Performance Analysis of LTE Systems with Artificial Intelligence-Assisted Interpretation

The proposed platform is built around a three-layer modular architecture inspired by the Model-View-Controller (MVC) design pattern. The Presentation Layer manages all user interaction through a Streamlit-based web interface running directly in the browser, requiring no local installation. The Logic Layer encapsulates the analytical estimation engine and the AI interpretation pipeline, keeping computational concerns strictly separated from display logic. The Services Layer provides specialized support functions including interactive visualization, step-by-step metric documentation, and independent numerical validation (Liu et al., 2020). This layered structure ensures that any component can be modified, replaced, or extended without creating side effects elsewhere in the system, a critical property for a research platform intended to evolve over time.

The platform comprises eight interconnected Python modules totaling approximately 3,200 lines of code. The main orchestration is handled by streamlit_app.py (~650 LOC), the analytical engine by sim_engine.py (~180 LOC), and the AI pipeline by claude_analysis.py (~550 LOC) paired with validation_engine.py (~450 LOC). All inter-module communication relies exclusively on structured Pandas DataFrames passed through typed function interfaces, eliminating hidden state dependencies and guaranteeing full reproducibility of results (Mohammed et al., 2020). Figure no. 1 illustrates the end-to-end execution workflow, from user parameter input to interactive browser rendering.

Figure no. 1:

Three-layer MVC-inspired architecture of the proposed LTE simulalion platform. Solid arrows indicate primary data flow between layers

(Source: Author’s own conception)

The core analytical engine (sim_engine.py) employs a parametric sigmoid model to compute LTE link-level performance curves without the computational overhead of Monte Carlo methods (3GPP, 2010; Dahlman et al., 2011). It is important to emphasize that this engine does not perform full physical-layer simulation; instead, it approximates BLER as a sigmoid function of SNR, providing a rapid and computationally efficient estimation of link-level behavior suitable for educational analysis and parameter space exploration. The three-stage computation pipeline is illustrated in Figure no. 2.

Figure no. 2:

Three-stage computation pipeline of the LTE simulation engine

(Source: Author’s own conception)

In Stage 1, the engine receives a configuration dictionary containing MIMO mode, modulation scheme, channel model, Doppler frequency, HARQ retransmissions, and antenna correlation (Goldsmith, 2005; Proakis & Salehi, 2008), and maps each parameter to its physical contribution in the SNR_pivot formula via pre-defined lookup tables.

In Stage 2, the pivot point of the sigmoid transition is determined by a composite formula accounting for modulation penalty (SNR_mod), channel degradation (P_channel), MIMO diversity gain (Foschini & Gans, 1998; Lee et al., 2009) (G_MIMO), and HARQ benefit, yielding a single scalar that uniquely characterizes the configuration. For the MIMO 2×2, 16-QAM, EVA reference case this evaluates to 2.8 dB.

In Stage 3, for each of the 16 SNR points sampled uniformly over [−10, 20] dB, the engine computes BLER via the sigmoid model (Proakis & Salehi, 2008) and effective spectral efficiency as shown in equations (1) and (2): 1BLER(SNR)=11+ek(SNR−SNRpivot)BLER(SNR) = {1 \over {1 + {e^{k\left( {SNR - SN{R_{{\rm{pivot }}}}} \right)}}}} 2Tbps/Hz=N(Tx)·bsymbol·(1−BLER){T_{bps/Hz}} = {N_{\left( {{T_x}} \right)}}\,\cdot\,{b_{{\rm{symbol }}}}\,\cdot\,(1 - BLER) where k controls the slope of the transition region, N_Tx is the number of spatial streams (1 for SISO, 2 for MIMO 2×2), and b_symbol is the bits-per-symbol of the modulation scheme (2 for QPSK, 4 for 16-QAM, 6 for 64-QAM). Results for each configuration are stored in a Pandas DataFrame of 16 rows × 5 columns (SNR_dB, BLER, spectral efficiency, Shannon capacity, and Shannon efficiency ratio), with a total execution time under 0.05 seconds per configuration, compared to 5-60 minutes for equivalent Monte Carlo approaches (Dahlman et al., 2011; Ghosh et al., 2010). The parametric nature of the model implies that results represent analytical approximations of expected LTE behavior rather than measured or simulated bit-level outcomes; accordingly, quantitative conclusions should be interpreted within this educational scope.

