Glossary

Alphabetical list of acronyms and terms used across the rfgen documentation. Click any cross-link to reach the authoritative concept page.

---

ABC: Abstract Base Class. A Python class that uses abc.ABC (or abc.ABCMeta) to declare abstract methods, preventing direct instantiation and enforcing a contract on every concrete subclass. The rfgen framework defines one primary ABC per layer (e.g., BaseEmitter, BaseChannel).

AGC: Automatic Gain Control. A feedback circuit that adjusts receiver amplifier gain to maintain a roughly constant output power level across a wide input-power range. In rfgen, AGC is the RX-hardware transformation modeled by BaseAGC.

AoA: Angle of Arrival. The incident direction of a radio wave at a receive antenna, typically measured in azimuth and elevation. Required for TDOA, beamforming, and direction-finding tasks. See Channels / Propagation for which backends preserve joint spatial correlation across RX antennas.

AWGN: Additive White Gaussian Noise. A standard noise model where each sample is drawn independently from a Gaussian distribution with flat power spectral density. In rfgen, additive thermal noise is no longer a standalone stage; it enters as LNA_NOISE inside Group.RX_CAPTURE via BaseLNANoise, where the noise power is set by the LNA noise figure and the receiver bandwidth. See Concepts / Channels / RX capture.

BS: Base Station. A fixed radio access node (e.g., a cellular tower or Wi-Fi access point). In 3GPP terminology, BS refers to the network side; the mobile side is the UE.

CDL: Clustered Delay Line. A 3GPP TR 38.901 small-scale fading model that adds angular cluster structure (angles of arrival and departure) on top of a tapped delay line. Useful when antenna arrays, beamforming, or MIMO decoding are part of the downstream task. Sionna exposes this as sionna.phy.channel.tr38901.CDL. See Concepts / Channels / Propagation.

TDL: Tapped Delay Line. A 3GPP TR 38.901 small-scale fading model for SISO simulation. Each tap fades independently; no spatial information. Simpler than CDL; preferred for controlled link-level experiments. Sionna exposes this as sionna.phy.channel.tr38901.TDL.

CFO: Carrier Frequency Offset. A small difference between the nominal TX carrier frequency and the frequency the receiver synthesizer locks to, caused by reference-oscillator drift. CFO appears as a slow phase ramp on the received IQ. In rfgen, CFO is applied per emitter as the CFO transformation inside Group.TX via BaseCFO.

EIRP: Effective Isotropic Radiated Power. Total power radiated by a transmitter relative to an isotropic antenna, equal to conducted power times antenna gain. Used to characterize the TX power budget in link budget calculations. Configured through emitter and scene power parameters.

ENOB: Effective Number of Bits. A measure of the actual resolution of an ADC or DAC, accounting for distortion and noise. Lower ENOB means more quantization noise. Modeled in rfgen through BaseDACQuantization on the TX side and BaseADCQuantization on the RX side.

EW: Electronic Warfare. The use of the electromagnetic spectrum for military purposes: jamming, spoofing, detection, and interception. The contested_ew_lband scenario preset targets EW scene understanding research.

FEC: Forward Error Correction. Channel coding that adds redundancy to transmitted bits so receivers can correct errors without retransmission. Not modeled inside rfgen emitters (emitters generate baseband waveforms at the PHY layer; upper-layer coding is outside scope).

FHSS: Frequency-Hopping Spread Spectrum. A transmission technique in which the carrier hops among many frequencies according to a pseudorandom sequence. Drone control links (e.g., OcuSync, Lightbridge) and some radars use FHSS. Emitters that sweep frequency intrinsically receive f_offset_hz directly from the scene composer rather than being mixed post-hoc. See Coordinate Systems.

HIL: Hardware-in-the-Loop. A validation technique that connects real radio hardware (e.g., USRP, HackRF) to the simulation so synthetic waveforms can be compared against actual SDR captures. rfgen’s HIL harness is planned for v1.5. See Background / Validation.

IF: Intermediate Frequency. A frequency to which an RF signal is shifted before final demodulation. rfgen generates and processes baseband IQ; IF stages are considered part of the analog hardware not modeled here.

IQ: In-phase / Quadrature. The standard representation of a complex baseband signal as two real-valued channels: the real part (I) and imaginary part (Q). In rfgen the canonical in-memory shape is (2, N) float32, channel-first. See Reference / IQ Layout Policy.

ITU: International Telecommunication Union. The United Nations agency responsible for global radio frequency allocation and telecommunications standards. ITU propagation models (ITU-R P series) inform some of the reference path-loss curves used to validate Sionna backends.

