The present invention relates to the field of resource assessment or resource mapping. In particular, it relates to a method for use in assessing and/or mapping resources such as (poly)metallic minerals in sub-sea (or marine) locations.

It is known to perform resource mapping by taking one or more samples (e.g. box cores, drill cores, or grab samples, depending for example on the mineral) from a seabed then analysing the samples to determine which resources (e.g. minerals, metals) are present and in what quantities or densities (e.g. in kg/m ² over a seabed). Sample locations may be selected based on bathymetry and/or multi beam echo sound (MBES) back-scatter amplitude data, for example. Box core data, for example, may be supported by back-scatter amplitudes and/or video assessments. Kriging methods may be used to interpolate the analysis results.

The present invention relates to a method for use in evaluating or mapping a sub-sea metallic mineral resource or resources over a region comprising: a. providing sample data relating to at least one sample taken from the region; b. providing at least one set of measured data relating to the region; c. inputting the sample data and the at least one set of measured data into a machine-learning algorithm, the machine-learning algorithm comprising a model relating the measured data to a quantity, fraction and/or density of the mineral resource or resources; and d. running the machine-learning algorithm to train the model.

Thus, the present invention provides a method of training a model for a machine-learning algorithm using both sample data relating to a region, and (other) measured data relating to the region. Using such a combination of different data to develop and train a machine-learning algorithm can provide an improved tool for evaluating or mapping a sub-sea metallic mineral resource or resources over a region compared, for example, to prior art methods which are mainly only based on box core data or drill core data and do not involve machine learning.

The present invention relates to a method of providing or training a model in a machine-learning algorithm. Once provided or trained (as described above), such a model may subsequently be used to evaluate or map a sub-sea metallic mineral resource or resources over the region. Such subsequent use of the trained model is described in more detail below. The sub-sea metallic mineral resource or resources preferably comprises one or more polymetallic minerals.

The sub-sea metallic mineral resource or resources may comprise nodules, such as metallic or polymetallic nodules, as these may be of particular interest.

Alternatively or additionally, the sub-sea metallic mineral resource or resources may comprise sulphides such as polymetallic sulphides.

Alternatively or additionally, the sub-sea metallic mineral resource or resources may comprise cobalt, more preferably one or more cobalt-rich crusts.

The sample data preferably is or comprises data from one or more box cores, from one or more drill cores, and/or one or more grab samples.

The sample data preferably comprises measurements of the (e.g. quantity or density of the) sub-sea metallic mineral resource or resources, e.g. for a certain location(s) within the region.

The sample data type(s) used (e.g. box core, drill core and/or grab sample) may depend on the type of sub-sea metallic mineral resource to be evaluated or mapped.

In cases where the sub-sea metallic mineral resource is or comprises nodules, the sample data preferably is or comprises data from one or more box cores.

In cases where the sub-sea metallic mineral resource is or comprises sulphides, the sample data preferably is or comprises data from one or more drill cores and/or one or more grab samples.

In cases where the sub-sea metallic mineral resource is or comprises one or more cobalt-rich crusts, the sample data preferably is or comprises data from one or more grab samples and/or one or more other samples from the sea bed (e.g. obtained with a remotely-operated tool or device, for example by breaking off the sample(s) from the sea bed). Cobalt-rich crusts are typically found in thin layers (around 10-20 cm thick) on the seabed, typically on seamounts. As such, grab samples and/or other samples obtained with remotely-operated tool or device are a suitable kind of sample for this resource.

In some cases (e.g. where the sub-sea metallic mineral resource is or comprises nodules), the sample data may comprise data from between around 100 and 1000 box cores. Where box core data is used, the spacing between the box cores may be around 2-6 k , or 4 km, for example on a regular grid. However, in other cases, the spacing between box cores may be larger and/or varying.

In some cases (e.g. where the sub-sea metallic mineral resource is or comprises sulphides), the sample data may comprise data from between around 10 and 40 drill cores. Such drill cores may be taken from a mound(s) (potentially) containing the mineral of interest (e.g. sulphides).

