PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Calibration of RF Digital Twins via Budget-Constrained Radio Measurement Selection

Yalin E. Sagduyu, Tugba Erpek, Hossein Jafari

arXiv:2610.36406v1cs.NIeess.SP

TL;DR

Physics-based RF digital twins can be systematically biased relative to real measurements, while collecting calibration measurements is costly under strict budgets. This paper compares online strategies for choosing measurements to estimate a calibration offset. Adaptive and structure-aware strategies reduce calibration errors, particularly when budgets are low.

  • Problem

    Systematic bias from geospatial uncertainty, geometric simplifications, and approximate material parameters makes calibration necessary, but radio measurements are costly under strict budgets.

    Source: §I · p. 1

  • Method

    The paper compares online measurement-selection strategies for estimating a global calibration offset, including rank-structured, progressive structure-aware, and uncertainty-aware approaches.

    Source: §I · p. 1, §III · p. 3

  • Results

    AUS achieves the fastest error decay and smallest worst-case error, especially at low budgets, while adaptive and structure-aware schemes improve calibration accuracy with limited measurements.

    Source: §I · p. 1

  • Takeaways & Limitations

    Adaptive and structure-aware selection improves calibration accuracy when radio measurements are limited, particularly in the low-budget regime.

    Source: §I · p. 1, §VI · p. 5

Abstract

from arXiv · show

Radio-frequency (RF) digital twins (DTs) enable site-specific wireless propagation prediction by explicitly reconstructing the physical environment and modeling electromagnetic interactions. Although physics-based DTs reduce reliance on exhaustive measurements, systematic bias relative to real radio data arises from environmental uncertainty, material inaccuracies, and geometric abstractions. Calibration using field measurements is therefore necessary but costly under limited sampling budgets. This paper studies active RF DT calibration, in which measurement locations are selected sequentially and in an online manner to estimate a calibration offset under strict measurement budget constraints. A physics-based RF DT pipeline is developed, integrating semantic geospatial data from OpenStreetMap, 3D modeling in Blender, and deterministic ray tracing in Sionna to capture dominant building and foliage effects. This RF DT is calibrated using path-loss measurements collected with software-defined radios. Multiple online calibration strategies that select radio measurement locations are compared, including random sampling, rank-structured sampling with balanced coverage, quasi-Monte Carlo ranked selection, progressive structure-aware allocation, and fully adaptive uncertainty-driven sampling with estimator stabilization. Results demonstrate that adaptive and structure-aware schemes significantly improve both mean and worst-case calibration accuracy, particularly in the low-budget regime for radio measurements.

I. INTRODUCTION

The paper addresses biased RF digital-twin predictions under costly measurement budgets by selecting calibration measurements online. It develops a physics-based RF digital twin and finds that adaptive and structure-aware selection improves calibration, especially at low budgets.

I. INTRODUCTION

  • Geospatial uncertainty, geometric simplifications, and approximate material parameters create systematic RF digital-twin bias, while radio measurements are expensive and time-consuming.Even when spatial trends are captured, dominant bias can degrade prediction accuracy.Source: §I · p. 1
  • Active calibration selects measurement locations sequentially under a bounded budget, using DT predictions and observed discrepancies to guide later measurements.The study compares five schemes, from random sampling to adaptive uncertainty-aware selection with estimator smoothing.Source: §I · p. 1
  • AUS achieves the fastest error decay and smallest worst-case error, especially at low budgets; adaptive and structure-aware schemes improve calibration accuracy with limited measurements.RS has the largest mean and worst-case errors; BRS, QRS, and PSS improve performance through balanced coverage, systematic ranked coverage, and progressive stabilized allocation, respectively.Source: §I · p. 1

II. PHYSICS-BASED RF DIGITAL TWIN CONSTRUCTION

  • The physics-based RF digital twin combines OSM geospatial data, Blender 3D modeling, and deterministic Sionna ray tracing to represent geometry, materials, and radio-wave interactions.The scene uses geometric abstractions for vegetation and buildings and models propagation paths including line of sight, reflection, diffraction, scattering, and refraction.Source: §II · p. 2

III. DIGITAL TWIN CALIBRATION

The calibration problem estimates a global path-loss offset from a limited set of real measurements. Online estimates are evaluated against an all-location oracle offset that is unavailable to the selection schemes.

III. DIGITAL TWIN CALIBRATION

  • The DT supplies predicted path loss at each candidate location, while the true path loss is available only through a real radio measurement.The residual is the difference between DT prediction and true measurement.Source: §III · p. 3
  • Active calibration uses k of N measured locations to estimate an offset that approximates the oracle offset, which is computed from residuals across all locations.The oracle is used for evaluation but is unavailable to online schemes.Source: §III · p. 3
  • Calibration fidelity is measured by absolute offset error over randomized runs, summarized using mean and worst-case errors.These metrics capture typical accuracy and worst-case deviation from the oracle offset.Source: §III · p. 3

IV. ACTIVE DIGITAL TWIN CALIBRATION SCHEMES

The schemes operate online, selecting unmeasured locations and updating the calibration estimate as radio residuals become available.

IV. ACTIVE DIGITAL TWIN CALIBRATION SCHEMES

  • At each step, a scheme selects an unmeasured location, obtains its radio path loss, forms a residual, and updates the offset estimate.The selection is made from locations not measured in earlier steps.Source: §IV · p. 3
  • Every scheme uses only DT predictions and residuals observed in earlier steps when choosing the next location.Source: §IV · p. 3
  • Each scheme jointly designs measurement selection and offset estimation.Source: §IV · p. 3

A. Random Sampling (RS)

Random Sampling (RS) uses uniform sampling without replacement, while Balanced Rank Sampling (BRS) structures selection by rank to cover the ordered DT prediction range. BRS uses rank reflection but remains non-adaptive.

