A structured path from your current baseline (O-RAN split, Shannon theorem) to confidently discussing RF in 5G AI-RAN simulator architecture.
Even in empty space, signal power spreads spherically. Energy per unit area drops as the sphere grows. Higher frequency = more loss. Doubling distance = +6 dB loss.
FSPL (dB) = 20·log₁₀(d) + 20·log₁₀(f) + 20·log₁₀(4π/c)
Every link budget starts here. Without FSPL you cannot compute received signal strength → no SNR → Shannon gives you nothing. It's the entry point to the whole chain.
λ = c / f. Higher frequency = shorter wavelength. The trade-off: mmWave has massive bandwidth but terrible propagation. Sub-6 GHz penetrates walls, travels far.
λ (meters) = 3×10⁸ / f(Hz)
Your simulator must model different frequency bands with different propagation behaviors. Carrier frequency changes path loss, Doppler magnitude, and antenna array element spacing (λ/2).
When you amplify a sine wave, nonlinearities in the amplifier create energy at 2f, 3f, 4f... These harmonics can fall into adjacent bands and cause interference.
The Power Amplifier (PA) in the RU must operate at a "back-off" point to avoid generating harmonics. This limits transmit power efficiency.
When your team says "PA nonlinearity compensation" or "DPD (Digital Pre-Distortion)", they're talking about mitigating harmonics digitally before the analog stage.
Signal powers span 12+ orders of magnitude. Multiplying gains/losses becomes addition in dB. Link budget = a chain of additions.
dB = 10·log₁₀(P₂/P₁) // power ratio
dBm = 10·log₁₀(P_mW/1mW) // absolute power
dBi = antenna gain vs isotropic radiator
Every RF simulator parameter is in dB/dBm. Link budget: Tx Power + Antenna Gain - Path Loss - Noise = SNR → Shannon capacity.
An antenna doesn't amplify power — it focuses it. Gain = how much more power goes one direction vs a perfect sphere radiator. A 15 dBi antenna is a flashlight vs a candle.
Azimuth = horizontal pointing direction (0–360°). In a 3-sector site: 0°, 120°, 240°. Tilt = how much the beam angles down toward the ground.
In your simulator: antenna gain pattern is an input parameter (3GPP standard model). For massive MIMO the pattern changes dynamically per UE.
Multiple antennas can send different data streams simultaneously in the same frequency (spatial multiplexing), or combine to boost SNR (diversity gain).
64–256 antenna elements in the RU enables: beamforming toward specific UEs, MU-MIMO (serve many UEs at once), and interference nulling.
C = B · log₂(1 + SINR) per stream × N streams
In Cell-Free Massive MIMO (your INFOCOM work), many distributed APs cooperate. Your simulator must model the channel H from each AP to each UE — this is where digital twins become essential.
Splits bandwidth into many narrow subcarriers (15 kHz in 5G NR). Each carries a tiny part of the data. Robust against multipath fading.
12 subcarriers × 1 slot in time. The DU scheduler assigns RBs to UEs every slot (~0.5ms for µ=1).
Throughput = RBs assigned × bits per RB. Bits per RB is determined by MCS, which is determined by SINR/CQI. You need this to map Shannon capacity to real 5G NR scheduling.
The DU/CU split in O-RAN exists partly because RB scheduling (DU, real-time) must be separate from session management (CU, non-real-time).
Licensed spectrum is scarce. Operators must share with legacy LTE deployments.
Spectrum sharing makes your interference model more complex — a UE can see interference from both NR and LTE transmissions. Your SINR must account for all of them.
A transmitted signal reaches the receiver via multiple paths: direct LOS, reflections off buildings, diffraction over rooftops. Each path has different delay, amplitude, phase.
Standard channel models: 3GPP TR 38.901 (UMa/UMi/InH), CDL (Clustered Delay Line), TDL — each represents different environments and compute complexity.
FSPL assumes empty space. Real environments have buildings, terrain, vegetation. Path loss models add empirical correction factors calibrated to measurement data.
