvature of the dispersion relation:

m* = ℏ² / (d²ω/dk²)

This quantity governs wave packet spreading, tunneling through barriers, and the response to perturbations.

Numerical Method

Computed via 5-point central second derivative (O(h&sup4;) accuracy):

d²f/dk² = [−f(k+2h) + 16f(k+h) − 30f(k) + 16f(k−h) − f(k−2h)] / (12h²)

At segment boundaries, a 3-point stencil is used:

d²f/dk² = [f(k+h) − 2f(k) + f(k−h)] / h²

Physical Insights

  • Positive curvature (m* > 0) — normal dispersion, wave packets spread normally
  • Negative curvature (m* < 0) — anomalous dispersion, enables negative refraction (Sukhovich et al., PRL 102, 154301, 2009)
  • Near-zero curvature (|m*| → ∞) — flat bands, heavy effective mass, strong localization
  • Infinite curvature (m* → 0) — at Dirac cones, analogous to massless particles

The plot displays sign(m*) × log10(|1/curvature| + 1) to compress the dynamic range for visualization.

Density of States (DOS)

The phononic density of states counts the number of modes available at each frequency:

g(ω) = ∑nk δ(ω − ωn(k))

In practice, this is approximated using Gaussian broadening or binning the eigenfrequencies into histogram bins. The DOS drops to zero inside bandgaps, providing a complementary visualization of forbidden frequency ranges.

Transmission Spectrum

The transmission through N layers of the phononic crystal is estimated using the transfer matrix method. The transmission coefficient drops exponentially inside bandgaps:

T(ω) ∝ e−Nκ(ω)a

where κ(ω) is the imaginary part of the wave vector (evanescent decay rate) inside the bandgap. More layers produce deeper attenuation in the gap region.

Finite Structure Mode Shapes

When a phononic crystal has finite extent (Nx × Ny unit cells), the allowed wavevectors are quantized and the response depends on whether the excitation frequency falls inside or outside a bandgap.

In-Gap: Evanescent Modes

Inside a bandgap, no propagating Bloch modes exist. Waves launched from the boundary decay exponentially toward the interior (Laude, Phononic Crystals, De Gruyter, 2015, Sec. 5.4):

u(x,y) ~ exp(−κ · dmin)

where dmin is the distance from the nearest boundary (in unit cells) and κ is the evanescent decay rate:

κ = (π/a) √(1 − (Δf / (Δfgap/2) − 1)²)

Decay is strongest at the gap center and weakest near band edges, consistent with complex band structure theory.

Pass Band: Standing Waves

Outside bandgaps, finite-size boundary reflections create standing wave patterns. For a plate with fixed-fixed boundaries:

u(x,y) = A · sin(nxπx/Lx) · sin(nyπy/Ly) · [1 + ε cos(2πx/a) cos(2πy/a)]

where (nx, ny) are the mode indices determined by the band index, and the cosine term represents the sub-cell Bloch modulation.

Frequency Response Function (FRF)

The displacement amplitude at the plate center as a function of frequency is computed from the Green’s function approach (Vasseur et al., PRL 86, 3012, 2001):

|u(f)|² ~ ∑n 1 / √((f² − fn²)² + (η f fn)²)

where fn are the natural frequencies of the finite structure (quantized from the band structure) and η is a structural loss factor. The FRF shows resonance peaks at pass-band frequencies and deep dips (orders of magnitude attenuation) inside bandgaps.

References: Laude, V., Phononic Crystals, De Gruyter (2015), Ch. 5 & 8; Khelif, A. & Adibi, A. (eds.), Phononic Crystals: Fundamentals and Applications, Springer (2016), Ch. 8; Vasseur, J.O. et al., PRL 86, 3012 (2001).

Phononic Crystal Waveguide

A W1 waveguide is formed by removing one row of holes from a perfect phononic crystal, creating a line defect that can guide acoustic/elastic waves at frequencies within the bulk bandgap (Khelif et al., APL 84, 4400, 2004).

