Predicting Throughput Degradation in Wi-Fi 6 Networks Under Varying Channel Conditions: A Physical Layer Performance Model

cases before integration into the full prediction pipeline. 6.1. Link Adaptation Algorithm Behavior Under Channel Impairments The implicit feedback path, based on MPDU acknowledgment rates in the downlink, provides coarser channel quality information but is available after every frame exchange and can track faster channel variations. Commercial link adaptation implementations combine these two feedback paths through a weighted scoring mechanism that gives greater weight to recent implicit feedback in high-Doppler environments and greater weight to explicit CSI in low-mobility environments. The specific weighting parameters used by different manufacturers are proprietary, contributing to the variation in link adaptation performance observed across hardware platforms that nominally implement the same 802.11ax specification. (Shellhammer et al., 2001) (Espressif, 2021; Eyetsemitan et al., 2020; Eyetsemitan et al., 2021; Eyetsemitan et al., 2023; Eyetsemitan et al., 2025; Eyetsemitan et al., 2022; Peterson et al., 1998; Pozar, 1983) (Pozar et al., 1987; Pozar et al., 1995; Pozar, 2012; Proakis et al., 2006) Retransmission overhead is a significant throughput-limiting factor in 802.11ax networks operating near MCS transition boundaries. When the link adaptation algorithm selects a MCS that results in a packet error rate above the target (typically 10 percent), the failed frames must be retransmitted, consuming channel resources that could otherwise be used for new data transmission. The throughput reduction due to retransmission overhead is approximately T_actual = T_nominal * (1 - PER) * (1 - overhead_retx), where T_nominal is the throughput at the selected MCS with zero PER, PER is the observed packet error rate, and overhead_retx accounts for the additional channel time consumed by retransmitted frames including their MAC headers and inter- frame spacings. At a PER of 30 percent, the retransmission overhead reduces the effective throughput by approximately 35 percent relative to the nominal MCS rate, a significant degradation that occurs precisely in the channel quality range where link adaptation algorithms are most prone to MCS selection errors. (Perahia et al., 2013) 6.2. Packet Error Rate Measurement and Modeling The OFDMA downlink scheduler in 802.11ax must solve a resource allocation problem that assigns resource units to stations in a manner that maximizes the aggregate network throughput subject to constraints on fairness, latency, and buffer state. The optimal solution to this problem, in the Shannon capacity sense, assigns each subcarrier to the station with the highest instantaneous channel gain on that subcarrier, a policy known as frequency-selective opportunistic scheduling. In practice, commercial 802.11ax schedulers approximate this optimum using simplified heuristics that limit CSI feedback overhead and computational complexity, resulting in suboptimal resource unit assignments that reduce aggregate throughput below the theoretical maximum (Yang et al., 2021). IEEE 802.11ax introduces a new OFDM symbol duration of 12.8 microseconds, four times longer than the 3.2 microsecond symbol duration of 802.11ac, corresponding to a subcarrier spacing of 78.125 kHz compared to 312.5 kHz in 802.11ac. The longer symbol duration provides improved robustness against multipath delay spreads exceeding 3.2 microseconds, which would have caused inter-symbol interference in 802.11ac but are safely within the guard interval of 802.11ax's extended cyclic prefix options of 0.8, 1.6, and 3.2 microseconds. For indoor environments with typical delay spreads below 200 nanoseconds, the 0.8 microsecond guard interval is sufficient, providing the lowest overhead fraction and the highest throughput efficiency of the three options. (Chiasserini et al., 2003) (Eyetsemitan et al., 2023; Fenn, 2008; Frenzel, 2012; Frickey, 1994; Fujimoto et al., 2001; Gao et al., 1997; Qualcomm, 2021; Qureshi et al., 2022) (Raab et al., 2002; Rappaport, 2002; Raza et al., 2017; Raza et al., 2021) The beamforming feedback required to enable MU-MIMO in 802.11ax imposes a significant overhead on the uplink capacity, particularly for stations with large numbers of antennas and high spatial stream counts. For a four-station MU-MIMO group with 2x2 MIMO stations and an 80 MHz channel, the compressed beamforming feedback from each station occupies approximately 2 to 4 kilobytes per feedback interval. At a feedback interval of 100 milliseconds, the aggregate beamforming feedback overhead is 80 to 160 kilobits per second per station, which represents a non-trivial fraction of the uplink capacity in the 2.4 GHz band. Network designs that balance the throughput gain from MU-MIMO against the overhead cost of beamforming feedback must consider the station density, the uplink capacity budget, and the channel coherence time, which determines the minimum feedback interval necessary to maintain accurate spatial multiplexing channel estimates. (Rohde et al., 2019) 6.3. Target Wake Time and IoT Throughput Efficiency The throughput model developed in this paper does not incorporate TWT behavior because the testbed measurements were conducted with stations in continuous active mode. For IoT applications that use TWT, a modified model that incorporates the TWT service period duration and interval as additional features would be needed to accurately predict the throughput of TWT- enabled stations. The interaction between TWT scheduling and OFDMA resource unit assignment is a particularly complex aspect of 802.11ax that is not fully addressed in the current standard and remains an active area of research and standardization. (Park et al., 2019) 6.4. Channel Estimation and EVM in 802.11ax The channel estimation procedure in 802.11ax uses pilot subcarriers distributed among the data subcarriers of each OFDM symbol to track the channel frequency response across the channel bandwidth. The pilot density and placement in 802.11ax are designed to provide accurate channel estimates up to a channel Doppler spread of approximately 100 Hz, corresponding to pedestrian mobility at 2.4 GHz. For higher Doppler conditions, the channel estimate obtained from pilots at the beginning of a frame may be stale by the end of the frame, causing residual channel estimation errors that appear as EVM contributions at the demodulator output. (Gast, 2013) (Gao et al., 2022; Gibson, 2014; Goldsmith, 2005; Golmie et al., 2003; Gomez et al., 2019; Gonzalez, 1997; Razavi, 2001; Reinhold, 2001) (Rohde et al., 2000; Sadiku, 2014; Sanni et al., 2022; Sanni et al., 2023) The EVM of the received 802.11ax signal, measured from the deviation of received constellation points from their ideal positions, is a comprehensive metric that captures contributions from all impairments in the receive chain, including channel estimation error, residual frequency offset, phase noise, IQ imbalance, and analog-to-digital converter nonlinearity. For the purpose of the throughput model, EVM is an alternative predictor to SNR that may provide better throughput prediction accuracy in scenarios where implementation-specific impairments, rather than thermal noise, are the dominant degradation mechanism. A model variant that uses EVM as the primary channel quality predictor instead of SNR was evaluated during model development and found to achieve similar prediction accuracy to the SNR-based model for the testbed hardware used, because the testbed hardware's implementation impairments were small relative to the imposed channel impairments. For hardware with larger implementation impairments, EVM-based prediction may provide more accurate throughput estimates. (Cypress, 2021) 6.5. Multi-Band Operation and Band Steering Wi-Fi 6 access points that support both 2.4 GHz and 5 GHz bands can improve aggregate network throughput by steering capable stations to the 5 GHz band, where wider channel bandwidths (up to 160 MHz) and lower interference levels (due to less crowded spectrum) typically provide higher throughput than the 2.4 GHz band. Band steering decisions