Evaluating Inversion Robustness: A Review of the InterPACIFIC Project Blind Test Results

The 2014 InterPACIFIC (Intercomparison of methods for site PArameter Characterization) project stands as a seminal benchmarking exercise in the field of applied geophysics and geotechnical engineering. Coordinated to assess the reliability and consistency of non-invasive seismic methods, the project invited fourteen international research teams to process identical sets of raw experimental data collected from three geographically and geologically distinct sites in Europe. The primary objective was to quantify the variability inherent in surface wave methods (SWM) when determining the time-averaged shear-wave velocity in the upper 30 meters ($V_{s,30}$), a critical parameter for seismic site classification in building codes such as Eurocode 8.

By employing a blind-test methodology, the organizers ensured that participants remained unaware of the intrusive borehole data available for each site until after their final models were submitted. This approach isolated the influence of human interpretation and algorithmic selection on the resulting velocity profiles. The findings, published across several peer-reviewed reports, revealed that while the extraction of experimental dispersion curves showed high levels of agreement, the subsequent inversion process—where material properties are inferred from wave velocities—introduced significant variability in the estimated subsurface models.

At a glance

• Participating Teams:14 international academic and commercial groups specialized in seismic processing.
• Test Sites:Grenoble (France), Mirandola (Italy), and Lisbon (Portugal).
• Methods Evaluated:Multi-channel Analysis of Surface Waves (MASW), Extended Spatial Autocorrelation (ESAC), Refraction Microtremor (ReMi), and Full Waveform Inversion (FWI).
• Primary Metric:$V_{s,30}$ and the depth to the seismic bedrock.
• Data Types:Passive microtremor recordings and active-source seismic data (hammer/weight drop).
• Outcome:Identification of a 10% to 20% variability in $V_{s,30}$ estimates derived solely from the choice of inversion algorithm and parameterization.

Background

In seismic hazard assessment, the local geological condition plays a decisive role in the amplification of ground motion during earthquake events. The $V_{s,30}$ value serves as a proxy for this amplification, categorizing sites from hard rock to soft soil. Historically, these measurements were obtained through invasive techniques such as cross-hole or down-hole testing, which require the drilling of expensive boreholes. Surface wave methods emerged as a cost-effective alternative, leveraging the dispersive nature of Rayleigh and Love waves in layered media to reconstruct the shear-wave velocity ($V_s$) profile.

However, the application of SWM is mathematically classified as an ill-posed inverse problem. This means that multiple, vastly different velocity-depth models can theoretically produce the same dispersion curve—a phenomenon known as non-uniqueness. Before the InterPACIFIC project, the extent to which this non-uniqueness affected the practical outputs of geophysicists was not fully quantified on an international scale. The project sought to bridge this gap by providing a standardized dataset to evaluate whether different expert teams would arrive at similar conclusions when faced with the same empirical evidence.

The Methodology of the InterPACIFIC Project

The project was structured into two distinct phases: the processing phase and the inversion phase. In the processing phase, teams were provided with raw signal data from geophone arrays. They were tasked with identifying the fundamental and higher modes of surface wave propagation to create a dispersion curve, which plots wave velocity against frequency. The results from this phase showed a surprising level of consensus, suggesting that the signal-processing techniques used to extract information from seismic noise or active sources are relatively strong across different software platforms.

The divergence occurred during the inversion phase. In this step, teams had to define a "search space" of possible earth models, including the number of layers, their thicknesses, and ranges for density and Poisson's ratio. Some teams employed deterministic algorithms, which seek a single best-fit solution through local optimization, while others used stochastic or global search methods, such as genetic algorithms or Monte Carlo simulations, which explore thousands of potential models to find a suite of acceptable solutions.

Site-Specific Challenges and Geological Profiles

The three European sites were selected specifically to test the algorithms under varying degrees of complexity. Each site presented a unique challenge for the inversion of Rayleigh and Love waves.

The Grenoble Site (France)

The Grenoble site is located in an Alpine valley characterized by a high impedance contrast. It features a layer of stiff clay and gravel overlying a very deep, dense crystalline basement. The primary challenge here was the resolution of the deep soil-bedrock interface. Because surface waves are only sensitive to depths proportional to their wavelength, characterizing the deep structure required extremely low-frequency data from large-scale passive seismic arrays. The blind test revealed that while most teams correctly identified the stiff top layer, estimates for the depth of the bedrock varied by as much as 25% among the participating teams.