PL(dB) = FSPL + 10·n·log₁₀(d/d₀) + χ_σ
n = path loss exponent (2=free space, 3–4=urban), χ_σ = log-normal shadowing
Path loss model is the first thing you implement. It determines received SNR for every UE position. Everything downstream (scheduling, beamforming, handover) depends on it.
Shannon uses SNR. But in any real deployment other base stations and UEs transmit in the same frequency. The operative metric is SINR:
SINR = S / (I + N)
S = desired signal, I = sum of all interference, N = thermal noise. In dense 5G, I often dominates N by 20–30 dB at cell edge.
Remove interference from step 3 and your throughput will be wildly optimistic — the simulator becomes useless for real deployment planning.
In conventional cells you model interference as "N nearest BSs." In Cell-Free, every AP is simultaneously a potential server AND a potential interferer for every UE. The entire point of joint precoding and AP clustering is to convert interferers into cooperating servers. Interference modeling IS the system design problem.
Interference coordination is one of the top AI-RAN use cases: multi-cell RL agents learn power/beam policies to minimize aggregate I across the network. Your SINR model is what those agents are optimizing.
No geometry. No multipath. No spatial info. One distance → one number.
Every bounce. Every delay. Full spatial structure for beamforming.
Ray tracing launches geometric rays from the transmitter. Each ray that reaches the receiver — direct, reflected off walls, or diffracted over edges — travels a different distance and therefore arrives with a different delay and amplitude.
The CIR h(τ) is a set of spikes — one per ray path — at the delay τ equal to that path's total travel distance divided by the speed of light. FSPL gives you only the first spike (LOS, τ=0). Ray tracing gives you all of them.
h(τ) = Σₙ αₙ · e^(jφₙ) · δ(τ − τₙ) // sum over all N ray paths
The delay spread (max τₙ − τ₀) tells you the channel's memory length — and how large an OFDM cyclic prefix must be to prevent inter-symbol interference. FSPL cannot give you this.
Ray tracing replaces the channel model block only. SINR, Shannon capacity, and scheduling logic stay exactly the same — they just receive richer input.
In your Cell-Free + digital twin work: ray tracing generates H between each distributed AP and each UE for a specific campus layout. AP placement, clustering, and precoding can be evaluated against actual geometry — not a random channel realization. The digital twin is the offline ray-trace database feeding the real-time simulator.
Ray tracing is compute-heavy. The practical split: run it offline for all (AP, UE-grid) pairs → store a channel database → replay in real-time simulation. This offline/online architecture is the foundation of the digital twin approach. Ray launching (outward from Tx) is faster and good for coverage maps; ray tracing (backward from Rx) is more accurate for CIR.
A UE moving through the network eventually gets closer to a neighbor cell. Without handover the UE would experience collapsed SINR or drop the connection.
Trigger: neighbor RSRP exceeds serving cell by threshold (A3 event) for Time-to-Trigger (TTT).
Handover in the simulator affects throughput continuity and generates signaling load. In AI-RAN, ML models predict when to trigger proactively. Requires path loss + mobility models working first.
RSRP = P_tx + G_tx - PL - M_fade + G_rx - NF
SINR = RSRP - (kTB + NF) - I_total
kTB = -174 + 10·log₁₀(B) dBm // B in Hz
C = B · log₂(1 + SINR_linear)
Now you see why the order matters: FSPL → path loss → antenna gain → MIMO → interference/SINR → link budget → Shannon. Each concept adds one term to this chain. Your simulator IS this chain, evaluated for every UE position at every time step.
The 7-2x split: beamforming weights computed in DU, applied in RU. Fronthaul carries time-domain IQ samples. Latency requirement: <100µs. This is why fronthaul bandwidth and latency are critical design constraints in O-RAN.
The RF question in every AI-RAN architecture discussion: "What is the channel input X to this ML model, and what RF output Y does it produce?" Knowing path loss, SINR, CSI, and beamforming lets you ask and answer those questions precisely.