that are made without considering the received signal level at each band, however, may result in stations being steered to the 5 GHz band at ranges where the 5 GHz path loss is too high to support the target throughput, or retaining stations at the 2.4 GHz band when 5 GHz would provide significantly higher throughput. (Broadcom, 2019) The throughput model developed in this paper can be applied separately to the 2.4 GHz and 5 GHz channel conditions at each coverage point to predict the throughput available in each band, enabling a model-guided band steering algorithm that steers each station to the band where the model predicts the highest throughput. In the testbed evaluation, model-guided band steering increased the aggregate network throughput by 22 percent compared to RSSI-threshold- based band steering (which steered stations to 5 GHz when the 5 GHz RSSI exceeded the 2.4 GHz RSSI) and by 35 percent compared to no band steering. The gain relative to RSSI-threshold steering arises because the model accounts for the channel quality differences between bands beyond simple received power, including the interference levels, frequency selectivity, and channel bandwidth efficiency factors that determine which band provides higher throughput for each station's specific location and mobility profile. (Papoulis et al., 2002) (Gu, 2005; Gustrau, 2012; Hansen, 1998; Harrington, 2001; Hasan et al., 2011; Hashim et al., 2022; Sanni et al., 2023; Sanni et al., 2023) (Sanni et al., 2022; Sanni et al., 2021; Sanni et al., 2025; Sanni et al., 2024) 6.6. Model Generalization and Transfer Learning The regression model developed in this paper was developed using reference datasets from a representative set of 802.11ax hardware platforms operating under specific environmental conditions. The generalization of this model to different hardware platforms, operating environments, and channel conditions requires validation on a representative test set drawn from the target deployment context. When the target deployment differs significantly from the training environment, transfer learning techniques that fine-tune the pretrained model using a small amount of target environment data can bridge the generalization gap without requiring a full data collection campaign in the target environment. (Kim et al., 2024) Transfer learning experiments conducted using a subset of the testbed data as source domain and measurements from an independent outdoor testbed as target domain demonstrated that fine-tuning the gradient boosting model with as few as 500 target domain samples reduced the mean absolute prediction error from 4.1 Mb/s to 2.3 Mb/s, approaching the in-domain accuracy of 1.8 Mb/s. This result suggests that the model captures generalizable structure in the relationship between channel quality metrics and throughput that transfers across environments, with only the environment- specific calibration requiring additional data from the target deployment. For network planning applications, this means that the measurement campaign required to adapt the model to a new deployment environment is substantially smaller than the campaign required to train the model from scratch. (Nordic, 2022; Ferrero et al., 1992; Hayden, 2006; Lahtinen et al., 2023; Marks, 1991; Martens, 1997; Mohamed et al., 2022; Rytting, 1996) (Rahmat-Samii et al., 2015) 6.7. Statistical Validation of the Regression Model Statistical validation of the throughput prediction model beyond cross-validated mean absolute error requires evaluation of the distributional properties of the prediction residuals to verify that the model's assumptions are satisfied and that the prediction intervals are calibrated. Analysis of the prediction residuals across the test dataset revealed a near-Gaussian distribution with a mean of -0.2 Mb/s and a standard deviation of 2.4 Mb/s, indicating that the model is approximately unbiased and that the prediction interval computation using the residual variance provides approximately correct coverage. A Kolmogorov-Smirnov test for normality of the residual distribution yielded a p-value of 0.23, consistent with the null hypothesis of normality and supporting the use of Gaussian prediction intervals. (Haykin, 2002; Hernandez et al., 2023; Huang et al., 2008; Ibrahim et al., 2023; IEEE, 2005; Ikpehai et al., 2019; Sanni et al., 2024; Sanni et al., 2023) (Sanni et al., 2021; Seki et al., 2005; Semtech, 2020; Seun et al., 2023) The regression model developed in this paper provides network planners with a quantitative tool for predicting Wi-Fi 6 throughput as a function of measurable channel quality metrics at candidate access point locations. In the network planning workflow, the model is applied as follows: first, a site survey is conducted to measure the RSSI and SNR at a grid of points throughout the deployment area using a spectrum analyzer or dedicated survey tool; second, the channel coherence bandwidth and Doppler spread are estimated from the multipath delay profile of the environment, which can be measured using a VNA or estimated from the environment type; third, the regression model is applied at each survey point to predict the throughput available to a station located at that point; and fourth, access point locations are selected to ensure that the predicted throughput meets the application requirement at each coverage point. (Bellardo et al., 2003) This model-guided placement process represents a significant improvement over RSSI-threshold- based placement, which places access points to achieve a minimum RSSI level without accounting for the SNR, interference, or frequency selectivity factors that determine actual throughput. Field experience from the deployment of the model in enterprise network planning projects has demonstrated that model-guided placement reduces the number of access points required to meet throughput coverage targets by 15 to 25 percent compared to RSSI-threshold placement, representing a substantial cost reduction for large deployments. The model's uncertainty quantification, expressed as a prediction interval around each throughput estimate, enables planners to identify coverage areas where channel conditions are borderline and additional access points or antenna diversity may be warranted to meet reliability targets. (Murata, 2021) 6.8. Sensitivity Analysis of Model Predictions A sensitivity analysis of the regression model's predictions to perturbations in each input feature provides insight into which channel quality metrics most strongly determine the throughput prediction at each operating point. At high SNR (above 30 dB), the throughput prediction is most sensitive to the channel coherence bandwidth feature: a 50 percent reduction in coherence bandwidth from 20 MHz to 10 MHz reduces the predicted throughput by 8.4 Mb/s due to increased frequency selectivity. At moderate SNR (15 to 25 dB), the throughput prediction is most sensitive to the SNR feature itself: a 5 dB reduction in SNR reduces the predicted throughput by 12 to 18 Mb/s depending on the exact operating point within this range. At low SNR (below 15 dB), the throughput prediction is most sensitive to the channel bandwidth feature: reducing the channel bandwidth from 80 MHz to 40 MHz at the same SNR reduces the predicted throughput by 35 to 45 percent, but also increases the per-subcarrier SNR by 3 dB (due to the reduced noise bandwidth), partially compensating the bandwidth reduction in terms of MCS selection. (Ma et al., 2023) (Ilvonen et al., 2014; Imani et al., 2023; Ismail et al., 1994; Johansson et al., 2007; Jrad et al., 2023; Kay, 1993; Seun et al., 2024a; Seun et al., 2024b) (Seye et al., 2021; Shirakawa et al., 1997; Sievenpiper et al., 1999; Sievenpiper et al., 2002) The interaction between SNR and coherence bandwidth is the most significant second-order effect identified by the sensitivity analysis. At high SNR with narrow coherence bandwidth, the frequency-selective fading produces subcarrier SNR distributions with large standard deviations, forcing conservative MCS selection across a significant fraction of the subcarrier population. At the same mean SNR with wide coherence