#Physics-informed graph models: cardiac paper summary and relevance to speech

Prepared 5 October 2026 for Speech Correlation Research.

Assessment: This paper is useful as an example of combining geometry, learned dynamics and physical constraints. For this project, the strongest connection is to acoustic propagation and geometry-aware spatial correlation. It does not establish a speech-separation result.

#1. Paper summary

Paper: Vu Anh Tran et al., Physics-Informed Surrogate Model Using Graph Neural Network for Cardiac Electrophysiology, MICCAI 2026. Full paper · Official record.

The authors approximate cardiac electrical propagation on anatomical meshes. Nodes contain voltage, recovery state and tissue orientations; edges encode displacement and direction-dependent conductivity. An LSTM processes temporal history, graph blocks exchange spatial information, and a decoder predicts state increments. Training combines data error with discretized monodomain PDE and ionic ODE residuals, using multi-step rollout and decreasing teacher forcing.

Evaluation uses simulations from 130 biventricular meshes, with geometry-disjoint splits and finer test meshes. Reported activation-time MAE is 4.20 ± 3.79 ms. Voltage-map Pearson correlation drops from 0.96 at 20 ms to 0.33 at 100 ms. Reported computation is 0.128 s/map versus approximately 2.15 s/solver step; timing comparability needs checking.

The useful evidence is improvement over an otherwise matching data-only model. Clinical validation is absent; long rollouts drift and predictions are clipped for stability. Table 1's activation-time caption actually accompanies voltage/recovery metrics. Geometry transfer does not establish transfer across arbitrary physical parameters. The abstract's attention description is less specific than the GraphNet implementation account.

Reading basis: full 11-page PDF; architecture and result figures inspected visually. Results are authors' reports, not reproductions.

#2. What can transfer to speech?

The transferable principle is to make a network's representation and training objective reflect the process that generated its observations.

For speech recorded in a room, a useful starting model is:

$x_m(t) = \sum_s (h_{ms} * s_s)(t) + n_m(t)$

Here, each microphone receives source signals filtered by source-to-microphone room impulse responses (RIRs). This separates speech content from acoustic propagation. It suggests several distinct research directions:

These are proposed connections, not results from the cardiac paper.

#The physics must change

Cardiac excitation is a reaction–diffusion process. Ordinary room acoustics is a wave-propagation problem. Copying the cardiac PDE, recovery variable or tissue conductivity into a speech model would be unjustified.

For a homogeneous, source-free acoustic region, a candidate constraint is:

$\frac{\partial^2 p}{\partial t^2} - c^2 \nabla^2 p = 0$

A full room model additionally needs source terms, initial conditions and boundary conditions. A local source-free reconstruction region can instead be constrained by measurements without explicitly reconstructing every wall. Which formulation is appropriate depends on the inference task.

A microphone graph and a volumetric simulation mesh are also different objects. Six microphone nodes do not by themselves provide a trustworthy spatial discretization of a room's Laplacian. A graph model over microphones should use defensible delay/phase or transfer-function relations; a PDE residual needs spatial support or a continuous field representation that makes its derivatives meaningful.

These distinctions are supported by existing acoustic work: Olivieri et al. (2024) reconstruct a continuous pressure field with a wave-equation constraint in a source-free region, using available recordings. Their model is optimized on those observations; it should not be confused with an amortized model that transfers to a new room without adaptation.

#3. Direct connection to the correlation project

SR-CorrNet combines spatio-spectro-temporal correlation features with deep-filter estimation and an asymmetric separation/reconstruction architecture. This provides a natural place to ask whether explicit geometry improves the interpretation of observed correlations. The project’s previous numerical review used v1; the current arXiv record also lists v2. This note does not silently update the earlier benchmark values.

