# Research study and publication plan

Proposed work for a CS portfolio project — 8 October 2026.

This document defines proposed experiments and publication routes. As of 9 October 2026, the browser demonstrator has local tagged versions and implementation QA, and the [technical report](research-report.md) records a completed modal-reconstruction and refinement comparison. No participant study, calibrated plate experiment, public deposition, or scholarly submission is asserted. The [literature review](literature-review.md) provides the original scoping context; the [model specification](mathematical-models.md) distinguishes proposed models from dated implementation notes.

## Choose one principal contribution

Two directions are feasible. They can share code, but the first paper should have one primary question.

### Direction A: physical dynamics and musical responsiveness

Working title:

**Physical fidelity and musical responsiveness in browser based Chladni visualization**

Question: how do different audio couplings affect the numerical response, perceived synchrony, visual coherence, and aesthetic experience of the same plate scene?

The contribution would be a clearly specified comparison, a browser implementation, validation of its computational components, and evidence about a tradeoff. It would not be the invention of Chladni figures, sound reactive particles, or music visualization.

“Fidelity” must identify its reference. Without measurements of a real plate, we can measure consistency with a specified mathematical model and a trusted numerical reference. We cannot infer accuracy for arbitrary physical hardware.

Prior visualization systems establish relevant design and evaluation precedents. [CymaSense](https://doi.org/10.1145/3064857.3079159), [Graf et al.](https://arxiv.org/abs/2106.10134), [Erdmann et al.](https://www.frontiersin.org/journals/virtual-reality/articles/10.3389/frvir.2025.1552321/full).

### Direction B: stochastic particles and continuum density

Working title:

**Comparing particle and density representations of sound driven granular patterns in the browser**

Question: how accurately and efficiently can two computational representations realize the same stated stochastic transport model?

This direction can yield a contained CS study without participant recruitment. Evaluate density error, mass conservation, convergence, memory, and computational cost. Its physical interpretation is limited by the transport model's validation domain.

A faster renderer alone does not establish a new numerical method. A contribution might instead be a careful implementation comparison, a reproducible benchmark, or a justified approximation with measured errors. The related stochastic and continuum literature needs positioning before making a novelty claim. [Arango and Reyes](https://doi.org/10.1017/S0013091515000139), [Grabec and Sok](https://www.sv-jme.eu/article/diffusion-equation-generalized-for-modeling-of-chladni-patterns/).

Our recommendation is Direction A if a supervised user study is practical; otherwise begin with Direction B and treat the perceptual comparison as later work.

## Controlled comparison for Direction A

Keep plate geometry, mode bank, renderer, camera, canvas size, and monochrome style fixed.

| Condition | Audio coupling | Intended interpretation |
| --- | --- | --- |
| P: specified dynamics | Waveform drives damped modal equations; a stated transport law moves grains | A physical model driven through an assumed transducer |
| A: direct musical control | Features select or blend patterns and control a designed particle rule | Mathematical art optimized for a chosen response |
| H: declared hybrid | Retains modal dynamics while adding a small, documented musical control | A simulation with explicit aesthetic modifications |

Initially keep the transport law identical across P, A, and H if the study question is about audio coupling. Compare attraction and diffusion transport in a separate experiment. Combining coupling and transport changes in every condition would make the observed differences difficult to attribute.

Record every modification in H. For example, a gain boost, a frequency remapping, and a faster particle timescale are distinct interventions. Change one at a time in early experiments. If a later preset combines them, report their composition.

Use the same fixed rendering transfer function. Choose reference gains using predetermined calibration signals and then freeze them. Do not independently normalize every condition until all motion looks equally strong; that could erase the effect under investigation.

### Hypotheses to test, not conclusions

- A may improve immediate perceived response to attacks.
- P may show more persistence, frequency selectivity, and continuity under slowly changing input.
- H may improve perceived synchrony while retaining some model structure.
- Greater motion or geometric complexity may help only over part of its range.
- Perceived synchrony, beauty, and understanding may change independently.

These are our proposed hypotheses. They should be refined with a pilot and registered before a confirmatory study. A hybrid is not assumed to be superior.

### Useful controls

Add a time shifted version of a condition to test synchrony, using predefined shifts. Use an audio independent visual control to test whether mere motion is enough. Include a simple waveform or spectrogram as a measurement baseline.