In some cases (e.g. where the sub-sea metallic mineral resource is or comprises sulphides and/or one or more cobalt-rich crusts), the sample data may comprise data from a large number of grab samples. Such grab samples may be taken from a seabed at or near a mound(s) (potentially) containing the mineral of interest (e.g. sulphides). Grab samples may be picked on the seabed and do not require drilling. As such, they are relatively cheap to gather. In some cases, sample data from around the order of 100 grab samples or more may be used.

The at least one set of measured data may comprise geological, geochemical, geophysical and/or oceanographic data. Preferably two or more of these types of data are provided and used in the method. In general, the more (different kinds) of measured data are used, the more accurately the machine learning algorithm (or its model) can be trained to predict or estimate a mineral resource or resources over a region.

The measured data preferably comprises numerical measured data. For example, the measured data preferably comprises data other than photographic, image and/or video data. In other words, the present invention preferably does not require the use of photographic, image and/or video data.

The at least one set of measured data may comprise one or more of: bathymetry data, (seabed) slope data, chlorophyll data, clay fraction data, carbonate fraction data, carbon-compensation depth data, the great circle distance from the mineral resource to the region of interest, (Bouguer or free air) gravity data, gravity vertical derivative data, isostatic compensation gravity data, magnetic data, ocean current data (e.g. for North and/or South Atlantic central water,

Antarctic intermediate water), oxygen level, seabed oxygen data, seawater pH data, MBES data, (seabed) curvature data, back-scatter data, fault density data, tectonic stress field data (e.g. including tectonic stress field anisotropy and/or maximum and/or minimum stress), distance from mid-oceanic ridge, geology data (e.g. relating to the presence of mafic and/or ultramafic rocks in the seabed), geoid anomaly data, electromagnetic data, sedimentation data (e.g. sedimentation rates such as pelagic clays or Sahara dust flux) and/or distance to hydrothermal venting. As above, the more (different kinds) of measured data are used, the more accurately the machine-learning algorithm can predict or estimate a resource or resources over a region.

Bathymetry data, (seabed) slope data and/or back-scatter data may be provided from multi-beam echo sounder data, for example.

Preferably, the location (e.g. area or region) to which the at least one set of measured data corresponds, at least partially correspond(s) to (e.g. overlaps, encompasses or falls within) the location(s) or point(s) to which the sample data corresponds.

The location (e.g. area or region) to which the at least one set of measured data and/or the sample data corresponds preferably at least partially correspond(s) to (e.g. overlaps, encompasses or falls within) the region for which sub-sea metallic mineral resource or resources is (are) to be mapped or evaluated.

In cases where the mineral resource or resources comprise polymetallic nodules, for example, preferably the at least one set of measured data comprises one or more of: bathymetry data, seabed slope data, chlorophyll data, clay fraction data, carbonate fraction data, carbon-compensation depth data, the distance from the mineral resource(s) to the region of interest, (Bouguer or free air) gravity data, gravity vertical derivative data, isostatic compensation gravity data, magnetic data, ocean currents data, seabed oxygen data, and/or seawater pH data. More preferably, in such cases, the at least one set of measured data comprises at least chlorophyll data, carbon-compensation depth data and/or the distance from the mineral resource(s) to the region of interest.

In cases where the mineral resource or resources comprises polymetallic sulphides, for example, preferably the at least one set of measured data comprises one or more of: MBES data, bathymetry (slope) data, (seabed) curvature data, back-scatter data, magnetic data, (Bouguer or free air) gravity data, fault density data, tectonic stress field data (e.g. including tectonic stress field anisotropy, maximum and/or minimum stress, and/or a ratio between maximum and minimum stress), distance from mid-oceanic ridge, geology data (e.g. relating to the presence of mafic and/or ultramafic rocks in the seabed), geoid anomaly data, and/or electromagnetic data. Preferably, the at least one set of measured data comprises one or more of: MBES data, (high resolution) magnetic data, and/or electromagnetic data.

