$Φ$asmaNet: A Staged Spectral Token-Refinement Architecture for Fast Gravitational-Wave Signal-to-Noise Ratio Prediction

Ioannis Tsioulis, Anastasios Tefas, Nikolaos StergioulasarXiv 2026 · arXiv

TL;DR

$\Phi$asmaNet is a novel neural network architecture designed to predict gravitational-wave signal-to-noise ratios (SNR) two orders of magnitude faster than traditional methods. By employing a staged refinement process with attention-based modules and top-down feedback, it achieves over 99% accuracy, enabling the massive population studies essential for modern astrophysics that were previously computationally prohibitive.

What Problem Does This Paper Address?

In gravitational-wave (GW) astronomy, researchers don't just care about individual detections; they want to understand the entire population of merging black holes and neutron stars. However, our detectors have "selection effects"—we are much more likely to see loud, nearby, or favorably oriented events. To correct for this bias, scientists must simulate millions of synthetic binary systems and determine if they would have been detected.

The gold standard for "detectability" is the Signal-to-Noise Ratio (SNR). Traditionally, calculating SNR requires generating a complex waveform and performing a noise-weighted inner product. While this takes only a fraction of a second per source, doing it $10^7$ times for a single population study creates a massive computational bottleneck. $\Phi$asmaNet acts as a high-speed "surrogate model," replacing the expensive physics integration with a lightning-fast neural network evaluation.

Methods

The authors introduce $\Phi$asmaNet, a compact architecture specifically tuned for structured tabular data (like the 11 parameters defining a binary black hole merger). Unlike standard Transformers or MLPs, $\Phi$asmaNet enforces a "staged" refinement process:

  1. Tokenization: The 11 input parameters (masses, spins, etc.) are projected into a set of learned "tokens."
  2. $\Phi$-Stages: An ordered sequence of blocks that refine these tokens. Each stage uses multi-head attention to capture interactions between parameters.
  3. $\Phi$-Refractor: A novel subspace-mixing module within each stage. It partitions the token embeddings into parallel subspaces, transforms them independently to promote feature specialization, and fuses them back together using learned gating.
  4. $\Phi$-Feedback: A top-down recursive mechanism. Once the model reaches the "highest" level of abstraction, it passes that global context back down to lower levels to refine the initial representations before a final re-computation.
  5. PCHIP Calibration: A post-processing step using Piecewise Cubic Hermite Interpolation to ensure the final predictions are monotonically calibrated, further reducing systematic bias.
Role of the SNR surrogate in population studies

Datasets

The primary dataset consists of $10^5$ to $10^6$ simulated Binary Black Hole (BBH) mergers.

  • Waveform Model: IMRPhenomXPHM, which includes complex effects like spin-precession and higher-order modes.
  • Parameters: 11 inputs, including chirp mass, mass ratio, six spin components, and antenna pattern projections.
  • Noise Curve: Advanced LIGO design sensitivity (aLIGOZeroDetHighPower).

To test the architecture's versatility, the authors also evaluated it on standard tabular benchmarks: YearPredictionMSD (regression), Covertype (multiclass classification), and Epsilon (high-dimensional binary classification).

Key Insights and Ideas

  • Explicit Inductive Bias: While most deep learning models rely on depth to build hierarchy, $\Phi$asmaNet builds it into the architecture. The "staged" approach mimics the way a physicist might think about the problem—moving from local parameter interactions to a global signal descriptor.
  • The "Spectral" Metaphor: The $\Phi$-Refractor allows the network to look at the data from multiple "perspectives" simultaneously, which is particularly useful for GW data where different physical domains (like mass vs. spin) influence the SNR through very different coupling mechanisms.
  • Distance Invariance: The authors leverage a key physical property: SNR scales exactly as $1/d_L$ (luminosity distance). By training the model at a fixed reference distance, they create a surrogate that is analytically rescalable to any distance at inference time.
Comparison of different GW events against the detector noise curve

Key Results

  • Accuracy: On a test set of 1 million samples, $\Phi$asmaNet achieved a post-calibration accuracy of 99.29% (defined as 100 - Mean Absolute Percentage Error). In the critical detection regime ($\rho > 8$), accuracy rose to 99.39%.
  • Speedup: The model sustains $\sim 1.2 \times 10^5$ evaluations per second on a single GPU. It can process a million-injection catalog in 8.3 seconds—roughly 100x faster than a 16-core CPU using traditional methods.
  • Efficiency: With only $\sim 732\text{k}$ parameters, it outperformed much larger models, including a 1D-ResNet with nearly 4 million parameters.
  • Generalization: Despite being designed for GW data, the model ranked 2nd out of 9 established deep learning architectures on the Covertype tabular benchmark without any dataset-specific hyperparameter tuning.

Limitations and Future Work

  • Fixed Sensitivity: The current model is tied to a specific detector (LIGO-Hanford) and a fixed noise power spectral density (PSD). Future versions will need to handle varying PSDs across different observing runs.
  • Single Detector: While the authors demonstrate how to extend this to networks via quadrature sums, the current study focuses on a single-detector configuration.
  • Hard Thresholds: The study uses a deterministic SNR threshold for detectability. The authors plan to integrate $\Phi$asmaNet into noise-averaged selection function pipelines to better model real-world search pipeline effects like non-stationarity.
  • Ablation Depth: While the results are strong, more exhaustive ablation studies are needed to precisely quantify the marginal gains of the $\Phi$-Feedback vs. the $\Phi$-Refractor components.

Summary updated

Abstract

Quantifying selection effects in gravitational-wave population studies requires millions of costly signal-to-noise ratio (SNR) calculations. We present $Φ$asmaNet, a compact neural architecture for structured tabular regression, with staged spectral refinement as an explicit architectural prior. Learned input tokens are refined through ordered attention-based $Φ$-Stages, each augmented by a $Φ$-Refractor subspace-mixing module, then revisited through gated top-down $Φ$-Feedback. Applied to the 11-dimensional parameters of precessing, higher-mode binary black hole signals (IMRPhenomXPHM), it predicts phase-averaged optimal SNR for a single detector at fixed Advanced LIGO design sensitivity. With approximately 732k parameters, it outperforms MLP, 1D-ResNet and Transformer-encoder baselines under matched training across $10^5$--$10^6$ samples. At $10^6$ samples, it achieves $100-\mathrm{MAPE}=98.73\%$ on the held-out test set, $99.29\%$ after rank-preserving monotonic (PCHIP) calibration, and $99.39\%$ ($R^2=0.9998$) in the detection-relevant regime $ρ>8$, with stable accuracy across independent dataset realizations. On a single GPU, it sustains approximately $1.2\times10^5$ SNR evaluations per second and processes a $10^6$-injection catalogue end-to-end in approximately 8.3 s, about two orders of magnitude faster than conventional waveform-and-inner-product calculations on a 16-core CPU. Exact distance rescaling through $ρ\propto1/d_L$ supports repeated catalogue-wide detectability evaluations. The fully documented pipeline is intended as a reproducible benchmark, while results on public tabular datasets indicate competitiveness beyond gravitational-wave applications. Although demonstrated for a fixed single-detector configuration, the construction supports extensions to detector networks and epoch-specific sensitivities. (Abridged)
BibTeX
@misc{tsioulis2026asmanet,
  author = {Ioannis Tsioulis and Anastasios Tefas and Nikolaos Stergioulas},
  title = {{$Φ$asmaNet: A Staged Spectral Token-Refinement Architecture for Fast Gravitational-Wave Signal-to-Noise Ratio Prediction}},
  year = {2026},
  eprint = {2611.05309},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05309},
}
Figure