Supercell Band Folding

To analyze a waveguide, a supercell of N unit cells perpendicular to the propagation direction is used. This folds the Brillouin zone N times:

kfolded,m = (k + 2πm/N) mod (π/a),    m = 0, 1, ..., N−1

Each bulk band generates N sub-bands. PhononIQ approximates the folded spectrum by shifting and interpolating the ML-predicted bulk dispersion, with small perturbations to simulate anti-crossing (mode coupling) at zone boundaries (Laude, 2015, Ch. 4).

Guided Defect Mode

Inside the complete bandgap, the line defect introduces a guided mode. Its frequency depends on the effective medium of the empty channel:

fdefect ≈ fgap,center − Δfshift

where Δfshift is proportional to the filling fraction (more material removed → lower frequency). The guided mode has a nearly flat dispersion with slight positive curvature, characteristic of slow-light operation near band edges (Pennec et al., Physica Status Solidi (c) 6, 2080, 2009).

Transmission Spectrum

The transmission through the waveguide shows:

  • Near 0 dB at the guided mode frequency — efficient wave transport through the defect channel
  • Lorentzian peak shape centered on fdefect with linewidth γ determined by confinement:
    T(f) = Tmax / (1 + ((f − f0) / γ)²)
  • Strong attenuation (< −30 dB) elsewhere in the gap — evanescent regime
  • Partial transmission with Fabry-Pérot fringes outside the gap — interference from multi-reflection in the finite crystal
Approximation note: The waveguide analysis uses a parametric model based on zone folding and Lorentzian transmission, not a full supercell eigenvalue calculation or FDTD simulation. The defect mode frequency and bandwidth are estimated from the bulk bandgap properties. The transmission values outside the guided mode bandwidth are qualitative approximations. For quantitative waveguide design, a full supercell FEM calculation (e.g., COMSOL) is recommended.
References: Khelif, A. et al., APL 84, 4400 (2004); Laude, V., Phononic Crystals, De Gruyter (2015), Ch. 10; Pennec, Y. et al., Physica Status Solidi (c) 6, 2080 (2009).

Effective Medium Properties

At long wavelengths (λ ≫ a), the phononic crystal behaves as a homogeneous effective medium with frequency-dependent properties:

Effective Density

ρeff(ω) = ρhost(1 − FF) + ρincl · FF · [1 + δρ(ω)]

where FF is the filling fraction and δρ accounts for resonant corrections near bandgap frequencies.

Effective Bulk Modulus

1/Keff = (1 − FF)/Khost + FF/Kincl

Effective Refractive Index

neff(ω) = chost / ceff(ω) = chost √(ρeff / Keff)

The refractive index becomes imaginary inside bandgaps (evanescent regime) and can become negative in anomalous dispersion regions.

References: Torrent, D. & Sánchez-Dehesa, J., New J. Phys. 10, 063015 (2008); Fokin, V. et al., Phys. Rev. B 76, 144302 (2007); Milton, G.W. et al., New J. Phys. 8, 248 (2006).

SHAP Feature Attribution

SHAP (SHapley Additive exPlanations) is a game-theoretic method for explaining individual ML predictions. For each prediction, SHAP assigns a contribution value to every input feature:

f(x) = φ0 + ∑i=1M φi

where φ0 is the baseline (average prediction), and φi is the SHAP value for feature i. The sum of all SHAP values equals the difference between the prediction and the baseline.

In the Sensitivity page, the SHAP waterfall chart shows:

  • Each bar = the contribution of one feature (TR, FF, TR², FF², TR×FF, etc.) to the predicted bandgap width
  • Green bars push the prediction higher (larger bandgap)
  • Red bars push the prediction lower (smaller bandgap)
  • The features are sorted by magnitude, so the most important features appear at the top

This tells you why a particular design has a large or small bandgap, and which parameters have the most influence.

Inverse Design & Optimization

Traditional design (forward problem): given (TR, FF) → predict bandgap. Inverse design reverses this: given a target bandgap → find the (TR, FF) that produces it.