In cases where the mineral resource or resources comprises cobalt-rich crusts, for example, preferably the at least one set of measured data comprises one or more of: bathymetry data, (seabed) slope data, ocean current data (e.g. for North and/or South Atlantic central water, Antarctic intermediate water), oxygen level, seabed oxygen data, back-scatter data, sedimentation data (e.g. sedimentation rates such as pelagic clays or Sahara dust flux) and/or distance to hydrothermal venting.

A play model may be used when selecting which set(s) of measured data to use, e.g. for a particular mineral resource(s). The play model may comprise a model of the system that creates the mineral resource(s).

If the mineral resource or resources comprises polymetallic nodules, the play model may comprise a nodule play model, for example modelling metal sources (e.g. cation and/or anion), metal transport and/or nodule accumulation or (favourable) conditions for nodule growth.

If the mineral resource or resources comprises polymetallic sulphides, the play model may comprise a model modelling metal sources, fluid pathways and/or accumulation and/or preservation of polymetallic sulphides.

The model of the machine-learning algorithm described above may (thus) be trained to predict or estimate a mean of the quantity, fraction and/or density of the mineral resource or resources.

In some cases, the model of the machine-learning algorithm may be trained to predict or estimate a mean of the quantity, fraction and/or density of the mineral resource or resources by regression.

Alternatively or additionally, the model of the machine-learning algorithm may be trained to predict or estimate a mean of the quantity, fraction and/or density of the mineral resource or resources by classification, e.g. by (semi-quantitative) classification of areas of predicted or estimated e.g. high, intermediate and low resource density. In other words, the model of the machine-learning algorithm may be trained to predict or estimate a mean of the quantity, fraction and/or density of the mineral resource or resources by classifying areas according to their predicted or estimated resource density.

Preferably, the method further comprises providing further sample data relating to at least one further sample taken from the region, and inputting the further sample data into the machine-learning algorithm to test the model relating the measured data to the mineral resource or resources. This can allow the trained model, trained as described above, to be tested and evaluated.

The method may comprise using further sample data to further train the model to predict the variance (or uncertainty) of the predicted mean.

The sample data used in training the model, and the further sample data used in testing and/or further training (e.g. to predict the variance or uncertainty of) the model, may relate to one or more separate or distinct areas of the region.

Alternatively (or additionally), the sample data used in training the model, and the further sample data used in testing and/or further training the model, may relate to overlapping areas of the region. In some cases, the sample data used to train the model, and the sample data used to test and/or further train the model may be selected randomly from a set of sample data relating to the region. Typically, around 80% of the available sample data relating to the region may be used to train the model, and around 20% of the available sample data relating to the region may be used to test the model, for example.

The machine-learning algorithm may comprise one or more Python and/or SciKit-Learn Python modules, or other equivalent modules in e.g. R or Matlab.

The present invention also relates to a method of evaluating or mapping a resource or resources over a region (e.g. a different area of the region to that used for training and testing the model of the machine-learning algorithm as described above), the method comprising: a. providing at least one set of measured data relating to the region; b. inputting the at least one set of measured data into a machine-learning algorithm comprising a model, wherein the model has been trained according to method described above (with any of its optional or preferred features), to evaluate or map the resource or resources over the region.

The machine-learning algorithm may (then) be run to evaluate or map the resource or resources over the region.

As above, the mineral resource or resources preferably comprises polymetallic minerals. The mineral resource or resources may comprise nodules, such as metallic or polymetallic nodules, sulphides such as polymetallic sulphides, and/or cobalt-containing mineral(s) such as cobalt-rich crusts.

The at least one set of measured data may comprise any of the types described above in relation to training the model. Preferably, the at least one set of measured data comprises the same type(s) of data as that (those) used for training the model.

For example, the at least one set of measured data may comprise geological, geochemical, geophysical and/or oceanographic data. Preferably, two or more of these types of data are provided and used in the method. In general, the more (different kinds) of measured data are used, the more accurately the machine learning algorithm can predict or estimate a mineral resource or resources over a region.

