Our dv/v estimates suggest that effective stress changes, attributed to changes in pond levels, have the strongest influence on dv/v, as no meaningful correlations are observed between dv/v and borehole temperature or barometric pressure data. This is consistent with other results27,28. Using a two-dimensional thermoelastic strain model, Tsai27 demonstrated that hydrologic variations likely dominate dv/v estimates over thermoelastic effects. Furthermore, Clements and Denolle28 modelled seasonal thermoelastic strains for dv/v monitoring and found that these strains had a much lower effect on the wavespeed perturbation than hydrological effects. Fokker et al.29 present a physics-based model to show that changes in shear wave velocities are primarily caused by fluctuations in effective stress through changes in the shear modulus. Their approach relies on modelling changes in seismic surface wave velocity using a spectral element method30. In comparison, our model is based on empirical equations (Eq. (7)) from fundamental principles of soil mechanics, effective for near surface applications in granular soils. The application of empirical relationships is attractive for tailings dams, since site-specific sCPT data are typically available.
Our effective stress model is able to replicate the three main trends visible in the dv/v curve (Fig. 2); however, some discrepancies remain. Higher variability in individual dv/v traces is visible from July 15 to 17, 2020. We attribute this variability to nearby construction at the south of the geophone array, resulting in lower waveform coherencies over this period. Furthermore, due to a lack of groundwater level measurements near the geophone array, groundwater levels were inferred using data from the nearby tailings pond. This may also contribute to discrepancies between dVs/Vs predictions and the dv/v curve. Material heterogeneities (e.g. dam fill and tailings) along the seismic wave propagation paths are likely to further increase this discrepancy (Fig. 4).
We attribute the lack of meaningful correlations between dv/v estimates and temperature to the shorter (~ 41 day) data acquisition period. Over this period, the daily mean surface temperature fluctuated between 16˚C and 26˚C, with a mean temperature of 21˚C. Had our data acquisition continued during winter with more extreme temperature changes, these effects may have been visible in dv/v, as observed by others31. We chose to neglect the effects of barometric pressure within our effective stress model, considering the non-linearity of barometric pressures at depth (e.g. damping and time delay effects), dependent on soil properties and groundwater levels32. We also neglected the effects of suction in the unsaturated zone; assuming the soil behaves either fully saturated below the inferred ground water level or fully unsaturated, above the inferred ground water level. Further research and model development may improve understanding of the sensitivity of dv/v measurements to suction in the partially saturated zone.
The depth sensitivity of the dv/v estimates is dependent on the frequency band applied to the cross-correlations (5 to 15 Hz in this work). In an isotropic and homogeneous medium, most of the Rayleigh-wave energy at a frequency f is contained from surface to a depth z at approximately one third of a wavelength λ33. The average Vs in the dam fill (compacted tailings) and underlying tailings can be approximated as 300 m/s and 200 m/s, respectively, based on nearby sCPTs. As such, we estimate an average Vs of ~ 220 m/s over the topmost 39 m (i.e., from surface to bedrock) near our geophone array. Assuming a homogeneous medium, wavelengths for this Vs are between 15 m and 44 m, which suggest depth sensitivity between 5 m and 15 m. Our model estimates a depth sensitivity over the top ~ 16 m, accounting for the inhomogeneity of the structure, with a higher-velocity dam fill layer overlying lower-velocity tailings.
The α and β parameters obtained for the compacted tailings, coarse tailings and clay units were observed to have significantly higher uncertainties than for the fine tailings unit (Fig S-5). This is attributed to an overall lower number of samples and the heterogeneity of these layers. Heterogeneity, such as increased cobbles or gravel, as well as higher resistance, due to compaction of the compacted tailings, prevents advancement of the sCPT. This lowers the reliability of the sCPT measurements obtained through these layers.
Active mine sites are prime locations to study emerging monitoring methods such as ANI, as complementary site information (e.g. sCPTs, historical boreholes, weather station data), can be used to compare and validate results. At this site, data from 52 downhole sCPT measurements was used to parametrize an effective stress model to compare with dv/v estimates. As sCPTs are routinely carried out at many mine sites to assess the liquefaction potential of tailings, incorporation of sCPT data for model constraint allows for site-specific adaptation by mining practitioners. At sites without access to similar datasets, alternative geophysical techniques (e.g. multichannel analysis of surface waves) could be used to estimate site-specific constraints (α, β) to parametrize an effective stress model. Based on the general agreement between the dv/v measurements and the dVs/Vs model, our results demonstrate how ANI can be applied at a tailings dam site to provide highly sensitive (< 1%) measurements of in-situ changes of Vs alongside an approximation of depth sensitivity, without requiring advanced geomechanical models. By comparing the changes in Vs obtained from the model to the dv/v estimates, deviations between the modelled dVs/Vs and the dv/v could be used to indicate potential anomalies (e.g. internal erosion) that aren’t attributable to changes in groundwater levels. This approach could be used to alert to potential areas of concern, indicating where additional inspection and monitoring may be warranted. Furthermore, as Vs measurements are used in the evaluation of the liquefaction resistance of a soil16, combining modelled dVs/Vs with ANI could inform on changes in liquefaction potential of the in-situ tailings material and underlying foundation. For instance, as liquefaction involves a phase-change in the medium from a solid to liquid state, it follows that the shear wave velocity will also decrease dramatically34. Further research of liquefaction-type failures is needed to improve understanding of whether adequate warning time could be provided by monitoring changes in Vs35.
The data processing steps described do not require high computing power, and could be used to efficiently process incoming data for ongoing monitoring purposes. For example, in an operational setting, geophone data could be processed using a three-day rolling average to reduce errors and limit uncertainties14. Combining information on changes in Vs with other monitored geotechnical parameters (e.g. pore pressures, deformation, seepage) could be used to advance overall tailings dam monitoring performance.
This methodology can be expanded towards monitoring greater extents (e.g. many kilometers) along linear infrastructure by combining ANI with distributed acoustic sensing (DAS) fiber optic technologies. At this mine site, existing telecoms optical fiber has been installed in the dam crest and research on the DAS dataset is underway using the methodology presented. The results of this research may have significant implications beyond monitoring tailings dams for other near-surface applications, such as permafrost engineering and landslide investigations. Even though some open questions remain, the advances presented in this manuscript show the potential to use ANI as a quantitative real-time tool and increase our understanding of the temporal evolution of the internal state of tailings dams.