PhononIQ uses a genetic algorithm (GA) for this optimization:

  1. Initialize: Create a random population of N candidate designs (TR, FF pairs)
  2. Evaluate: For each candidate, use the ML model to predict the bandgap and compute a fitness score based on how close the predicted gap matches the target
  3. Select: Tournament selection — pick the best individuals to become parents
  4. Crossover: Combine parameters from two parents to create offspring
  5. Mutate: Apply small random perturbations to maintain diversity
  6. Repeat: Over many generations, the population converges toward the optimal design

The fitness function minimizes:

F = w1|Δωpred − Δωtarget| + w2c,pred − ωc,target|

where Δω is bandgap width and ωc is center frequency.

t-SNE Dimensionality Reduction (Design Atlas)

t-SNE (t-distributed Stochastic Neighbor Embedding) is an algorithm that maps high-dimensional data to 2D while preserving local neighborhood structure. It works in three steps:

Step 1: Pairwise Similarities in High-D

For each pair of designs (i, j), compute the probability that i would pick j as a neighbor using a Gaussian kernel:

pj|i = exp(−||xi − xj||² / 2σi²) / ∑k≠i exp(−||xi − xk||² / 2σi²)

The bandwidth σi is set per point so each point has an effective neighborhood size called perplexity (~30 neighbors). Symmetrize: pij = (pj|i + pi|j) / 2N.

Step 2: Pairwise Similarities in 2D

In the 2D embedding, use a heavier-tailed Student-t distribution (1 degree of freedom):

qij = (1 + ||yi − yj||²)−1 / ∑k≠l (1 + ||yk − yl||²)−1

The heavy tail allows moderate high-D distances to map to larger 2D distances, reducing the “crowding problem”.

Step 3: Minimize KL Divergence

Move the 2D points by gradient descent to minimize the Kullback-Leibler divergence between p and q:

KL(P||Q) = ∑i≠j pij log(pij / qij)

This ensures that points close in the original 5D space (similar TR, FF, bandgap) remain close in the 2D map.

Implementation note: PhononIQ uses a vanilla t-SNE with adaptive gains and early exaggeration. To keep the browser responsive, the algorithm subsamples to 500 designs (stratified by bandgap presence) from the full 3,000-design dataset. The perplexity can be tuned with the slider (default 30).

Fourier Shape Descriptors

The Shape Explorer uses Fourier descriptors to parameterize the cross-sectional shape of the inclusion (air hole). Any closed 2D curve r(θ) can be decomposed as:

r(θ) = a0 + ∑n=1N [an cos(nθ) + bn sin(nθ)]

where a0 is the mean radius (related to filling fraction), and the higher harmonics control the shape:

HarmonicEffect
n = 0Mean radius (circle size)
n = 2Elliptical distortion
n = 3Triangular perturbation
n = 4Square-like perturbation
n = 6Hexagonal symmetry (matches lattice)

By varying these coefficients, you can explore how breaking the circular symmetry of the holes opens, closes, or shifts the bandgap.

Dirac Cones

A Dirac cone is a point where two bands touch with linear (conical) dispersion, forming a degeneracy analogous to massless Dirac fermions in graphene. Near the Dirac point:

ω(δk) ≈ ωD ± vD|δk|

where ωD is the Dirac frequency and vD is the Dirac velocity.

PhononIQ detects Dirac cones automatically by checking four conditions between adjacent bands:

  1. Small gap: The frequency gap between two bands is less than 3 kHz at that k-point
  2. High frequency: The touching point is above 50 kHz (avoids false positives at low frequencies)
  3. Local minimum: The gap at that k-point is a local minimum compared to neighboring k-points
  4. Opposing slopes: The two bands have opposite slopes (one increasing, one decreasing) with magnitude exceeding 200 kHz per k-unit, confirming the conical (linear) dispersion

Dirac cones are marked with ○ circle indicators on the band structure plot. They indicate points where two bands nearly touch with linear dispersion, which is a necessary condition for topological phase transitions.

Credits & Acknowledgments

© NewFOS — New Frontiers of Sound

Developed by Samarjith Biswas, PhD

Supported by the National Science Foundation (NSF) Engineering Research Center for New Frontiers of Sound (NewFOS).
University of Arizona • Caltech • UCLA • Georgia Tech • CUNY • Wayne State • Spelman College • UC Boulder • CU Denver • University of Alaska Fairbanks

Noise / Vibration Filter Designer

Enter a frequency range you want to block and physical constraints. The tool finds the best phononic crystal designs with full fabrication specs.

CNC/Laser limit
Feasibility Map
Design Comparison