The at least one set of measured data may comprise one or more of: bathymetry data, (seabed) slope data, chlorophyll data, clay fraction data, carbonate fraction data, carbon-compensation depth data, the great circle distance from the mineral resource to the region of interest, (Bouguer or free air) gravity data, gravity vertical derivative data, isostatic compensation gravity data, magnetic data, ocean current data (e.g. for North and/or South Atlantic central water, Antarctic intermediate water), oxygen level, seabed oxygen data, seawater pH data, MBES data, (seabed) curvature data, back-scatter data, fault density data, tectonic stress field data (e.g. including tectonic stress field anisotropy and/or maximum and/or minimum stress), distance from mid-oceanic ridge, geology data (e.g. relating to the presence of mafic and/or ultramafic rocks in the seabed), geoid anomaly data, electromagnetic data, sedimentation data (e.g. sedimentation rates such as pelagic clays or Sahara dust flux) and/or distance to hydrothermal venting. As above, the more (different kinds) of measured data are used, the more accurately the machine learning algorithm can predict or estimate a resource or resources over a region.

In cases where the mineral resource or resources comprise polymetallic nodules, for example, preferably the at least one set of measured data comprises one or more of: bathymetry data, seabed slope data, chlorophyll data, clay fraction data, carbonate fraction data, carbon-compensation depth data, the distance from the mineral resource(s) to the region of interest, (Bouguer or free air) gravity data, gravity vertical derivative data, isostatic compensation gravity data, magnetic data, ocean currents data, seabed oxygen data, and/or seawater pH data. More preferably, in such cases, the at least one set of measured data comprises at least chlorophyll data, carbon-compensation depth data and/or the distance from the mineral resource(s) to the region of interest.

In cases where the mineral resource or resources comprises polymetallic sulphides, for example, preferably the at least one set of measured data comprises one or more of: MBES data, bathymetry (slope) data, (seabed) curvature data, back-scatter data, magnetic data, (Bouguer or free air) gravity data, fault density data, tectonic stress field data (e.g. including tectonic stress field anisotropy, maximum and/or minimum stress, and/or a ratio between maximum and minimum stress), distance from mid-oceanic ridge, geology data (e.g. relating to the presence of mafic and/or ultramafic rocks in the seabed), geoid anomaly data, and/or electromagnetic data. Preferably, the at least one set of measured data comprises one or more of: MBES data, (high resolution) magnetic data, and/or electromagnetic data.

In cases where the mineral resource or resources comprises cobalt-rich crusts, for example, preferably the at least one set of measured data comprises one or more of: bathymetry data, (seabed) slope data, ocean current data (e.g. for North and/or South Atlantic central water, Antarctic intermediate water), oxygen level, seabed oxygen data, back-scatter data, sedimentation data (e.g. sedimentation rates such as pelagic clays or Sahara dust flux) and/or distance to hydrothermal venting.

The method may further comprise interpreting the output of the method.

Evaluating or mapping resources preferably comprises determining an amount, fraction (e.g. volume fraction of the metal or resource in an ore), or density (e.g. a spatial density measured, for example, in kg / m ² , or a volume density measured, for example, in kg / m ³ ) of the resources.

Preferably, the uncertainty of the output of the method (e.g. the estimated amount, fraction or density of the resources) is also assessed. The uncertainty may be assessed by running (training and/or testing) the machine-learning algorithm for a further time, e.g. to train and/or test a (second or further) model for estimating or determining the uncertainty of the output of the method (e.g. the estimated amount, fraction or density of the resources). Thus, the machine-learning algorithm may comprise a further model that may be trained and/or tested (e.g. by running the machine-learning algorithm for a further time) to estimate or determine the uncertainty of the output of the method (e.g. the estimated amount, fraction or density of the resources). The uncertainty may comprise the squared error of the prediction(s) from the first model.

Thus, aspects of the present invention provide a method of training (and optionally also testing) a machine-learning algorithm for mapping or evaluating a mineral resource or resources over a region. The method uses both sample data relating to at least one sample taken from the region and (different) measured data relating to the region to train the algorithm. Further sample data may then be used to test the machine-learning algorithm.