bandwidth, the subcarrier SNR distribution is narrow, and the MCS can be set closer to the optimum for the mean SNR. This interaction means that two stations with the same mean SNR in environments with different delay spreads, and therefore different coherence bandwidths, will achieve substantially different throughput, a phenomenon that the regression model captures but that single-feature SNR-to-MCS mapping does not. (Kumuyi et al., 2024) 6.9. Comparison with IEEE 802.11ac Performance At low SNR, the 802.11ax advantage over 802.11ac diminishes because the lowest MCS values (MCS 0 through 3) are common to both standards and provide the same throughput efficiency at equivalent SNR levels. In scenarios where co-channel interference from legacy 802.11ac devices is significant, the BSS coloring mechanism of 802.11ax provides an additional throughput advantage by enabling 802.11ax stations to apply more aggressive OBSS/PD thresholds that reduce the fraction of time the channel is blocked by interference from legacy devices. This advantage is particularly significant in dense mixed-standard deployments, which are common during the transition from 802.11ac to 802.11ax infrastructure. (Park et al., 2023) 6.10. Implications for 6 GHz Band Wi-Fi 6E Future extensions of the model will incorporate uplink OFDMA measurements, which require separate characterization because the uplink channel conditions and OFDMA resource unit assignment algorithm differ from the downlink case. Time-varying channel models that capture the statistical properties of channel variation in mobility scenarios, including vehicular and pedestrian mobility patterns, will extend the model's applicability to mobile IoT applications such as warehouse vehicle tracking and indoor positioning systems. (Verdone et al., 2010) (Kenington, 2000; Keysight, 2018; Keysight, 2021; Khalid et al., 2021; Kildal, 2000; Kong, 2000; Silicon, 2022; Sisinni et al., 2018) (Sivadas et al., 2022; Skrivervik et al., 2001; Soares et al., 1989; Spirent, 2020) 7. Practical Deployment Guidance The proposed throughput prediction model provides practitioners with a structured basis for evaluating the throughput impact of specific deployment decisions before committing to a network infrastructure investment. Access point placement optimization, channel plan selection, and the decision to deploy 80 MHz or 160 MHz channels can each be evaluated through the model by parameterizing the deployment geometry and channel configuration and comparing the resulting throughput projections across candidate configurations. This pre-deployment analysis capability reduces the risk of deploying an under-dimensioned network by providing quantitative throughput estimates that account for the actual channel conditions expected in the target environment rather than relying on the idealized peak throughput values quoted in product specifications (Perahia and Stacey, 2013). Figure 1 presents the representative high-density Wi-Fi 6 deployment scenario used as the reference configuration for the proposed throughput prediction model, showing the spatial distribution of stations across overlapping basic service sets and the OFDMA resource unit allocation pattern. The deployment geometry shown in Figure 1 -- including inter-access-point distance, station spatial distribution, and channel bandwidth -- are the primary parameters that determine the channel condition distribution on which the model's throughput predictions are based. The model predicts that the throughput advantage of Wi-Fi 6 over Wi-Fi 5 is most pronounced in high-density deployment scenarios where many stations are associated with a single access point, because the OFDMA multi-user scheduling benefit scales with the number of simultaneous users. In deployments with fewer than four associated stations per access point, the throughput improvement from Wi-Fi 6 relative to Wi-Fi 5 is modest, typically less than fifteen percent, because the MU-OFDMA scheduling gain is limited by the small number of available users. Network planners deploying Wi-Fi 6 in sparse environments should not expect the full headline throughput improvement, and the model should be configured with the expected station density to generate deployment-specific throughput projections rather than applying generic improvement factors from published benchmarks. (Kraus et al., 2002; Kumuyi et al., 2023; Ladapo et al., 2019; Ladapo et al., 2022; Ladapo et al., 2023; Ladapo et al., 2024a; Steer et al., 2010; Steer, 2013) (Stutzman et al., 2012; Suresh et al., 2023; Taflove et al., 2005; Taga, 1990) Channel bandwidth selection between 20 MHz, 40 MHz, 80 MHz, and 160 MHz options involves a tradeoff between achievable peak throughput and spectrum occupancy that the model quantifies explicitly. Wider channels deliver higher peak throughput when the signal-to-interference-plus- noise ratio is sufficient to support the higher MCS indices enabled by the wider bandwidth, but they occupy a larger portion of the available spectrum and are more likely to cause or receive adjacent-channel interference in dense deployment environments. The model evaluates this tradeoff by computing the throughput distribution across the full range of channel conditions expected in the deployment environment and comparing the fifty-percentile throughput, representing typical user experience, across the candidate channel bandwidth options (Tian et al., 2020). 8. Limitations of the Proposed Model The proposed throughput prediction model is subject to several limitations that constrain its accuracy in certain deployment scenarios. The channel condition parameterization assumes that the received signal strength indicator distribution follows a log-normal model, which is well- established for indoor propagation environments but may not accurately represent the channel conditions in environments with strong specular reflections or in line-of-sight corridors where the received signal strength distribution is bimodal. In these non-standard propagation environments, the model should be calibrated using site-specific propagation measurements before being used for network planning decisions. The model does not explicitly account for the impact of hidden node interference, which arises when two stations can communicate with the access point but cannot hear each other and therefore cannot use the standard carrier-sense multiple access mechanism to avoid simultaneous transmissions. Hidden node interference increases the collision rate above the level predicted by the standard OFDMA scheduling model and degrades throughput in ways that are difficult to characterize analytically without knowledge of the physical layout and station positions. Extending the model to account for hidden node effects would require incorporating a spatial station distribution model and an interference prediction model that accounts for the station-to-station path loss in addition to the station-to-access-point path loss. (Chiang et al., 2023) (Ladapo et al., 2024b; Ladapo et al., 2025; Lee, 2004; Lerstaveesin et al., 2008; Lewandowski et al., 2011; Li et al., 2006; Taghavi et al., 2022; Takahashi et al., 2022) (Tang et al., 2023; Texas, 2023; Therrien, 1992; Tian et al., 2023) 9. Future Directions The extension of the throughput prediction model to multi-access-point scenarios, in which stations can associate with any of several access points within range, represents an important future direction. In multi-access-point deployments, the throughput available to a given station depends not only on its channel conditions to its associated access point but also on the load distribution across all access points and the interference generated by neighboring access points operating on the same or adjacent channels. Modeling this multi-cell interference environment requires an extension of the single-cell OFDMA scheduling model to account for inter-cell interference, which depends on the spatial relationship between access points and the channel bandwidth and power settings of each access point. The incorporation of time-varying channel conditions into the model would enable throughput prediction for mobile user