The AI interpretation subsystem processes analytical outputs through a two-stage pipeline, as illustrated in Figure no. 3.

Figure no. 3:

Two-stage AI interpretation and validation pipeline: Stage 4 generates a structured academic report via Claude Sonnet 4; Stage 5 independently verifies all eight IEEE metrics against NumPy-computed reference values

(Source: Author’s own conception)

In Stage 4, claude_analysis.py implements the first computation path of the platform’s dual-path verification architecture. It pre-computes eight IEEE-standard performance metrics (Proakis & Salehi, 2008; Shannon, 1948) using NumPy and Pandas, SNR at 10% BLER, SNR at 1% BLER, peak throughput, mean throughput, knee SNR, good SNR minimum, peak SNR, and Shannon efficiency, and embeds these values as explicit anchor entries in a structured prompt, alongside the raw CSV simulation data, submitted to Claude Sonnet 4 via the Anthropic API (Anthropic, 2024) (max_tokens = 4,000, temperature = 0.7). The prompt explicitly instructs the language model to use each anchor value exactly as provided and warns that an automated validation system will cross-check numerical precision. Critically, the language model does not independently calculate any metric; it receives pre-computed numerical values and is tasked with constructing a coherent academic narrative that correctly attributes each value to its respective metric definition. An MD5-based caching mechanism eliminates redundant API calls for repeated configurations, reducing operational cost by approximately 70% (Huang et al., 2025).

In Stage 5, validation_engine.py implements the second, fully independent computation path: it recalculates all eight metrics using exclusively NumPy-based computation applied directly to the same raw CSV data, without access to the anchor values embedded in the AI prompt, extracts the values reported by the AI from the generated text via calibrated regular expressions, and cross-validates them against IEEE-standard tolerances: ±0.3 dB for SNR thresholds, ±0.05 bps/Hz for throughput, and ±5% for Shannon efficiency (Proakis & Salehi, 2008; Shannon, 1948). This dual-path design - Path A: analytical engine pre-computes and anchors values in the AI prompt; Path B: validation engine independently recalculates from raw data and audits the AI output – ensures that both the pre-computed anchors and the AI-generated report are verified against the same ground-truth source, eliminating the possibility that a calculation error in Path A could propagate undetected into the final report. Each metric is labelled VALID, INVALID, or MISSING. Evaluation on the MIMO 2×2, 16-QAM, EVA reference configuration confirmed 7 out of 8 metrics as VALID (87.5% metric attribution accuracy). The single failure is attributable to a semantic attribution error: the language model was provided the correct value for Good SNR Minimum (12.00 dB, corresponding to BLER < 1%) as an anchor in the prompt, but incorrectly attributed the SNR @ 10% BLER threshold value (8.04 dB) to the Good SNR Minimum definition in the generated narrative – confusing two conceptually adjacent operating-point definitions despite having the correct value available (Gartlehner et al., 2025; Singh et al., 2024).

Three systematic scenarios were designed to validate platform performance across the fundamental dimensions of LTE link-level behavior, following an isolated-variable methodology: in each scenario, one parameter is varied while all others remain fixed, enabling direct attribution of performance changes to the factor under study (Joseph, 2021). The complete parameter configurations are consolidated in Table no. 1.

Scenario I examines the impact of spatial multiplexing and diversity by comparing four MIMO configurations, SISO, SIMO 1×2, MISO 2×1, and MIMO 2×2 (Foschini & Gans, 1998; Lee et al., 2009), under fixed 16-QAM modulation and EVA channel conditions. The primary metric is SNR gain at 10% BLER relative to the SISO baseline.