Once such a machine-learning algorithm has been developed, it may then be used to map or evaluate the resource or resources over (a preferably different but nearby area of) the region, for example by inputting (different/further) measured data relating to (a preferably different but nearby area of) the region into the machine-learning algorithm, and running the algorithm such that it outputs an estimate (e.g. a mean) of a quantity, fraction and/or density of the resource or resources for the (preferably different but nearby area of the) region and optionally also an estimate of the variance or error of the estimate of the quantity, fraction and/or density of the resource or resources for that (different but preferably nearby) region.

As such, one or more sets of measured data and/or samples corresponding to a first area (or first set of areas) of the region may be used to train and test the model of the machine-learning algorithm as described above. Then, a further set (or sets) of measured data relating to a (different) second area (or second set of areas) of the region may be used to map or evaluate the resource or resources over the second area (or second set of areas) of the region.

The second area (or second set of areas) are preferably relatively close to (e.g. in a same ocean or sea as) the first area or set of areas. This helps to ensure that the trained model would work for the second area or set of areas.

The second area or second set of areas may be or comprise an area or areas for which there is no (or limited) sample data, for example. As such, training a model for a region based on an area(s) of the region for which sample and measured data (as described above) is available may provide a model which may then be used to map or evaluate the resource(s) for other area(s) of the region, e.g. for which no sample data is available.

The method may comprise using standard machine learning techniques such as the known “random forest” and/or Bayes techniques.

The region over which the present invention is applied may be a relatively large region such as an area of around 6,000,000 km ² (around 20 times the size of Norway), for example. Thus, the present invention may be used to highlight an area(s) of interest, e.g. where resource density may be high, within such a region. Such an area(s) (of the region) may then be focussed on in further investigation to provide a more detailed evaluation of the resource(s) in that (those) area(s).

The method may thus further comprise, based on the evaluation or mapping of the resource or resources over the region, identifying an area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density (e.g. above a particular threshold). For example, in cases where the resource is or comprises polymetallic nodules, if an area of the region is estimated to have at least 8-15 kg / m ² of nodules, then it could be identified as an area with a relatively high resource amount or density. In cases where the resource is or comprises polymetallic sulphides, if an area of the region is estimated to have at least 2-4% (by weight) of copper and/or zinc, then it could be identified as an area with a relatively high resource amount or density.

The method may comprise displaying its output (e.g. the estimate (e.g. a mean) of the quantity, fraction and/or density of the resource or resources for that (different) region and optionally also an estimate of the variance or error of the estimate of the quantity, fraction and/or density of the resource or resources for that (different) region) graphically, e.g. in a map format. For example, areas with the highest estimates for the quantity, fraction and/or density of the resource or resources (e.g. above a first particular threshold) may be shown in one colour (e.g. green), areas with mid-range estimates (below the first particular threshold and above a second particular threshold which is lower than the first particular threshold) may be shown in a second colour (e.g. yellow), and areas with the lowest estimates (below the second particular threshold) may be shown in a third colour (e.g. red).

The method may then further comprise (further) investigating such an area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density (e.g. above a particular threshold). Such (further) investigating may comprise, for example obtaining (further) data or measurements from the area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density. The (further) data or measurements may comprise, for example, bathymetry data, MBES data, back-scatter amplitudes and/or photographic and/or video assessments, for example. A further (e.g. improved or more accurate) estimate of the resource amount or density for the area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density may then be obtained, e.g. from or based on the (further) data or measurements from the area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density.

Based on the initial evaluation or mapping of the resource or resources over the region, and/or on the further (e.g. improved or more accurate) estimate of the resource amount or density for the area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density (e.g. if either or both of these provide an estimate of the resource amount or density for the area or areas which is sufficiently high, e.g. above a certain threshold such as described above), a mining operation may be performed to mine for the resource or resources in (the) identified area or areas. The decision to perform a mining operation may of course be based on one or more further factors (e.g. further testing of predicted targets by core drilling, the presence or absence of any existing infrastructure in the area(s), the accessibility of the area(s)) besides just the initial evaluation or mapping of the resource or resources over the region and/or the further (e.g. improved or more accurate) estimate of the resource amount or density for the area or areas of the region which may have (or are estimated to have) a relatively high resource amount or density.