JCAS: Joint Communications and Sensing. A dual-function radio system that reuses the same waveform and hardware for both data communication and environmental sensing (e.g., radar). A planned roadmap area for rfgen’s emitter catalog.

kTBF: Thermal noise floor, computed as k × T × B × F, where k is Boltzmann’s constant (1.38 × 10⁻²³ J/K), T is temperature in Kelvin (~290 K standard), B is bandwidth in Hz, and F is the noise figure. This gives the minimum achievable noise floor at the receiver input. See Reference / Noise Floor Table.

LNA: Low-Noise Amplifier. The first active component in most receiver chains, placed immediately after the antenna to amplify the received signal before any lossy components introduce noise. LNA noise figure (NF) is the primary parameter of BaseLNANoise inside Group.RX_CAPTURE.

LO: Local Oscillator. An oscillator within a radio that provides the reference tone for frequency translation. The scene center frequency (scene_center_frequency_hz) in rfgen corresponds to the virtual LO that defines 0 Hz in the wideband scene IQ. See Coordinate Systems.

LOS: Line of Sight. A propagation condition where transmitter and receiver have an unobstructed geometric path between them. LOS paths typically have stronger signal levels (higher K-factor) and less delay spread than NLOS.

NLOS: Non-Line of Sight. A propagation condition where the direct geometric path is blocked by obstacles. NLOS channels typically show stronger multipath and lower received power. The 3GPP UMa, UMi, and RMa scenario models include stochastic LOS/NLOS state transitions.

mAP: mean Average Precision. The primary detection metric used in object-detection benchmarks (COCO, PASCAL VOC). Averaged over IoU thresholds and object classes. Used to evaluate wideband detection models trained on rfgen datasets.

MIMO: Multiple Input Multiple Output. Radio systems with multiple transmit and receive antennas that exploit spatial multiplexing or diversity. Multi-RX rfgen scenes with RecordAxis.JOINT expose the joint tensor needed for MIMO decoding.

mIoU: mean Intersection over Union. The primary segmentation metric: average over classes of (intersection area) / (union area) between predicted and ground-truth masks. Used to evaluate time-frequency segmentation models trained on rfgen datasets.

NF: Noise Figure. A measure of the noise added by a component, in dB above the thermal noise floor. Lower NF means less added noise. The LNA NF is the dominant NF source in the RX chain.

NR PUSCH: New Radio Physical Uplink Shared Channel (5G). The 5G NR uplink data channel. In rfgen, NRPUSCHEmitter wraps sionna.phy.nr.PUSCHTransmitter to generate protocol-conformant 5G NR PUSCH waveforms.

NTIA: National Telecommunications and Information Administration. The US agency that manages federal spectrum use (the FCC manages non-federal). NTIA frequency allocation charts are useful for planning real-world RF scenarios.

OSM: OpenStreetMap. A community-maintained geographic database. Sionna RT can load OSM-derived 3D scene meshes for site-specific ray-traced propagation. rfgen uses OSM scenes (e.g., Munich downtown) for the dense_urban_2_4ghz preset.

PA: Power Amplifier. The final amplifier stage in a transmitter. PA nonlinearity (modeled by the Rapp or Saleh model) causes spectral regrowth and EVM degradation, and is a key source of per-device fingerprint. In rfgen, PA nonlinearity is the PA_NONLINEARITY transformation inside Group.TX via BasePANonlinearity.

PAES: Physical Attribute Extraction Score. A metric introduced in this framework for measuring annotation quality. The verifier inference pass re-extracts physical attributes (modulation, SNR, bandwidth, channel, time placement) from a text annotation and compares to ground-truth metadata. Score = fraction of attributes recovered within tolerance. Target: PAES ≥ 0.80 on all five annotation types.

PRI: Pulse Repetition Interval. The time between the start of successive radar pulses. PRI determines the unambiguous range of a pulsed radar. rfgen’s radar emitters accept pri_us as a parameter; the TimePlacementStrategy.PRI_PULSE_TRAIN strategy schedules pulses at the specified interval.

PSD: Power Spectral Density. The distribution of signal power per unit frequency (in dBm/Hz or W/Hz). Used in SNR calculations, spectral mask compliance checking, and AWGN noise floor specification.

RFML: Radio Frequency Machine Learning. The research discipline of applying machine learning to RF signal processing: modulation classification, signal detection, specific emitter identification, and related tasks. rfgen is designed to generate training data for RFML tasks.

RIS: Reconfigurable Intelligent Surface. A passive array of tunable scattering elements that can be programmed to reflect or focus radio waves. An emerging propagation technology; Sionna RT supports RIS scenes. Listed as a planned roadmap item for rfgen.