Scenario II investigates the effect of mobility and multipath severity by varying the 3GPP channel model across EPA (5 Hz Doppler, pedestrian), EVA (70 Hz, urban vehicle), and ETU (300 Hz, high-speed train) profiles (3GPP, 2016; Goldsmith, 2005), with MIMO 2×2 and 16-QAM held constant. For the ETU configuration, LDPC channel coding is substituted for Turbo coding. While Turbo coding remains the mandatory scheme in 3GPP LTE TS 36.212 (3GPP, 2016), LDPC is introduced here as an analytical benchmark: at 300 Hz Doppler frequencies, the rapid temporal variations of the ETU channel produce burst error patterns that degrade Turbo decoder performance, whereas LDPC’s sparse parity-check structure and belief propagation decoding yield an approximately 1.5 dB SNR gain under equivalent conditions (Bjerke, 2011; Dahlman et al., 2011). This choice reflects coding strategies employed in practical high-speed railway LTE deployments and provides a forward-looking performance bound consistent with the 5G NR transition, where LDPC was adopted as the data channel coding standard in 3GPP TS 38.212 precisely due to its superior performance in high-mobility fading environments. Within the platform, the LDPC gain is modelled as a fixed SNR offset applied to the sigmoid pivot parameter.

Scenario III evaluates the spectral efficiency–robustness trade-off across the three LTE modulation orders, QPSK, 16-QAM, and 64-QAM, under fixed MIMO 2×2 and EVA channel parameters, with Shannon efficiency (Shannon, 1948) as the primary comparison metric.

Figure no. 4 presents the BLER vs. SNR and throughput vs. SNR curves generated by the platform for the MIMO 2×2, 16-QAM, EVA reference configuration. The BLER curve exhibits the characteristic sigmoid shape (Proakis & Salehi, 2008) with three operationally distinct regions: a cutoff zone (SNR < 6 dB) where BLER ≈ 100% and the link is effectively unusable; a steep transition zone (6–12 dB) where error rate drops by three orders of magnitude; and a saturation zone (SNR > 12 dB) where BLER falls below 0.1%. The platform’s analytical model places the 3GPP operational threshold of BLER = 10% at SNR = 8.04 dB, which is consistent with the expected behavior of a MIMO 2×2 system under EVA channel impairments as described in the literature (3GPP, 2016; Lee et al., 2009).

Figure no.4:

Simulated BLER vs. SNR (up) and spectral efficiency vs. SNR (down) for the MIMO 2×2, 16-QAM, EVA reference configuration. Dashed horizontal line at BLER = 10% indicates the 3GPP operational threshold. Dashed blue curve represents the Shannon capacity bound

(Source: LTE Simulator)

The throughput curve computed by the platform rises asymptotically toward a saturation plateau of 7.9999 bps/Hz, which corresponds exactly to the theoretical maximum of N_Tx × b_symbol = 2×4 = 8 bps/Hz for two-stream MIMO with 16-QAM, as established in the spatial multiplexing capacity literature (Foschini & Gans, 1998; Lee et al., 2009). The superimposed Shannon capacity bound (Shannon, 1948), computed as C = 2·log₂(1 + SNR_linear), reaches 13.32 bps/Hz at SNR = 20 dB. The platform’s computed peak throughput of 7.9999 bps/Hz thus represents 60.08% of the Shannon bound, a value characteristic of real LTE deployments accounting for pilot overhead, coding rate constraints, and Turbo decoder imperfections (Dahlman et al., 2011; Ghosh et al., 2010).