As an illustrative feature, consider a normalized complex cross-product:

$C_{ij}(t,f) = \frac{X_i(t,f)\,\overline{X_j(t,f)}}{|X_i(t,f)|\,|X_j(t,f)| + \varepsilon}$

This is a proposed feature example, not a transcription of SR-CorrNet's exact implementation. It retains pairwise phase information. For a single direct-path source with delay $\tau_i$ at microphone $i$:

$\arg(C_{ij}) \approx -2\pi f(\tau_i - \tau_j) \pmod{2\pi}$

with the sign determined by the Fourier and channel conventions. Geometry constrains plausible delays. Reverberation, noise and competing speakers make observed cross-products depart from this idealization.

Research hypothesis: supplying relative microphone coordinates alongside correlation features may help a network distinguish transferable propagation structure from training-array-specific patterns. A target-conditioned weighting mechanism might further suppress unreliable time–frequency observations.

Important limits:

• A single cross-product is not an averaged covariance or a reliable coherence estimate.
• Time–frequency bins are not spatial locations. Applying a room wave-equation loss to arbitrary spectrogram-neighborhood edges would require a separate derivation.
• All sources obey acoustics. Physical consistency alone cannot identify which speaker is the target; identity, visual evidence or another target cue remains necessary.
• In a single-channel setting, array geometry is unavailable. Temporal and spectral correlations remain relevant, but the spatial proposal does not directly apply.
• Physics losses are soft preferences. Incorrect calibration or assumptions can make them harmful.

#4. Existing speech/acoustic evidence

The general topic already has speech-related prior art, so “apply physics-informed learning to speech” is not by itself a novelty claim.

• Olivieri et al., Physics-informed neural network for volumetric sound field reconstruction of speech signals (2024): the closest direct bridge, with real-environment speech-field reconstruction and a wave-equation regularizer. Reconstruction is distinct from separating overlapping speakers.
• Karakonstantis et al., Room impulse response reconstruction with physics-informed deep learning: combines limited measured RIRs with sound-propagation physics, reporting particular strength on early responses.
• Pezzoli et al., Implicit neural representation with physics-informed neural networks for the reconstruction of the early part of room impulse responses: studies missing early RIR reconstruction in a linear array with real measurements.

The latter two were screened at abstract level for this note, not subjected to a full numerical comparison.

#5. Recommended first experiment — proposal only

Question: Does geometry-aware correlation improve generalization across arrays and rooms at matched model capacity?

Start with one fixed separation backbone and output filter support. Avoid changing the separation/reconstruction architecture at the same time.

Keep context, training data, target definitions and parameter budget comparable. Include a capacity-matched control if adding geometry increases model size.

Data split: hold out rooms, speakers and array layouts independently. Do not split adjacent frames from the same scene across train and test. Test microphone dropout, coordinate error, channel permutation with corresponding coordinate permutation, quiet single-speaker audio, and severe reverberation.

Evaluate: separation improvement with clearly defined references; fixed-ASR WER/CER where appropriate; target-selection failures; localization angular errors only if a localization output is actually present; runtime and memory. Report paired uncertainty across independent scenes, not artificially independent frames.

Acceptance: improvements must survive unseen-room/array evaluation without worsening single-speaker preservation. A smaller physics residual alone is insufficient. If benefits disappear under matched-context and capacity controls, reject the claimed mechanism.

For RIR-surrogate work, run a separate experiment against an appropriate acoustic baseline and measure early-arrival timing, complex transfer-function error and downstream effect. Count fitting/precomputation cost as well as inference latency.

#6. Implications for Mino's work

My recommendation is to explore geometry-aware spatial correlation first. It is a smaller, more diagnostic extension of the existing project than building a full room simulator. RIR surrogates are a longer-term alternative if data generation or room adaptation becomes the actual bottleneck.

Any link to BYD remains conditional on verified microphone geometry and channel identity. This note does not authorize integration into Dolphin + TIGER. The current requirement to discuss architecture with Mino before further localization integration or experiments remains in force.

Decision to make: prioritize transferable correlation features, or make sound-field/RIR reconstruction a separate research question. They need different supervision, baselines and success criteria.