The Mirandola Site (Italy)

Located in the Po Plain, the Mirandola site consists of a thick succession of alluvial sediments. Unlike Grenoble, the transition from soil to rock is not sharp but follows a gradual increase in velocity with depth. This site tested the ability of inversion algorithms to resolve velocity gradients. Teams using a small number of discrete layers in their models often oversimplified the profile, leading to significant differences in the calculated $V_{s,30}$ when compared to the borehole logs. The results emphasized the importance of "parameterization"—the way a geophysicist chooses to describe the subsurface layers before the computer begins the calculation.

The Lisbon Site (Portugal)

The Lisbon site presented the most complex geological stratigraphy, featuring weathered volcanic rocks (basalt) interbedded with limestone and calcarenite. This created velocity inversions, where a faster layer sits on top of a slower layer. Such structures are notoriously difficult for standard Rayleigh wave dispersion analysis because the wave energy can jump between modes, leading to the misinterpretation of data. The InterPACIFIC results for Lisbon showed the highest degree of scatter, highlighting the limitations of non-invasive methods in urban environments with complex lithology.

Statistical Analysis of Inversion Variability

The project quantified the uncertainty of $V_{s,30}$ using the coefficient of variation (COV). The analysis demonstrated that the COV for the experimental dispersion curves was generally below 5%, indicating that researchers are very good at measuring wave speeds. However, the COV for the resulting $V_{s,30}$ values jumped to 15-20% at the more complex sites. This increase is attributed to the "interpreter effect"—the subjective choices made by the analyst during the inversion process.

The data presented in the final reports suggests that the uncertainty is not merely a product of the software used but is deeply rooted in the lack ofA prioriInformation. When teams were given a small amount of geological context (such as the expected depth to a major lithological boundary), the scatter in their results decreased significantly. This finding has led to the recommendation that non-invasive seismic surveys should always be integrated with at least one local borehole or a well-constrained geological map.

The Soil-Bedrock Interface and Resolution Limits

One of the most critical aspects of the InterPACIFIC project was investigating the resolution of the soil-bedrock interface. In seismic engineering, the bedrock is often defined as the point where the shear-wave velocity exceeds 800 m/s. The project's findings indicated that surface wave methods are highly effective at characterizing the upper 15 to 20 meters, but their resolution decays exponentially with depth. For deep interfaces, the choice of inversion algorithm becomes the dominant factor in model accuracy. Global search methods were found to be more reliable at capturing the uncertainty of these deep boundaries, whereas deterministic methods often provided a false sense of precision by converging on a single, potentially incorrect, model.

Implications for Engineering Practice

The results of the InterPACIFIC blind test have had a profound impact on how geophysical data is interpreted for infrastructure projects. It is now widely accepted within the community that a single "best-fit" profile is insufficient for critical design. Instead, practitioners are encouraged to provide a range of possible profiles that all satisfy the observed data. This allows engineers to account for the uncertainty in $V_{s,30}$ when performing site-response analyses for bridges, tunnels, and high-rise foundations.

Furthermore, the project underscored the necessity of using both active and passive seismic sources. Active sources (like sledgehammers) provide high-resolution data for the near-surface, while passive sources (like ambient urban noise) provide the long-wavelength information needed to see deeper into the earth. The cooperation between these two data types was essential for achieving even a moderate level of accuracy at the Grenoble and Lisbon sites.

"The variability observed in the InterPACIFIC project highlights that the inversion of surface waves is as much an art of geological interpretation as it is a mathematical exercise. The robustness of a site model depends heavily on the analyst's ability to constrain the search space with physical reality."

The InterPACIFIC project provided a benchmark that continues to guide the development of inversion algorithms. By documenting the discrepancies between different approaches, it has spurred the creation of more sophisticated, multi-modal inversion techniques that simultaneously consider Rayleigh and Love waves, as well as the horizontal-to-vertical spectral ratio (HVSR). These advancements aim to reduce the interpreter effect and move the discipline toward a more standardized, reproducible methodology for characterizing the shallow subsurface.