A phase intervention can preserve individual frequency magnitudes while changing coherent superposition, but its construction must be defined. Randomizing phases or shifting visuals can change more than one perceptual cue. Controls should document what is preserved and what is deliberately changed.

## Validate the model before evaluating its appearance

### Plate response

Use the analytically solvable rectangle first.

1. Verify mode frequencies, shapes, normalization, and boundary conditions.
2. For a numerical eigenproblem, report residuals and convergence with refinement.
3. Compare repeated eigenmodes as subspaces rather than demanding one particular eigenvector orientation.
4. Check sinusoidal gain and phase against the modal response formula.
5. Check free decay and linear superposition.
6. Test in-phase and opposite-phase excitation.
7. Measure the effect of mode truncation.

For a chosen centrally mounted plate, supports and forcing should be modeled separately. Comparison with laboratory photos requires dimensions, material, mounting, driving conditions, and image calibration. The three supplied photos are inspiration, not sufficient quantitative ground truth.

### Particle or density transport

Use a static, positive diffusivity field first. Compare the stationary density with the stated analytical equilibrium. Check conserved grain number and reflecting boundary behavior.

For transient tests, refine timestep, grid, particle number, and spatial interpolation. Run multiple random seeds. Compare particle histograms with continuum density at the same physical times. Separate finite particle sampling noise from solver discretization error.

A visually convincing picture is not a substitute for these checks. Conversely, agreement with a mathematical reference does not independently validate the transport law for real sand.

### Optional physical experiment

If laboratory access becomes available, begin with controlled sine sweeps and isolated resonances. Record geometry and material, mounting, excitation location, frequency, amplitude measurement where possible, grain properties, and time to settle.

Image comparisons could use nodal set distance or a carefully defined overlap metric. Compare frequencies and transient behavior as well as a final pattern. Quantify alignment and threshold sensitivity. Do not adjust the simulated support or contrast separately for each photo until it matches.

## Benchmark the browser implementation

Record CPU and GPU, operating system, browser version, backend, canvas resolution, grid size, particle count, number of modes, audio configuration, and quality setting.

| Measure | What to report |
| --- | --- |
| Frame timing | Median and p95 frame time, dropped frames, and run duration |
| Simulation cost | Per-step CPU/GPU time, update rate, and numerical settings |
| Memory | Relevant buffer sizes and observed memory limits |
| Numerical quality | Error or convergence at each quality setting |
| Audio alignment | Feature timestamp, availability, simulation update, and presentation delay |
| Model response | Resonance rise/decay and grain redistribution times |
| Robustness | Silence, clipping, long files, resize, seek, and unsupported capabilities |

A 60 FPS target implies a frame budget near 16.7 ms; 30 FPS implies 33.3 ms. These are targets, not achieved results. Present cost versus error curves where possible.

Distinguish transport/analysis latency from the model's intentional physical response. A plate taking time to build resonance is not the same thing as delayed presentation of already computed frames.

Uploaded audio with lookahead and causal microphone input should have separate results. Use the audio clock and synthetic impulses to test scheduling. Correlation between audio energy and visual motion can supplement inspection, but correlation alone does not establish causation or perceived synchrony.

## A perceptual study, if pursued

A small exploratory pilot could involve roughly 12–20 participants to debug the procedure and estimate variability. That number is a proposal, not a claim of adequate statistical power. A confirmatory sample should follow a power or precision analysis informed by the expected effect and design.

Use a within-participant design with randomized, counterbalanced condition order. Choose a modest set of licensed excerpts with varied attacks, sustained textures, rhythm, and spectral content. A possible pilot uses 6–12 excerpts of 15–30 seconds, but settling dynamics may require longer examples.

Use identical audio excerpts across conditions. Standardize playback level, display size, and instructions. Give the system the same warm-up or reset policy for every trial; initial grain state can otherwise confound the comparison.

Ask separate questions about perceived synchrony, visual coherence, aesthetic liking, distraction, and understanding. A brief explanation task can check whether participants actually understand nodes or resonance. Do not combine every outcome into one score without justification.