A. Random Sampling (RS)

  • RS draws locations uniformly without replacement and estimates the calibration offset with the empirical mean.This estimator is unbiased under uniform sampling without replacement, though its variance may be large early in heterogeneous environments.Source: A. Random Sampling (RS) · p. 3

B. Balanced Rank Sampling (BRS)

C. Quasi-Monte-Carlo Ranked Sampling (QRS)

Quasi-Monte-Carlo Ranked Sampling (QRS) selects locations to cover the ordered DT prediction range systematically, using a low-discrepancy sequence. It reverts to uniform selection when few candidates remain.

C. Quasi-Monte-Carlo Ranked Sampling (QRS)

D. Progressive Structure-Aware Sampling (PSS)

Progressive Structure-Aware Sampling (PSS) allocates measurements across prediction bins using both coverage and dispersion, with a fallback to uniform selection as candidates dwindle. Its shrinkage-stratified estimator is adjusted to limit late-budget artifacts.

D. Progressive Structure-Aware Sampling (PSS)

  • PSS partitions locations into empirical-quantile bins and measures within-bin dispersion with the interquartile range.The bins are characterized by their sizes and proportions of the full index set.Source: D. Progressive Structure-Aware… · p. 3
  • PSS balances coverage and dispersion in its target allocation, selecting the bin with the largest allocation deficit and then a point maximizing within-bin DT-space diversity.Source: D. Progressive Structure-Aware… · p. 3
  • PSS falls back to uniform random selection when few candidates remain and uses a shrinkage-stratified estimator with a weight that decreases toward zero as the budget nears its end.The late-stage weighting is intended to prevent artifacts.Source: D. Progressive Structure-Aware… · p. 4

E. Adaptive Uncertainty-Aware Sampling (AUS)

Adaptive Uncertainty-Aware Sampling (AUS) combines residual-informed uncertainty allocation with estimator stabilization. It selects diverse points within bins and applies EWMA stabilization under limited budgets.

E. Adaptive Uncertainty-Aware Sampling (AUS)

F. Complexity

The selection schemes have low online computational cost compared with DT ray tracing and radio measurements, though bin-based methods can become more expensive when bins are highly unbalanced.

F. Complexity

  • Online selection is negligible beside ray tracing and radio measurements; RS is O(k), while BRS and QRS add O(N log N) initial sorting before O(k) selection.PSS and AUS have per-step cost O(B + Nb) and worst-case online complexity O(kN/B), approaching O(kN) with highly unbalanced bins.Source: F. Complexity · p. 4

V. PERFORMANCE EVALUATION

Across measurement budgets, rank-structured and adaptive selection improve RF DT calibration over random sampling, with PSS and AUS providing more reliable convergence and AUS performing especially well at small budgets. The evaluation uses real path-loss measurements and reports mean and worst-case offset errors as the budget grows.

V. PERFORMANCE EVALUATION

  • The evaluation uses USRP B210 radios at 30 receiver locations and 1.2 GHz, with DT simulations at the same locations.The simulations use ray interaction depth 5, 200,000 rays per source, and ITU-based material assignments.Source: §V · p. 4
  • Performance is assessed over randomized runs for each scheme and budget, with selected-location residuals revealed sequentially and mean and worst-case errors plotted against measurement budget.The figures identify the mean absolute offset error and worst-case absolute offset error, respectively.Source: §V · p. 4, Figure 3 · p. 5, Figure 4 · p. 5
  • QRS yields smaller mean and worst-case errors across most budgets through more uniform ranked-axis coverage, while PSS and AUS provide stronger adaptive or stabilized convergence.QRS does not use residual feedback; AUS has the fastest reduction in mean and maximum error in the intermediate regime.Source: §V · pp. 4–5
  • At very small budgets, RS can select unrepresentative extremes and produce large initial bias, while BRS and QRS reduce worst-case error through rank-structured coverage.BRS uses symmetric rank pairing; QRS applies low-discrepancy placement for more uniform coverage of the ranked DT axis.Source: §V · pp. 4–5
  • All schemes converge to zero error as the measurement budget approaches N, but PSS and AUS provide the most predictable late-budget behavior.PSS uses a late-stage fallback and fades structure; AUS maintains stable convergence through estimator smoothing.Source: §V · p. 5
  • PSS combines dispersion-driven allocation and stabilization, while AUS adds residual-informed allocation with EWMA smoothing; Table I summarizes average mean and worst-case error reductions relative to RS.AUS has the fastest error decay at small budgets and smallest worst-case deviations; PSS improves intermediate-regime robustness and convergence reliability.Source: §V · pp. 4–5, Table I · p. 5

VI. CONCLUSION

The paper concludes that active calibration can improve RF DT accuracy with limited measurements, particularly at low budgets, through adaptive and structure-aware selection.

VI. CONCLUSION

  • Adaptive and structure-aware schemes achieve faster error decay and smaller worst-case deviations, particularly in the low-budget regime where measurement efficiency is critical.The conclusion describes a site-specific DT built from semantic geospatial data, 3D scene construction, and deterministic ray tracing, calibrated with limited real radio measurements.Source: §VI · p. 5
Loading 2610.36406v1…