### Riskope Blog

- LATEST BLOG POST
- A decade of physical risks generated by industrial systems hacking is featuring a remarkable and worrisome acceleration. Indeed, it all…
- Read More

Recent tailings disasters such as the failure of the Mount Polley and Samarco dams have focused the attention on the Consequences of tailings dams failures.

Various classic approaches (Rico et al., 2008, Azam & Li, 2010) assume that the volume of released tailings is a function of storage volume. Dam height intervenes, particularly for the runout distance preliminary evaluation. They find “statistical correlations” using a set of dam break cases. The set is non-homogeneous, mixes various types of failure modes, disregards many potentially driving parameters. A small number of well documented dam break cases is the reason for this. Nevertheless many engineers apply those correlations in a deterministic way, using “one magic number” for the outflow volume.

Azam and Li stated that 20% of the total volume is a good approximation. Rico et al. equation delivers roughly one third of the storage volume. However, in many cases the outflow volume has been significantly larger than 20-30%. Indeed, none of the above accounts for liquefaction and Rico et al. rightly pointed out that some parameters contributing to the uncertainty in the predictions include:

- sediment load,
- fluid behaviour (depending on the type of failure),
- topography,
- the presence of obstacles stopping the flow, and
- the proportion of water stored in the tailings dam (linked to meteorological events or not).

More recently, in 2018, a paper by Paulina Concha Larrauri and Upmanu Lall (Columbia University ) and related app attempt to introduce probabilistic estimates (in terms of exceedance probabilities) using an updated set of data and “statistical techniques” and similar equations than prior authors. Unfortunately neither the outflow volumes nor the travel distances estimated using the recent study by Columbia University should be blindly applied. Indeed the database uses again:

- a mix all sorts of different failure types (even worse than in the original Rico’s paper), in different materials and
- excludes cases were the outflow was very small, biasing the results.

But perhaps the biggest bias is introduced by taking into account the Samarco failure’s travelled distance at 637km. In the Samarco case two distinct phenomena occurred:

- Tailings outflow
- River transportation

Reportedly in the Samarco case more than 90% of the tailings stopped within 120 Km of the dam, while the rest travelled in the Doce river to the Atlantic Ocean. The proper way to deal with a risk assessment in this case is not to include this data in the “statistical dataset”, but to model the outflow. Then one should consider how much of the outflow would end-up in the river, and finally model the river transportation.

In a paper from Rourke, H., & Luppnow, D. (2015) entitled The Risks of Excess Water on Tailings Facilities and Its Application to Dam-break Studies the authors consider a different approach. Indeed after the review of several documented tailings failures, they found necessary to consider the extent and depth of the supernatant pond. They state this it is just as important, if not more so, than the storage volume and dam height as driving parameters.

Their set of data is five catastrophic failures which “magically” line up in the graph with an astounding high correlation coefficient.

In the figure we can see the incredibly simple relationship, which would be amazing if it was true. Indeed it would allow for fast and cheap screening level analysis of the outcome of a dam break. But is it real or a fluke? You know what they say: when something looks too good to be true…

Let’s start by noting that Mount Polley, Kolontar and Stava all had more than 70% of ponded surface/total pond surface ratio. Tailings facilities with excess water storage have a higher probability of failure as they are more susceptible to overtopping, piping and liquefaction failures. They also pose greater failure consequences as the saturated conditions will lead to more material being mobilized. Indeed, their outflow volumes were all above 50% of the impounded volume.

It is generally accepted that the collapse of the Fundão tailings dam of November 5th 2015 released between 32Mm3 (Larrauri, Lall, Table 1, case 15) and 43Mm3 of tailings. That represent 60% to 80% of the total impounded volume. We note here how even in a actively studied and documented case significant uncertainties remain present.