For analysis, account for repeated observations from participants and excerpts. A mixed model with participant and excerpt effects is a candidate; clustered bootstrap intervals are another option when appropriate. Specify primary contrasts, handle multiple comparisons, report effect sizes and uncertainty, and separate exploratory analyses.

Do not treat thousands of video frames as thousands of independent human observations. Keep pilot material separate from confirmatory test material where feasible. Include preference variation rather than implying that one sample defines universal taste.

Follow the relevant institution's consent, data handling, and ethics process for human research. Collect only information needed for the study. Own or appropriately licensed audio makes the reproducibility package easier to share.

## Reproducible release contents

The proposed research repository should contain:

- A model description for each scene, including units, supports, coupling, and limitations.
- Exact solver and renderer configurations, software versions, and random seeds.
- Validation scripts and numerical reference outputs.
- Synthetic tones, impulses, sweeps, and silence fixtures.
- Redistributable music inputs or reproducible feature arrays with input provenance.
- Benchmark scripts and device metadata.
- Analysis notebooks, figure-generation code, and study materials.
- An explanation of failed cases and parameter ranges.
- A pinned release corresponding to each reported experiment.

Offline Python tools such as librosa can serve as a feature reference, with the actual installed version pinned. The original publication is an algorithmic starting point rather than a guarantee about today's library API. [McFee et al.](https://brianmcfee.net/papers/scipy2015_librosa.pdf).

## Publication routes and what each means

| Route | What it establishes | Appropriate project stage |
| --- | --- | --- |
| Public repository and portfolio case study | Inspectable code, documented engineering decisions, and your contribution | After a working, validated prototype |
| Static demo and research report | An accessible demonstration linked to its explanation and evidence | After reproducible results exist |
| Zenodo release archive | Persistent, versioned research artifact with a DOI | At a stable code/report release |
| Undergraduate project or thesis | A supervised academic investigation under the institution's process | When scope and supervision are agreed |
| arXiv preprint | Public scholarly manuscript subject to suitability/moderation | After a coherent original manuscript exists |
| Peer reviewed paper or demo venue | Evaluation under the chosen venue's review process | After a contribution matches its criteria |

Zenodo can archive GitHub releases. A version DOI identifies a particular artifact; a concept DOI covers its versions. Archive the exact version used in experiments. Receiving a DOI is not equivalent to peer review. [GitHub integration](https://help.zenodo.org/docs/github/), [DOI versioning](https://zenodo.org/help/versioning).

arXiv has category suitability, moderation, and sometimes endorsement requirements. It is not an automatic publication guarantee and does not replace peer review. [Endorsement](https://info.arxiv.org/help/endorsement.html), [moderation](https://info.arxiv.org/help/moderation/index.html).

NIME is a plausible community if the main contribution concerns musical expression or an expressive interface. ISMIR becomes a stronger fit when the contribution concerns music analysis or a meaningful MIR evaluation. A beautiful visualization with ordinary FFT features alone may not supply that contribution. Consult their actual tracks when results are ready. [NIME proceedings](https://nime.org/papers/).

The 2026 ISMIR late-breaking demo call inspected for this review is closed. It also distinguishes its abstracts from official proceedings papers. Treat this as information about formats, not a currently available submission route. [Official call](https://ismir2026.ismir.net/call-for-late-breaking-demo).

Our recommendation is to publish a reproducible project report and demo first, archive a stable release, and then choose a scholarly route based on the evidence. Do not advertise a literature review or a planned study as an accepted academic paper.

## Stages and completion criteria

| Stage | Deliverable | Completion criterion |
| --- | --- | --- |
| Research framing | Review, bibliography, models, and study proposal | Claims, access limits, and hypotheses are explicit |
| Analytical prototype | One plate scene and signal inspector | Frequency, phase, decay, and boundary checks pass |
| Transport prototype | Particle model and optional density solver | Conservation and convergence evidence exists |
| Browser demo | Local audio, tones, clear explanations, adaptive rendering | A specified device set runs within measured limits |
| Experiments | Numerical comparison and optional user study | Results answer the registered question with uncertainty |
| Public release | Report, code, demo, archived version | Another person can reproduce the reported outputs |

These are dependency stages rather than promised dates. The present research package completes the framing stage only. A strong portfolio can emerge from the prototype and validation stages even if a peer reviewed paper remains a later goal.