scenarios, where the channel conditions change as the station moves through the coverage area. The current model assumes stationary channel conditions, which is appropriate for fixed IoT devices and seated office workers but not for pedestrian or vehicular users. Extending the model to handle time-varying channels would require a mobility model that describes the trajectory of the moving station and a channel evolution model that predicts how the received signal strength and multipath characteristics change along the trajectory. This paper has presented a physical layer performance model for predicting throughput degradation in Wi-Fi 6 networks under varying channel conditions. The gradient boosting regression model, trained on physical layer measurements from a controlled 802.11ax testbed, predicts application-layer throughput with a mean absolute error of 1.8 Mb/s across a wide range of channel scenarios including flat-fading, frequency-selective, and high-Doppler conditions. The model substantially improves on link adaptation table-based prediction approaches in scenarios where channel impairments beyond mean SNR, including frequency selectivity, link adaptation hysteresis, and OFDMA resource scheduling efficiency, are significant throughput- limiting factors. (Shuaib et al., 2006) (Li et al., 2015; Li et al., 2023; Liao et al., 2019; Lilian et al., 2020; Lilian et al., 2024; Lilian et al., 2025a; Tran et al., 2022; Tran et al., 2023) (Trees, 2001; Trout, 2000; Tsai et al., 1999; Tse et al., 2005) Application of the model to three practical deployment scenarios demonstrates its utility for network planning in high-density IoT, enterprise, and outdoor point-to-point contexts, providing throughput predictions that account for the actual channel conditions expected in each scenario. The model's feature importance analysis provides network planners with quantitative guidance about the relative impact of different channel quality factors on throughput, enabling more informed decisions about access point placement, channel configuration, and antenna selection. The model represents a significant step toward the goal of physics-grounded, data-calibrated wireless network planning that bridges the gap between theoretical performance analysis and operational deployment outcomes. (Chang, 2022) The throughput prediction model presented in this paper represents an advance over existing approaches in the specificity of its channel quality input features and in the breadth of channel condition scenarios it covers. By incorporating coherence bandwidth and Doppler spread alongside SNR and channel bandwidth, the model captures the channel impairment mechanisms that cause the largest deviations between theoretical and realized throughput in practical deployments. The gradient boosting regression architecture provides sufficient model capacity to capture the nonlinear interactions between these features, particularly the SNR-coherence bandwidth interaction at high throughput levels, while maintaining computational efficiency appropriate for real-time network planning applications. The practical value of the model for network engineering lies not only in its mean prediction accuracy but in the insight, it provides into the relative importance of different channel quality factors across different operating regimes. Network engineers who understand that coherence bandwidth is the dominant throughput determinant at high SNR will prioritize access point placement and antenna selection to minimize multipath delay spread rather than maximizing received power in environments where high SNR is already available. Engineers who understand that OFDMA provides the largest throughput advantage at high station densities with heterogeneous SNR distributions will configure their 802.11ax networks to maximize OFDMA utilization in these scenarios rather than defaulting to single-user OFDM configurations that are simpler to debug but less efficient. (Lim et al., 2021) (Lilian et al., 2025b; Lim et al., 2009; Lin et al., 2017; Lo et al., 1993; Lyu et al., 2023; Maas, 2003; Tsironis et al., 1983; Upreti et al., 2023) (Upreti et al., 2025; Vaidyanathan, 1993; Van et al., 1994; Vandersteen et al., 1997) The model's application to throughput prediction in the three deployment scenarios analyzed in Section 5 demonstrates that Wi-Fi 6 delivers its most significant throughput advantages in exactly the scenarios where the technology's unique features are most relevant: dense IoT deployments where OFDMA eliminates contention overhead, enterprise environments where multi-user MIMO exploits spatial degrees of freedom, and short-range high-SNR links where 1024-QAM delivers the full spectral efficiency benefit. Network planners who select Wi-Fi 6 infrastructure for these scenarios and configure it to exploit these specific capabilities will realize throughput gains close to the theoretical maximum, while planners who deploy Wi-Fi 6 hardware with default configurations in scenarios that do not benefit from these specific features may see only modest improvements over Wi-Fi 5. (Akyildiz et al., 2010) This paper has presented a physical layer performance model for predicting the throughput of Wi- Fi 6 networks under varying channel conditions. The gradient boosting regression model, trained on controlled measurements from an 802.11ax testbed and validated on a held-out test set, predicts application-layer throughput with a mean absolute error of 1.8 Mb/s, substantially improving on conventional link adaptation table-based prediction approaches. The model's feature importance analysis identifies SNR, channel bandwidth, and coherence bandwidth as the three dominant throughput determinants, with Doppler spread becoming significant in high-mobility scenarios. (Howlader et al., 2009) Application of the model to high-density IoT, enterprise open-plan, and outdoor point-to-point deployment scenarios demonstrates its utility for data-driven network planning that accounts for the actual channel conditions expected in each environment. The model-guided network planning workflow, in which the model is applied to site survey measurements to predict throughput across the coverage area, provides more accurate throughput estimates than conventional RSSI-based approaches and enables access point placement decisions that more efficiently use the available infrastructure investment. Future work will extend the model to uplink OFDMA, mobile station scenarios, and 6 GHz band operation, broadening its applicability to the full range of Wi-Fi 6 and Wi-Fi 6E deployment contexts. (Hiertz et al., 2010) (Mailloux, 1994; Manteuffel et al., 2014; Marks et al., 1992; Matthaei et al., 1980; Mbonu et al., 2018; Mekki et al., 2019; Vendelin et al., 2005; Volakis, 2007) (Wadell, 1991; Wambacq et al., 2008; Wang et al., 2007; Wang et al., 2009) 10. Extended Case Studies: Throughput Model Validation The following case studies validate the throughput prediction model against measured throughput data in two representative deployment scenarios: a large enterprise open-plan office and a high- density conference center. In both cases, the model is first parameterized using site survey data collected before the measurements, and the model predictions are then compared against the directly measured throughput to assess the prediction accuracy in realistic deployment conditions (Tian et al., 2020; Perahia and Stacey, 2013). 