To contextualize these results within the broader MIMO comparison, Figures no. 5 and no. 6 present the BLER and throughput curves across all four antenna configurations evaluated in Scenario I. The platform’s output shows a progressive leftward shift of BLER curves, quantified as a 2.0 dB diversity gain for both SIMO 1×2 and MISO 2×1, and a 4.0 dB combined gain for MIMO 2×2, all relative to the SISO baseline. These gain values are consistent with theoretical predictions for MIMO diversity and multiplexing (Foschini & Gans, 1998; Lee et al., 2009). Additionally, the platform computes peak throughput doubling from 4.0 bps/Hz (SISO, SIMO) to 8.0 bps/Hz (MISO, MIMO 2×2), a direct consequence of spatial multiplexing (Rawat et al., 2021), while BLER at a fixed operating point of 10 dB SNR drops from 38.2% (SISO) to 3.0% (MIMO 2×2), representing a 92% reduction in block error rate (Joseph, 2021; Lee et al., 2009).

Figure no. 5:

BLER vs. SNR comparison for SISO, SIMO 1×2, MISO 2×1, and MIMO 2×2 under fixed 16-QAM, EVA, 70 Hz Doppler conditions

(Source: LTE Simulator)

Figure no. 6:

Spectral efficiency vs. SNR for the four antenna configurations. MISO and MIMO 2×2 achieve 8.0 bps/Hz peak through simultaneous two-stream spatial multiplexing

(Source: LTE Simulator)

Following the analytical computation phase, the AI interpretation pipeline described in Section 2.3 was executed on the MIMO 2×2 reference results. The complete generate-interpret-verify cycle completed in 40.2 seconds: 0.05 s for metric pre-computation, 0.05 s for prompt construction, 39.9 s for the Claude Sonnet 4 API response (Anthropic, 2024), and 0.2 s for independent validation. A cached response for an identical configuration is retrieved in under 4 seconds, reflecting the approximately 70% latency reduction achievable through MD5-based caching in repeated educational use (Mardan & Ahmed, 2020; Huang et al., 2025).

The validation engine independently recalculated all eight IEEE metrics using exclusively NumPy-based computation applied directly to the raw simulation data, then extracted and cross-referenced the values reported in the AI-generated text. As detailed in Section 2.3, the 87.5% figure measures metric attribution accuracy of the language model, the rate at which the AI correctly associates each pre-provided anchor value with its respective metric definition in the generated narrative, and does not reflect any independent computational capability of the language model. The dual-path verification architecture ensures that both the anchor values supplied to the AI and the values reported in the AI output are checked against the same raw-data ground truth, so an error in either path is independently detectable. Results, presented in Table no. 2, demonstrate that six metrics were reproduced with effectively zero error: SNR @ 1% BLER (0.00 dB deviation), peak throughput (0.0000 bps/Hz), peak SNR (0.00 dB), mean throughput (0.0000 bps/Hz), knee SNR (0.02%), and Shannon efficiency (<0.01%). The fundamental operational threshold, SNR @ 10% BLER, was reproduced at 8.04 dB versus the independently computed 8.0442 dB, a deviation of 0.05%, well within the ±0.3 dB IEEE tolerance (Proakis & Salehi, 2008).

Figure no. 7:

Independent validation dashboard for the MIMO 2×2, 16-QAM, EVA reference configuration, showing per-metric VALID/INVALID status, absolute deviations, and overall accuracy score

(Source: LTE Simulator)

Figure no. 8:

Side-by-side comparison of independently computed metric values (left column) versus AI-reported values extracted via regular expression pattern matching (right column), for the SNR @ 10% BLER metric

(Source: LTE Simulator)

The single INVALID result – Good SNR Minimum – was reported as 8.04 dB by the AI versus the computed 12.00 dB, representing a 3.96 dB error (33%) that exceeds the ±1.0 dB tolerance. Analysis of the AI output reveals a semantic confusion between two conceptually adjacent metrics: the AI correctly identified the SNR @ 10% BLER threshold (8.04 dB) but mistakenly assigned this value to the Good SNR Minimum definition (BLER < 1%). This systematic confusion between adjacent operating-point definitions constitutes a known failure mode of language model metric extraction (Gartlehner et al., 2025; Singh et al., 2024), and is addressable through structured JSON metric output in future iterations, which would eliminate reliance on regular-expression-based extraction entirely.