Characterization of metamorphic rocks to evaluate waste material acid rock drainage potential is particularly challenging as commonly used laboratory methods can result in substantial underprediction of acid generation potential (AGP) and overprediction of acid neutralization potential (ANP); this combination can lead to considerable underprediction of overall material acid rock drainage potential. Legacy waste and ore sulfur assays from the Touro copper project in Galicia, Northwestern Spain frequently relied on methods insufficient to fully digest metamorphosed sulfides. Additionally, the presence of both graphite and manganese–iron carbonates in Touro added considerable risk of ANP overestimation. Environmental characterization tests conducted on newer ore and waste samples aim to properly address these risks.

Innovative machine learning algorithms were employed to correct over 60,000 erroneous sulfur data points. When Atalaya Mining, Cobre San Rafael began exploring the Touro project, it inherited multiple legacy assay datasets with noticeable inconsistencies in sulfur assay data. The cost of such re-analysis is typically very high when such errors are repeated over the scale of thousands of samples. The innovative approach relied on re-analysis of a small subset of samples, and training an algorithm to 1) recognize relationships between the corrected parameter and other assay parameters in the subset, and 2) estimate corrected values for the larger dataset. Environmental characterization methods employed towards new waste and ore samples included Leco sulfur for full digestion of metamorphosed sulfides and correct AGP estimation. For improved ANP estimation, the Modified Sobek method was employed to more accurately account for the buffering capacity of non-metal carbonates while ignoring graphite and metal carbonates (although aluminosilicate dissolution rates need to be evaluated in the context of sulfide oxidation rates).

Machine learning algorithms trained on a dataset with correct sulfur data were able to derive a relationship between other assay variables which enabled reproducing the sulfur concentrations with 93% accuracy. Predictive success is largely a function of 1) the number of samples, 2) the number of assay parameters, and 3) material/deposit geochemistry. Additionally, use of LECO and Modified Sobek to quantify AGP and ANP resulted in considerable improvements over legacy data.

The innovative supervised sulfur prediction, LECO digestion and Modified Sobek titration methods employed in this study indicate that erroneous data generated from use of improper laboratory tests does not necessarily need to be discarded. Rather, such methods offer a pathway to correction of erroneous data.

Water quality data is collected throughout the mine project life cycle. During early stages, water quality is collected for baselining purposes. During feasibility, permitting, and mine planning, water quality data collected from material characterization tests is combined with baseline data to develop models predicting water quality associated with future mining activities. Prediction outcomes support material and water management strategies. During operations, water quality data is collected to monitor mine facility seepage, water treatment performance, and mine boundary compliance. At closure, water quality is collected to assess closure strategy effectiveness and employ adaptive management strategies. Water quality data is ubiquitous through the project life cycle and yet remains highly underutilized.

The abundance of this data is ideally suited to using machine learning approaches to maximize its value. It is typically collected at regular intervals throughout the mine life cycle. Furthermore, compliance and operational requirements dictate that water quality be collected at sufficient spatial density. Finally, water quality data is multi-dimensional, commonly including measurements for 20 to 40 (or more) different parameters. The large number of parameters makes for a “wide” dataset with considerable statistical variance, which facilitates application of innovative unsupervised machine learning methods.

Application of unsupervised methods to water quality data in four different parts of the mine project life cycle indicates that the multivariate approach is highly effective in identifying different water quality domains, breakthrough of MIW, mixing effects and various reactive processes that occur in water.

The four use cases have application and implications as follows:

• Baseline and compliance: machine learning methods are applied to separate pre-mining water quality in different hydrostratigraphic units and to detect vertical transmissivity between units.
• Feasibility and permitting: interpretation of a large laboratory humidity cell dataset to understand the formation and timing of acid conditions; this information can be applied at the field scale to predict acidic conditions in mine waste facilities and assess regional groundwater patterns in mineralized zones.
• Operations: forensic study to detect mine process water influence in both surface and groundwater, and to separate MIW from natural acid rock drainage.
• Closure planning: coupled unsupervised machine learning and isotope study is used to identify natural acid rock and neutral drainage and to separate these chemical fingerprints from groundwater influenced by pit and waste rock generated acid rock drainage; this work is all in support of development of the long-term site closure strategy.

In 2015, Minera La Zanja S.R.L. (a subsidiary of Compañía de Minas Buenaventura) began the planning, design and subsequent execution (2018) of the pilot closure project based on improved soils (Tecnosoles) over an area of approximately 8 ha, which proposed to refocus the criteria for coverings based on the Peruvian Ministry of Energy and Mines' mine closure guide.

Nature does not design or segment layers of materials in the soil, as proposed in the Peruvian closure guidelines, on the contrary, it is the soil itself that brings together the properties of waterproofing, water retention and evapotranspiration, so this pilot focused on working the soil properties in order to test its effectiveness in a waste deposit within an area of the San Pedro Sur open pit, operated at the La Zanja mine, in the Cajamarca region, Peru.

To test the effectiveness of the improved soil cover, a monitoring system was designed and installed. This monitoring system consists of 1) near-surface boreholes, just below the soil cover and isolated from the waste rock by a layer of quartz gravel; 2) deep boreholes, the purpose of which is to collect any water that may be present at the base of the reservoir after circulation through the reservoir; 3) instrumented wells, which are wells with a series of sensors (for continuous measurement of temperature, conductivity, suction pressure, humidity, oxygen and CO2) and suction lysimeters at different depths within the reservoir; and 4) external piezometers upstream and downstream of the reservoir and at different depths to monitor the potential effect on the aquifer. This project started reporting water quality indicators since 2018.

The evolution of water quality indicators, infiltration, runoff and evapotranspiration, have been similar or better than those obtained by traditional systems of layers with low permeability materials such as clays, which are used in many cases of mine closure in Peru, but without the added cost of their acquisition (often outside the mining units), transport and disposal, becoming a living pilot that continues to yield results that can be used to focus other cases of closure based on sustainable solutions.