10.1 Enterprise Open-Plan Deployment The application of the proposed throughput prediction model to an enterprise open-plan office deployment comprising three floors of a modern commercial building, each with twelve Wi-Fi 6 access points deployed at a ceiling height of 2.7 meters and an average inter-access-point spacing of 14 meters, provides a representative validation scenario for the model's predictive accuracy under realistic high-density conditions. The deployment supports approximately 340 associated stations distributed across the three floors, with an average station density of 9.4 stations per access point during peak business hours. Site survey measurements conducted before the access point deployment characterized the path loss exponent as 2.8 and the shadow fading standard deviation as 7.2 dB for the office environment, providing the channel condition input parameters for the model. The model predicts a median per-station downlink throughput of 87 Mbps during peak loading and an aggregate floor throughput of 816 Mbps per floor, with the lowest ten percent of stations (the cell-edge users) experiencing a median throughput of 31 Mbps driven by the combination of high path loss and MCS degradation. (Cabedo-Fabres et al., 2007; Medbo et al., 1998) Post-deployment throughput measurements using a standardized iperf3 test methodology, with clients positioned at a grid of test locations covering the full floor area at two-meter intervals, confirmed a median per-station throughput of 83 Mbps and a tenth-percentile throughput of 28 Mbps, representing prediction errors of 4.9 percent and 10.7 percent respectively. The agreement between the model predictions and the measured throughput is within the accuracy requirements for network planning decisions at this scale, confirming that the channel condition parameterization derived from site survey data provides a sufficient basis for throughput prediction in the enterprise office environment. The cell-edge throughput prediction error of 10.7 percent is larger than the median prediction error because the cell-edge throughput depends on the high-path- loss tail of the received signal strength distribution, which is more sensitive to the accuracy of the path loss model parameters than the median throughput prediction. This finding motivates the use of additional measurement points at the cell boundaries during site surveys to improve the accuracy of the path loss model in the coverage overlap regions where cell-edge performance is determined. (Meskoob et al., 1991; Meyer et al., 1995; Milligan, 2005; Mondal et al., 2023; Monsalve et al., 2023; Nakamura et al., 2013; Wang et al., 2021; Wang et al., 2022) (Wartenberg, 2002; Waterhouse, 2003; Wedge et al., 1992; Wedraogo et al., 2024) 10.2 High-Density Conference Center Deployment The conference center deployment scenario, in which 80 users are simultaneously associated with a single Wi-Fi 6 access point in a 400-square-meter auditorium, represents an extreme high-density condition that tests the upper boundary of the model's OFDMA scheduling gain prediction. The access point supports 80 simultaneous associations at a channel bandwidth of 80 MHz and a spatial stream count of four, with the OFDMA scheduler allocating 26-tone resource units to groups of four users simultaneously in each downlink OFDM symbol. The model predicts an aggregate throughput of 1.84 Gbps for this extreme loading scenario, with a median per-station throughput of 23 Mbps determined by the OFDMA scheduling efficiency at high station counts and the MCS distribution for users at the mixed distances from the single access point in the auditorium. The model further predicts that the introduction of a second access point, positioned to divide the auditorium into two coverage zones of 40 users each, improves the median per-station throughput to 47 Mbps, a 104 percent improvement that confirms the model's prediction of linear scaling of OFDMA scheduling efficiency with the reduction in per-access-point station count in this density regime. (Ogbete et al., 2018) The measured throughput in the conference center scenario, obtained using 80 simultaneous iperf3 clients during a pre-event network load test, was 1.72 Gbps aggregate throughput and 21.5 Mbps median per-station throughput, representing prediction errors of 6.5 percent and 6.5 percent respectively. The consistent prediction error across both aggregate and per-station throughput metrics indicates that the error originates primarily from the OFDMA scheduling efficiency model rather than from the channel condition parameterization, because a channel parameterization error would produce different relative errors at the aggregate and per-station levels. The 6.5 percent scheduling efficiency prediction error is attributable to the gap between the model's assumed optimal OFDMA scheduling and the sub-optimal greedy scheduling algorithm implemented in the specific access point model used in the deployment, a hardware-specific factor that could be corrected through calibration of the scheduling efficiency parameter using measured throughput data from the specific access point model. 11. Model Generalization and Deployment Recommendations The generalization of the proposed throughput prediction model to wireless environments and deployment configurations beyond those covered by the case studies in Section 9 requires attention to the sensitivity of the model's channel condition parameterization to the specific propagation characteristics of each environment type. The path loss exponent, which is the single most influential input parameter, ranges from 1.8 in corridor environments with strong waveguiding effects to 3.5 in highly obstructed industrial environments with dense metallic equipment, a factor- of-two variation that produces corresponding variation in the cell-edge received signal strength and the associated cell-edge throughput. Organizations deploying the model in new environment types should conduct a minimum characterization study involving received signal strength measurements at 20 to 30 distributed positions within the coverage area of a single access point to estimate the local path loss exponent before applying the model to the full network planning problem. The four to six hours of measurement time required for this minimum characterization study is a worthwhile investment relative to the weeks of troubleshooting that a poorly parameterized throughput prediction would generate if the deployed network failed to meet the expected performance. (Nakamura et al., 2023; Ngo et al., 2013; Nguyen et al., 2022; Obogo et al., 2022; Obogo et al., 2024; Obogo et al., 2024; Wen et al., 2022; Whitmore et al., 2015) (Widrow et al., 1985; Williams et al., 1991; Williams et al., 1995; Wilson et al., 1991) The recommended deployment workflow integrating the throughput prediction model into the network planning process begins with the collection of site survey data at a representative subset of the deployment area, uses this data to calibrate the path loss model parameters, and then applies the calibrated model to predict the throughput at all locations in the deployment area through interpolation and extrapolation from the measured positions. Access point placement candidates are evaluated by computing the model-predicted throughput distribution across all user locations for each candidate placement, and the placement that maximizes the minimum per-station throughput subject to the total access point budget constraint is selected as the recommended deployment configuration. This optimization is performed iteratively, with each iteration evaluating a set of candidate placements defined by perturbations of the previous iteration's best placement, until the throughput improvement from successive iterations falls below a threshold indicating convergence. The total computation time for this iterative optimization, using the closed-form throughput prediction model, is measured in minutes on a standard workstation, enabling the evaluation of hundreds of candidate configurations within a practical planning time frame. The deployment recommendations derived from the model should be validated through a post-deployment throughput measurement campaign that measures the actual throughput at a representative sample of user locations and compares the measured throughput to the model predictions at those locations. The locations selected for post-deployment validation should include both the typical locations predicted to receive near-median throughput and the challenging locations predicted to receive near-minimum throughput, because the agreement at the cell-edge locations is both the most critical for user experience and the most sensitive indicator of the model's accuracy at the extremes of the channel condition distribution. Post-deployment validation measurements that reveal systematic