According, to a further aspect, there is provided a computer program product comprising computer-readable instructions that, when run on one or more processors or a computer, cause the one or more processors or computer to perform the method described herein, with any of its optional or preferred features.

Aspects of the present invention may provide a method of assessing and/or mapping resources over a relatively large scale. This can help to identify areas (e.g. prospects or targets) for which it might be useful to obtain further data, with which to make a better assessment and/or mapping of the resource(s).

In embodiments of the invention, a multi-geoscientific approach may applied to regional resource assessment. This may be useful as part of an access project, or to evaluate one or more farm in opportunities, for example.

The method may takes a multidisciplinary approach, for example utilizing geological, geochemical, geophysical and/or oceanographic datasets, most of which may be (relatively easily) available from publications and public academic databases, i.e. without the need to take (further) specific measurements. Such a multidisciplinary approach means that users may make quantitative use of more data, with better predictions, including uncertainty estimates. Preferred embodiments of the invention will now be described by way of example only and with reference to the accompanying drawing in which:

Fig. 1 is a flow chart illustrating a method for evaluating or mapping a sub sea metallic mineral resource or resources over a region.

As illustrated in Fig. 1, a method for evaluating or mapping a sub-sea metallic mineral resource or resources over a region comprises the following steps:

1a - providing sample data relating to at least one sample taken from an area or areas of the region;

1b - providing at least one set of measured data relating to the area or areas the region;

2 - inputting the sample data and the at least one set of measured data into a machine-learning algorithm, the machine-learning algorithm comprising a model relating the measured data to a quantity, fraction and/or density of the mineral resource or resources, and running the machine-learning algorithm to train the model to predict the quantity, fraction and/or density of the mineral resource in the region;

3 - inputting further sample data into the machine-learning algorithm and running the machine-learning algorithm to test the model and/or to train the model to predict the variance or uncertainty of the quantity, fraction and/or density of the mineral resource;

4 - inputting measured data relating to a different area or areas of the region into the trained machine-learning algorithm and running the machine learning algorithm to predict the quantity, fraction and/or density of the mineral resource and its variance or uncertainty for the different area or areas of the region;

5 - displaying the result of step 4 graphically;

6 - based on the predicted quantity, fraction and/or density of the mineral resource and its variance or uncertainty (e.g. if the predicted quantity, fraction and/or density of the mineral resource is above a particular threshold for a particular area or areas), obtaining further data for the area or areas to obtain an improved estimate of the quantity, fraction and/or density of the mineral resource

7 - based on the improved estimate of the quantity, fraction and/or density of the mineral resource, making a decision to mine for the mineral resource.

In some embodiments, the mineral resource consists of polymetallic nodules. In other embodiments, the mineral resource consists of polymetallic sulphides. In other embodiments, the mineral resource consists of cobalt-rich crusts.

The above steps of evaluating or mapping these resources over a region will now be described in more detail.

At step 1a sample data relating to at least one sample taken from the region is provided. The sample data comes from a plurality of box cores.

At step 1b one or more sets of measured data relating to the region are provided.

In cases where the resource being evaluated or mapped consists of polymetallic nodules, the one or more sets of measured data comprise one or more of: bathymetry data, seabed slope data, chlorophyll data, clay fraction data, carbonate fraction data, carbon-compensation depth data, the distance from the mineral resource(s) to the region of interest, Bouguer or free air gravity data, gravity vertical derivative data, isostatic compensation gravity data, magnetic data, ocean currents data, seabed oxygen data, and/or seawater pH data. In particular, it is preferred that the measured data comprises at least chlorophyll data, carbon- compensation depth data and/or the distance from the mineral resource(s) to the region of interest.

