Selection effects need millions of SNR values: asmaNet predicts about 120,000 per second on one GPU with 732k parameters

A compact neural network can return the signal-to-noise ratio directly from binary parameters, removing the waveform generation and noise-weighted inner product from each evaluation. Earlier fast SNR surrogates were almost all plain multilayer perceptrons over low-dimensional intrinsic parameters; the production gwsnr package's neural estimator uses only five engineered scalar inputs and compresses the spin degrees of freedom into one effective spin. At 100k, 500k and 1M samples under one shared training protocol, asmaNet outperforms the MLP, 1D-ResNet and Transformer-encoder baselines at every scale; at 1M samples its post-calibration accuracy is 99.29%, with the ResNet, Transformer and MLP reaching correspondingly lower values.

The network first turns the 11-dimensional input into a small set of learned tokens, then refines those tokens through an ordered sequence of stages. Each stage pairs a standard Transformer encoder layer with a Φ-Refractor module that splits the embedding dimension into subspaces, passes each through an independent two-layer MLP, and recombines them by softmax-gated residual weighting. On the 1M-sample held-out test set, post-calibration accuracy in the detection-relevant regime (optimal SNR at least 8) is 99.36%, with a mean absolute percentage error of 0.64% in linear SNR space.

The refined tokens then pass through several top-down gated feedback iterations that inject higher-level context into lower levels before a forward recomputation. These feedback iterations reuse the same stage and gate weights, so adding iterations adds no parameters and only raises the cost of each forward pass. On a single RTX 4080, a full end-to-end evaluation of an independent 100k-injection dataset, including data loading, forward passes, metrics and calibration, takes about 8.3 s, roughly 120,000 SNR evaluations per second, versus about 980 s for the conventional calculation on a 16-core CPU.

The demonstrated setting is a single LIGO-Hanford detector at fixed Advanced LIGO design sensitivity, with IMRPhenomXPHM precessing, higher-mode binary black holes and an 11-dimensional parameter and antenna-pattern projection as input rather than detector strain. The optimal SNR scales exactly as inverse distance, so a surrogate trained at one reference distance can be rescaled to any luminosity distance at query time; the authors propose next training one surrogate per detector and combining them in quadrature, and retraining for or conditioning on the noise spectra of different observing epochs.

The authors state that attributing gains to individual components would require more exhaustive ablations; the Φ-Refractor on/off difference on the SNR task mostly falls within run-to-run variation, with only a modest and more robust off advantage at 100k samples, and the default-on choice rests mainly on its larger benefit on public tabular benchmarks. Training labels inherit the IMRPhenomXPHM approximant's systematic uncertainty at high mass ratio and large in-plane spin, and the uniform polar-angle sampling used in training is not the isotropic measure. Transfer to other noise spectra, observing epochs or detector networks is not established, and the hard-threshold detectability is only an accuracy diagnostic, with noise-averaged detection probability and sensitive-volume bias left to follow-up work.