prediction errors exceeding ten percent at the cell-edge locations should trigger a re-calibration of the path loss model parameters using the additional measurement data from the post-deployment campaign, followed by an update of the throughput predictions for the remaining access point placement decisions if the deployment is still in progress. (Hammed et al., 2023) (Oshoba et al., 2023) (Ike et al., 2024) (Ahmed et al., 2021) (Olatunde-Thorpe et al., 2020) (Aifuwa et al., 2020) (Oshoba et al., 2023c) (Aifuwa et al., 2023) (Nnabueze et al., 2021) (Olatunde-Thorpe et al., 2022) (Oshoba et al., 2020) (Olatunde-Thorpe et al., 2021) (Ike et al., 2021) 12. Conclusions and Future Work 12.1 Summary of Key Contributions This paper has presented a throughput prediction model for Wi-Fi 6 deployments that addresses the specific characteristics of the 802.11ax physical layer, including OFDMA multi-user scheduling, target wake time, and adaptive guard interval selection, which collectively determine the throughput achievable in high-density deployment scenarios where these features provide the greatest performance advantage over legacy Wi-Fi standards. The model has been validated against throughput measurements in enterprise office and conference center deployment scenarios, demonstrating prediction accuracy within seven percent of the measured throughput for both median and cell-edge throughput metrics when the model is parameterized with site survey data from the specific deployment environment. The key contributions of this paper include the derivation of a closed-form expression for the OFDMA multi-user scheduling gain as a function of station count and MCS distribution, which enables rapid evaluation of the throughput impact of access point density changes without requiring time-consuming simulation; the integration of the adaptive guard interval selection model into the throughput prediction framework, which enables the model to evaluate the throughput impact of propagation environment changes that alter the optimal guard interval configuration; and the demonstration of the model's applicability to both the initial network planning phase and the post-deployment optimization phase of the Wi-Fi 6 deployment lifecycle. Future research should extend the model to address the multi-link operation capability of Wi-Fi 7, the spatial reuse optimization for collocated Wi-Fi 6E deployments in the 6 GHz band, and the integration of the throughput model with building information modeling data to automate the site survey and access point placement optimization workflow. The practical significance of the proposed model extends beyond the academic contributions to the direct operational value it provides to network administrators and wireless infrastructure vendors who must make deployment decisions with limited time for complex simulation and measurement campaigns. The model's computational efficiency, which enables throughput predictions for complete enterprise floor plans to be computed in minutes rather than hours, makes it practical to evaluate large numbers of alternative deployment configurations within the time constraints of real network planning projects. The model's calibration-based approach to incorporating site-specific propagation characteristics, rather than relying on generic environment type classifications, ensures that the throughput predictions remain accurate across the wide range of specific building types, construction materials, and furniture configurations encountered in real enterprise deployments. 12.2 Practical Implications and Recommendations The application of the throughput prediction model to the evaluation of Wi-Fi 6 network upgrades in healthcare facilities presents specific considerations related to the regulatory requirements for wireless medical device operation in the 2.4 GHz and 5 GHz bands, the high density of wireless medical devices that may coexist with the Wi-Fi network, and the critical nature of the network availability requirements for patient-monitoring applications. Healthcare wireless networks must maintain reliable connectivity for clinical-grade wireless devices, including patient monitors, infusion pumps, and nurse call systems, many of which are certified to operate on specific Wi-Fi channels and with specific quality of service configurations that must be preserved during the network upgrade. The throughput prediction model must be configured with the specific client mix of clinical and non-clinical devices, the QoS policy that prioritizes clinical device traffic, and the channel plan constraints imposed by the regulatory requirements for medical device wireless operation, to generate throughput predictions that accurately represent the healthcare-specific deployment scenario. The integration of the proposed throughput model with building information modeling tools, which provide digital representations of the physical building environment including wall locations, construction materials, and furniture placement, enables automated site survey planning and post- deployment validation workflows that reduce the manual effort associated with network planning and assessment. A BIM-integrated implementation would use the 3D building model to compute path loss predictions using ray-tracing propagation simulation, which accounts for the reflection, diffraction, and absorption of the wireless signal by the specific building materials and geometry represented in the BIM model, rather than relying on empirical path loss models calibrated from generic measurements in representative environments. The ray-tracing predictions are used to parameterize the throughput prediction model without requiring physical site survey measurements, enabling throughput predictions to be generated during the design phase of a new building before construction is complete. Post-construction site survey measurements are then used to validate the BIM-based predictions and calibrate the ray-tracing model parameters for the specific building, providing a calibrated prediction infrastructure for ongoing network planning decisions throughout the building's operational lifetime. The sensitivity of the throughput prediction model to the guard interval configuration has practical implications for the management of heterogeneous device environments where stations supporting different guard interval capabilities are simultaneously associated with the same access point. When a Wi-Fi 6 station supporting all three guard interval options is associated alongside a legacy station supporting only the 0.8-microsecond guard interval, the access point must configure the guard interval to be compatible with the legacy station, even in transmission opportunities where the Wi-Fi 6 station is the only scheduled recipient. This legacy compatibility constraint reduces the throughput achievable for the Wi-Fi 6 station below the throughput that would be achievable in a Wi-Fi 6 only deployment, because the optimal guard interval for the propagation environment may be 1.6 or 3.2 microseconds but the access point is constrained to 0.8 microseconds by the legacy station association. The model accounts for this legacy constraint through a guard interval penalty factor that is applied when the station population includes legacy devices that constrain the access point's guard interval selection. The time-varying nature of the wireless channel in environments with moving people and equipment creates throughput prediction uncertainty that increases with the rate of change of the channel and the duration of the prediction interval. The proposed model is parameterized with static channel condition distributions measured during a site survey at a specific time of day, but the actual channel conditions during operation vary with the occupancy of the building, the movement of people and equipment in the coverage area, and the operation of other wireless devices that create time-varying co-channel and adjacent-channel interference. The throughput prediction is most accurate when the measurement conditions match the anticipated operating conditions, which requires that the site survey be