Glossary

ASD: Amplitude Spectral Density. The square root of the PSD, often used to characterise detector noise in units of strain per root Hz.
BBH: Binary Black Hole. A compact binary system consisting of two black holes. BBH mergers are the dominant source class in current LIGO/Virgo catalogues and the primary target of Sage’s production runs.
BCE: Binary Cross-Entropy. The classification loss used by Sage to distinguish signal from noise samples. Combined with heteroscedastic regression in BCEWithPEsigmaLoss.
BNS: Binary Neutron Star. A compact binary system consisting of two neutron stars. BNS mergers produce both gravitational waves and electromagnetic counterparts (kilonovae), as demonstrated by GW170817.
CBAM: Convolutional Block Attention Module. An attention mechanism applied after each residual block in Sage’s backend ResNet. Applies sequential channel attention and spatial attention to focus the network on merger-relevant features.
CBC: Compact Binary Coalescence. The inspiral, merger, and ringdown of a compact binary system (BBH, BNS, or NSBH). CBC signals are the primary target of current matched-filter and ML GW search pipelines.
Chirp mass: The combination of component masses \(m_1\) and \(m_2\) that governs the leading-order frequency evolution of a CBC signal: \(\mathcal{M} = (m_1 m_2)^{3/5} / (m_1 + m_2)^{1/5}\). It is the best-measured mass parameter in GW observations.
FAR: False Alarm Rate. The rate at which noise triggers exceed a given ranking statistic threshold, typically expressed in units of events per month or per year. Detection claims are usually made at FAR < 1 per year.
FD: Frequency Domain. Sage generates waveforms and performs whitening in the frequency domain for efficiency, converting to the time domain only for multirate compression.
Fiducial PSD: A fixed representative PSD used for whitening during training. Sage pre-computes a fiducial PSD from the noise dataset and applies it consistently across all batches to avoid per-sample PSD estimation overhead.
GWOSC: Gravitational-Wave Open Science Center. The public repository of LIGO/Virgo/KAGRA strain data, segment information, and event catalogues. Sage’s data primer downloads directly from GWOSC.
Heteroscedastic regression: A regression approach in which the model predicts both the mean and the variance of each output. Sage uses this to produce calibrated uncertainty estimates on parameters such as chirp mass and coalescence time.
IMRPhenomD: An aligned-spin, frequency-domain phenomenological waveform model for CBC signals. Sage provides a fully vectorised GPU implementation compatible with torch.compile.
IMRPhenomPv2: An extension of IMRPhenomD to precessing-spin binaries via the PhenomPv2 “twist-up” formalism. The default waveform model used in Sage production runs.
IST: Inverse Spectrum Truncation. A time-domain windowing technique applied to the inverse PSD before whitening, preventing spectral leakage from sharp PSD features such as spectral lines.
Mismatch: A measure of waveform similarity: \(1 - \langle h_1, h_2 \rangle\), where the angle bracket denotes the noise-weighted inner product maximised over time and phase. A mismatch of zero means identical waveforms.
MLGWSC: Machine Learning Gravitational-Wave Search Challenge. A standardised injection study (Schäfer et al. 2023) providing common noise and signal datasets for direct comparison between GW search pipelines.
Multirate sampling: A compression scheme that assigns different sample rates to different time intervals of a CBC signal, keeping high time resolution only near the merger where signal power is concentrated. Reduces the effective sequence length from ~24 576 to ~2 600 samples.
NSBH: Neutron Star – Black Hole. A compact binary system consisting of one neutron star and one black hole.
OTF: On-the-Fly. Sage generates every training batch from scratch at runtime — no pre-computed dataset is stored. Waveforms and noise windows are sampled and processed per batch to eliminate data-reuse biases.
PSD: Power Spectral Density. A frequency-domain characterisation of detector noise power per unit bandwidth, in units of strain² per Hz. Used for whitening and optimal SNR calculation.
Recolouring: A data augmentation technique in which a real noise segment is whitened with its original PSD and then re-coloured with a different PSD drawn from a library. This exposes the network to a wider range of noise spectral shapes during training.
ROC curve: Receiver Operating Characteristic curve. A plot of true alarm probability versus false alarm probability as the detection threshold is varied. Used to compare pipeline sensitivity across different methods.
Sensitive distance: The volume-averaged maximum distance at which a pipeline can detect signals from a given population at a fixed FAR. A standard MLGWSC benchmark metric.
SNR: Signal-to-Noise Ratio. The ratio of signal power to noise power in the detector data, optimally computed as the noise-weighted inner product of the signal template with itself. Sage uses optimal SNR rescaling during training.
TD: Time Domain. Real-valued strain as a function of time. Sage applies multirate sampling in the time domain after converting whitened frequency-domain data via inverse FFT.
Whitening: Division of the frequency-domain strain by the ASD, transforming coloured detector noise into approximately white (flat-spectrum) noise. Sage implements whitening via FiducialWhitening.