In cases where the resource being evaluated or mapped consists of polymetallic sulphides, the one or more sets of measured data comprises one or more of: MBES data, bathymetry (slope) data, (seabed) curvature data, back- scatter data, magnetic data, Bouguer or free air gravity data, fault density data, tectonic stress field data (e.g. including tectonic stress field anisotropy, maximum and/or minimum stress, and/or a ratio between maximum and minimum stress), distance from mid-oceanic ridge, geology data (e.g. relating to the presence of mafic and/or ultramafic rocks in the seabed), geoid anomaly data and/or electromagnetic data. In particular, it is preferred that the measured data comprises at least MBES data, (high-resolution) magnetic data, and/or electromagnetic data.

In cases where the resource being evaluated or mapped consists of cobalt- rich crusts, the one or more sets of measured data comprises one or more of: bathymetry data, (seabed) slope data, ocean current data (e.g. for North and/or South Atlantic central water, Antarctic intermediate water), oxygen level, seabed oxygen data, back-scatter data, sedimentation data (e.g. sedimentation rates such as pelagic clays or Sahara dust flux) and/or distance to hydrothermal venting. At step 2, the sample data and the at least one set of measured data are input into a machine-learning algorithm. The machine-learning algorithm comprises a model relating the measured data to a quantity, fraction and/or density of the resource being evaluated or mapped. The machine-learning algorithm is then run to train the model to predict the quantity, fraction and/or density of the resource.

In cases where the resource consists of polymetallic nodules, a nodule play model which models metal sources (e.g. cation and/or anion), metal transport and/or nodule accumulation or (favourable) conditions for nodule growth may be used to select which (types of) measured data to use.

In cases where the resource consists of polymetallic sulphides, a play model which models metal sources, fluid pathways and/or accumulation and/or preservation of polymetallic sulphides may be used to select which (types of) measured data to use.

The machine-learning algorithm may comprise one or more Python and/or SciKit-Learn Python modules, or equivalent or similar modules implemented in e.g. R or Matlab.

By performing step 2, the model of the machine-learning algorithm is trained to predict or estimate a mean x = E[x] of the quantity, fraction and/or density of the resource.

At step 3, further sample data is input into the machine-learning algorithm and the machine-learning algorithm is run again to test the model and to train the model (or a further model) to predict the variance 8 = E[(x - x)] or uncertainty of the quantity, fraction and/or density of the resource. The uncertainty may alternatively be expressed as P90, P50, P10 estimates.

The further sample data used at step 3 comes from a different area(s) of the region to that referred to in steps 1a and 2.

Once the model has been trained according to steps 1a, 1b, 2 and 3, the machine-learning algorithm can be used, at step 4, to predict the quantity, fraction and/or density of the resource and its variance or uncertainty for a different area of the region (e.g. an area for which there is no sample data) to that (those) area(s) of the region to which the sample and measured data used to train and test the model related.

As such, as step 4, measured data relating to a different area of the region (compared with the area(s) of the region to which the sample and measured data used to train and test the model related) is input into the trained machine-learning algorithm and the trained machine-learning algorithm is run to predict the quantity, fraction and/or density of the resource and its variance or uncertainty for the different area of the region.

The type(s) of measured data used at step 4 are the same as those obtained at step 1b and used to train the model. Alternatively, in some cases, a subset of the type(s) of measured data obtained at step 1b and used to train the model are used at step 4.

At step 5, the result of step 4, i.e. at least the predicted quantity, fraction and/or density of the resource is displayed graphically, for example as a map. Different colours (e.g. red, yellow, and green) are used to indicate areas of low, medium and high resource quantity, fraction and/or density.

If any areas of (sufficiently) high resource quantity, fraction and/or density are predicted at steps 4 and 5, at step 6 a decision may be made to obtain further data for that(those) area(s) in order to obtain an improved estimate of the quantity, fraction and/or density of the resource in that(those) area(s). Such further data may then be obtained. The further data can include any or all of bathymetry data, MBES data, back-scatter amplitudes and/or video assessments, for example.

If the improved estimate of the quantity, fraction and/or density of the resource in that(those) area(s) obtained at step 6 is (still) sufficiently high, the resource may be mined in that(those) area(s). The decision to perform a mining operation may of course also depend on further factors such as the presence or absence of any existing infrastructure in the area(s) and the accessibility of the area(s).