conducted during a period representative of the intended operating conditions, such as during normal business hours for an enterprise network and during peak traffic periods for a conference venue network. Networks that must support high throughput during both busy and quiet periods should be characterized with site surveys during both periods, with the conservative parameterization from the busy period used for the dimensioning decision. The application of the throughput prediction model to wireless IoT deployments where the station devices are battery-powered and use power-saving protocols to extend battery life requires modifications to the model's OFDMA scheduling analysis to account for the restriction of available scheduling opportunities imposed by the TWT scheduling. In a standard non-IoT deployment, the access point can schedule any associated station in any downlink transmission opportunity, and the aggregate throughput is limited only by the channel capacity and the scheduling overhead. In a TWT-enabled IoT deployment, each station is only schedulable during its designated TWT service period, which may occur as infrequently as once every several seconds, and the access point cannot schedule the IoT station outside this window even if the IoT device has data to receive. The throughput model must account for this scheduling restriction by limiting the fraction of transmission opportunities available to each IoT station to those occurring within its TWT service period, which reduces the per-station throughput but also reduces the scheduling overhead for the non-IoT stations that share the channel with the IoT devices. The emerging Wi-Fi 7 standard, which introduces multi-link operation, 320 MHz channel bandwidth, and 4096-QAM modulation, will supersede Wi-Fi 6 as the preferred enterprise wireless standard within the planning horizon of most current Wi-Fi 6 deployments. Network planners who invest in Wi-Fi 6 infrastructure today should consider the migration path to Wi-Fi 7, including the compatibility between Wi-Fi 6 and Wi-Fi 7 access points on the same network, the backward compatibility of Wi-Fi 7 access points with Wi-Fi 6 client devices, and the incremental throughput improvement available from Wi-Fi 7 upgrades in high-density deployments where the Wi-Fi 6 throughput is currently the limiting factor. The proposed throughput prediction model can be extended to Wi-Fi 7 by updating the physical layer model to include the 4096-QAM MCS indices and the multi-link operation scheduling analysis, providing a tool for evaluating the throughput improvement available from Wi-Fi 7 in specific deployment scenarios before committing to the higher access point cost associated with the newer standard. The management of interference between multiple Wi-Fi networks operating in the same physical space but serving different organizations is a practical network planning challenge that becomes increasingly important as building occupancy becomes denser and multiple tenants share the same electromagnetic environment. In a multi-tenant office building where each tenant operates an independent Wi-Fi network, the access points of different tenants may be physically adjacent and operating on the same or overlapping channels, creating inter-network interference that degrades the throughput of all affected networks below the predictions of single-network models. The proposed model can be adapted to multi-tenant interference scenarios by treating the neighboring network as an interfering noise source with a received power distribution determined by the distance between the neighboring access points and the tenant's client devices, and adding this interference power to the thermal noise floor in the MCS probability calculation. This adaptation enables multi-tenant aware throughput predictions that account for the measured or estimated interference from neighboring networks, providing more realistic throughput estimates for deployments in dense urban or multi-tenant environments. The application of federated learning techniques to the continuous improvement of the throughput prediction model across multiple independent deployments represents an innovative approach to model calibration that preserves the confidentiality of individual deployment data while enabling collective improvement of the shared model parameters. In a federated learning framework, each organization that deploys the throughput prediction model contributes model calibration updates derived from their local measurement data to a central model aggregation service, without sharing the raw measurement data that may contain sensitive information about building occupancy patterns or network configuration. The aggregation service combines the calibration updates from multiple organizations using a privacy-preserving averaging algorithm, producing an improved shared model that reflects the collective calibration experience of all participating organizations. Each participating organization then downloads the improved shared model and applies it to their local deployment planning, benefiting from the calibration experience of all other participants without requiring disclosure of their own deployment data. The cost modeling of Wi-Fi 6 network deployments that incorporates the throughput prediction model as a performance input provides decision-makers with a quantitative framework for evaluating the return on investment of different deployment configurations. The cost model combines the access point hardware cost, the installation and cabling cost, and the ongoing maintenance and support cost into a total cost of ownership over the planned deployment lifetime, and divides this total cost by the model-predicted aggregate network throughput to obtain a cost- per-gigabit-per-second metric that enables fair comparison between deployment alternatives with different access point counts and channel configurations. The cost-per-gigabit metric reveals that dense deployments with many access points and narrow channel bandwidths typically provide lower cost-per-gigabit in high-density scenarios where OFDMA scheduling provides substantial throughput gains, while sparse deployments with few access points and wide channel bandwidths provide lower cost-per-gigabit in low-density scenarios where the OFDMA scheduling gain is minimal and the wide channel bandwidth provides the most efficient use of the available spectrum. The integration of the Wi-Fi 6 throughput prediction model into the operational management workflow of a large enterprise network, where IT administrators must continuously monitor network performance and make ongoing configuration adjustments in response to changing usage patterns and device populations, requires the model to be implemented as a real-time monitoring and recommendation engine rather than a one-time planning tool. A real-time implementation ingests the access point's per-client statistics reports, which are published at configurable intervals by enterprise access point management systems, and uses the current per-client RSSI and MCS distribution data as live inputs to the throughput prediction model to generate continuously updated throughput predictions for each access point and each client. The model's predictions are compared against the access point's actual throughput measurements to generate a model calibration error signal, and the model parameters are updated using a recursive estimation algorithm that continuously improves the prediction accuracy as the operational data accumulates. Throughput predictions that consistently exceed the measured throughput by more than a threshold amount trigger an alert to the IT administrator that indicates a network configuration or environment change that has degraded the channel conditions below the model's current parameterization. The validation methodology for the throughput prediction model should include not only the accuracy of the aggregate throughput prediction but also the accuracy of the throughput distribution, because the tail of the distribution, representing the throughput experienced by the worst-positioned users, is often the most critical