RMa: Rural Macro (3GPP). A 3GPP TR 38.901 scenario for base-station deployments in rural areas with large inter-site distances and lower building density. Implemented in Sionna as sionna.phy.channel.tr38901.RMa.

UMa: Urban Macro (3GPP). A 3GPP TR 38.901 scenario for macro-cell deployments in urban areas with base-station height around 25 m. Implemented in Sionna as sionna.phy.channel.tr38901.UMa.

UMi: Urban Micro (3GPP). A 3GPP TR 38.901 scenario for small-cell or street-canyon deployments. Implemented in Sionna as sionna.phy.channel.tr38901.UMi.

RT: Ray Tracing. The propagation simulation method that traces discrete geometric paths between TX and RX, accounting for reflections, diffraction, and scattering from scene geometry. In rfgen, Sionna RT (SionnaRT) is the ray-tracing backend. See Concepts / Channels / Propagation.

SDR: Software-Defined Radio. A radio communication system that implements typical hardware components (amplifiers, modulators, mixers) in software. Hardware SDRs (USRP, HackRF, ADALM-PLUTO) are used in rfgen’s HIL validation harness.

SEI: Specific Emitter Identification. The RFML task of identifying a specific physical radio device from its waveform, relying on subtle hardware impairments (CFO, IQ imbalance, phase noise, PA nonlinearity) that are device-specific. The device_fingerprint_lora_25 preset targets SEI research.

SFO: Sample Frequency Offset. A small mismatch between the nominal and actual sampling clock frequency, causing the received IQ to appear to run at a slightly different rate than expected. SFO does not have a dedicated transformation slot in the current pipeline; the SEI literature treats it as a TX-side oscillator impairment that manifests as a small fractional resample, and rfgen realizes that as a (1 + ε) resample inside BaseResampler in Group.RX_CAPTURE. See Reference / Algorithms / Fingerprint Math for the math and Background / RF Fingerprinting (SEI) for context.

SigMF: Signal Metadata Format. An open standard for recording RF captures as paired binary data (.sigmf-data) and JSON metadata (.sigmf-meta). rfgen accepts SigMF captures as read-only input for HIL validation and replay emitters.

SNR: Signal-to-Noise Ratio. The ratio of signal power to noise power, typically expressed in dB. In rfgen, SNR is specified as a per-emitter target; the realized noise floor is set by LNA_NOISE inside Group.RX_CAPTURE via BaseLNANoise, and the per-emitter realized SNR is recorded in SignalMetadata.snr_db and SceneMetadata.realized_snr_distribution.

SSL (machine learning): Self-Supervised Learning. A family of ML training approaches that derive supervisory signal from the data itself without human labels, often by predicting masked or augmented versions. The ContrastiveCollator in rfgen supports SSL training on IQ data.

SSL (networking): Secure Sockets Layer. A cryptographic protocol for securing network communications, superseded by TLS. The two SSL acronyms are unrelated; context disambiguates them.

Stage: Retired. In the previous channel pipeline, Stage was a five-member IntEnum (TX_HARDWARE, FREQUENCY_SHIFT, PROPAGATION, RX_FRONTEND, AWGN) tagged onto every BaseChannel implementation. The current design replaces it with two finer-grained enums: Group (four members partitioning the chain at the pre-sum / post-sum boundary) and Transformation (fourteen members naming individual operations within a group). See Background / Design Decisions for the rationale and Concepts / Channels for the current model.

TDOA: Time Difference of Arrival. A localization technique that estimates the position of a transmitter by comparing the arrival time of its signal at multiple receivers. Requires joint multi-RX records (RecordAxis.JOINT).

TX / RX: Transmit / Receive. TX denotes the transmitter side (emitter, BS downlink, or UE uplink); RX denotes the receiver side. In 3GPP cellular contexts, TX and RX roles depend on direction: uplink means UE transmits, BS receives; downlink means BS transmits, UE receives.

UE: User Equipment. The mobile terminal in a cellular network (phone, modem, IoT device). In 3GPP uplink scenarios the UE is the transmitter; in downlink it is the receiver.

V2X: Vehicle-to-Everything. Encompasses V2V (vehicle-to-vehicle), V2I (vehicle-to-infrastructure), and related vehicular communications. V2X scenarios with distributed TX positions are a motivating use case for ChannelApplicationMode.PER_EMITTER in rfgen.

Zarr: A chunked array storage format optimized for cloud object stores (GCS, S3). Each Zarr chunk is a separate object, enabling concurrent writes by many workers without file-level locking. rfgen uses Zarr as the canonical Phase 1 storage format. See Concepts / Storage and Reference / Storage Layout.