The 80% value is really interesting as, with the proposed equation, it would only remain true for an impossible ratio of pool to impoundment surface (red arrow in the figure, would require to more than 100% ratio). We also note here that the complex interaction of the slimes and sands in the Fundao Dam impoundments makes the “ponded surface/total pond surface” ratio meaningless.

Interestingly Rourke and Luppnov themselves stated that the relationship shows a good correlation for the small number of cases. However more data is needed to increase confidence. Adding more case studies to the analysis may indicate that the relationship is steeper (more dependent on pond volume), shallower or even non-linear. It would also be worthwhile incorporating other parameters, such as embankment height, into the analysis.” We could not agree more with Rourke and Luppnov.

Conclusion: the model may have some merits, but its applicability should be carefully studied and stated. It is indeed too good to be true.

The failure of a dam is a serious event with the potential to disrupt people’s lives, the environment, infrastructures and human organizations. It should be studied in detail in order to ensure geo-ethical decisions are taken.

Although we are strong promoters of portfolio screening tools like ORE2_Tailings, we believe the proper way to perform dams portfolio analyses is the following:

- Use ORE2_Tailings to evaluate the probabilities of failure of all the dams in the portfolio.
- Decide for a land base (classic) or Space Based (satellites) monitoring program.
- Determine for the high probabilities dams within the portfolio the consequences of a dam break.

This procedure will highlight that the dams with the highest probability of failure within the portfolio are not necessarily the largest. In our experience, the “horrors” are oftentimes smaller dams, featuring poor design, building techniques and maintenance.

Then it will be time to start thinking at potential consequences of a break. In some papers we have seen authors using the adjective “imponderable” for dams break consequences. We do not believe “imponderable” is an appropriate adjective. In many instances we should replace it with “non evaluated” or “disregarded”. And that constitutes another ethical offense.

All that to say that the runout distance should only be evaluated using dam break analyses after careful probabilistic characterization of the tailings’ rheological parameters. The probabilistic characterization is necessary because the tailings rheology can change dramatically. Drivers of the change are, for example:

- variations in the degree of saturation,
- consolidation, and the
- increased presence of clay minerals in some layers.

“Wet” failure modes obviously drive up outflow volumes, while “dry” failures drive it down. For example, overtopping in a rainy day scenario is definitely a “wet” scenario.

The runout distance will also react to the conditions downstream:

- Rio Doce in the case of Samarco,
- steep narrow valley in the case of Stava,
- lake downstream for Mount Polley.

For balanced risk analyses when a dams portfolio has to be prioritized for sensible mitigation it is paramount to:

- Select appropriate parameters from the dam break outflows analyses and to
- definitive of the relative causality of various failure modes are paramount.

Dam break analyses are a scenario where there is little room for oversimplified approaches based on insignificant statistical samples.

Tagged with: Consequences, tailings dams failures

Category: Consequences, Hazard, Optimum Risk Estimates, Probabilities, Risk analysis, Risk management

Thanks for a nice blog. You have a comment on the inclusion of the Samarco data in the Larrauri and Lall paper. All the regressions reported in that paper were also done without the Samarco data, and using cross validation. You are right that there are many different types of dams, failures and environments. Using all of them will give a lower R2 than using just a sub-type. This is reflected in the analysis in that paper. Further, in the real world you have many types of dams and settings, so an honest estimate of uncertainty requires cross validation, analysis of leverage (undue effect of an exceptional situation) and an adequate sample size. This is reflected in that paper. Of course all models have caveats, but statistical models need to reflect the uncertainty and that uncertainty needs to be presented, as we did in that paper.

Incidentally, most topographic features have a very good relationship between surface area and volume, especially engineered structures, so the area vs volume figure you present is not a surprise. Indeed using remote sensing from satellites, one could monitor tailing dam volume and relative volume filled could effectively, and that would provide a basis for a rapid risk assessment using the model we presented. Of course 4 observations to characterize a relationship would worry me normally, and in this example one could indeed rely on remote sensing to get much more at site estimates.