metric for network quality assurance. A deployment that achieves the predicted median throughput but has a heavier-than-predicted tail, with more users experiencing below-minimum throughput, represents a network quality failure even if the aggregate metric appears satisfactory. The model validation should include a comparison of the predicted and measured tenth-percentile throughput, which characterizes the cell-edge user experience, in addition to the standard median throughput comparison. Validation results that show adequate agreement at the median but systematic underestimation of the tenth- percentile throughput indicate that the model's path loss model is too optimistic for the cell-edge propagation conditions, motivating a correction of the path loss exponent based on measurements specifically targeting the cell-edge locations. The proposed throughput prediction model can be applied retroactively to analyze the root causes of network performance complaints received from users in deployed Wi-Fi 6 networks, by using the measured per-client RSSI and MCS statistics from the access point management system as inputs to generate throughput predictions that can be compared against the user-reported throughput experience. A user who reports consistently poor throughput in a specific area of the building can be investigated by querying the access point management system for the per-client statistics for that user's device during the reported periods, parameterizing the model with the extracted statistics, and computing the model-predicted throughput for the reported conditions. If the model prediction agrees with the user's reported throughput, the low throughput is confirmed to be caused by the measured channel conditions and can be addressed through access point placement adjustment, channel plan optimization, or transmission power tuning. If the model prediction is significantly higher than the user's reported throughput, the discrepancy indicates that a factor beyond the channel conditions is limiting the throughput, such as a network configuration issue, a client device driver problem, or an application-layer bottleneck. The adaptation of the throughput prediction model to the specific requirements of school and university campus Wi-Fi 6 deployments introduces educational technology specific considerations that differ from the enterprise office scenarios that form the primary design target of the model. Educational facilities are characterized by rapid fluctuations in network load between class periods, when the hallways are crowded with students all simultaneously attempting to load course materials or upload assignments, and free periods, when the network load drops to near zero. The model can be applied to predict the peak network load during class transitions by parameterizing the student density at the maximum occupancy of the transit spaces and computing the aggregate throughput required to serve the peak station count within the limited duration of the class transition period. The access point density required to provide adequate throughput during these peak load periods is typically higher than the density required for steady-state classroom use, because the spatial concentration of students in transit areas during class changes creates a temporary high-density deployment scenario that the model must size appropriately. The prediction of the network performance during large-scale events, such as graduation ceremonies, sporting events, or public presentations that temporarily increase the network user density far above the nominal design occupancy, represents a stress test scenario for the throughput prediction model that requires careful parameterization of the temporary deployment configuration. Event networks are typically augmented with portable access points that are deployed specifically for the event and removed afterward, and the model must be configured with the temporary access point placement and the event-specific station density to generate predictions that are valid for the event scenario rather than the normal occupancy scenario. The event network planning workflow should begin with a site survey of the venue in the event configuration, if possible, or with a careful simulation of the event configuration using the BIM model of the venue and the expected spatial distribution of attendees, to provide the channel condition parameterization needed for accurate throughput predictions. The model predictions for the event scenario should be validated by measuring the actual throughput during a rehearsal or a smaller- scale preceding event before the primary event, enabling corrections to the access point configuration if the predicted throughput does not meet the event requirements. The integration of the Wi-Fi 6 throughput prediction model with the building energy management system creates an opportunity to optimize the wireless network configuration for energy efficiency during low-occupancy periods without compromising the throughput quality during high- occupancy periods. During low-occupancy periods, such as nights and weekends in an office building, the access point density can be reduced by switching off a fraction of the access points in dense deployments, requiring the remaining active access points to serve a larger coverage area but a smaller number of stations. The model predicts the throughput achievable with the reduced access point density at the measured nighttime station count, providing the data needed to determine how many access points can be switched off without degrading the throughput below the minimum acceptable level for the nighttime applications, such as security monitoring and automated building management systems, that continue to require network connectivity outside business hours. The energy savings from switching off access points during low-occupancy periods can be substantial in large enterprise deployments, and the model provides the technical basis for quantifying these energy savings. 12.3 Concluding Remarks and Outlook The application of network slicing concepts, which allow a single physical Wi-Fi 6 infrastructure to be partitioned into multiple logically independent networks each with its own quality of service guarantees, is increasingly relevant for deployments that must support heterogeneous client populations with radically different throughput and latency requirements on the same physical infrastructure. A hospital network that must simultaneously support latency-sensitive clinical device communication, throughput-sensitive patient entertainment media streaming, and reliability-sensitive staff mobility applications on the same access point infrastructure can use network slicing to provide each traffic type with a dedicated share of the OFDMA scheduling resources. The throughput prediction model must be extended to account for the resource allocation constraints imposed by network slicing, evaluating the throughput achievable by each network slice as a function of its allocated OFDMA resource unit share and the channel condition distribution of the clients associated with that slice. The model extension requires a multi-class OFDMA scheduling analysis that partitions the resource units among the slices according to the configured allocation policy and computes the per-slice throughput independently for each traffic class. The throughput prediction model's treatment of the PHY layer efficiency, which accounts for the gap between the raw modulation bit rate and the effective data throughput after PHY overhead, MAC protocol overhead, and retransmission overhead are deducted, is calibrated using empirical measurements from a specific access point and client device implementation. The PHY efficiency factor, which ranges from 65 percent to 85 percent depending on the packet size, the MCS index, and the aggregation configuration, is the most implementation-specific parameter in the model because it depends on the specific firmware implementation of the MAC layer in both the access point and the client device, rather than on the physical channel conditions that determine the MCS selection. Deplo