Advanced Search

Indexed by SCI、CA、РЖ、PA、CSA、ZR、etc .

Volume 32 Issue 2
Apr.  2021
Turn off MathJax
Article Contents

Donald A. Singer. How Deep Learning Networks could be Designed to Locate Mineral Deposits. Journal of Earth Science, 2021, 32(2): 288-292. doi: 10.1007/s12583-020-1399-2
Citation: Donald A. Singer. How Deep Learning Networks could be Designed to Locate Mineral Deposits. Journal of Earth Science, 2021, 32(2): 288-292. doi: 10.1007/s12583-020-1399-2

How Deep Learning Networks could be Designed to Locate Mineral Deposits

doi: 10.1007/s12583-020-1399-2
More Information
  • Whether using a shallow neural network with one hidden layer, or a deep network with many hidden layers, the training data must represent subgroups of the deposit type being explored to be useful. Published examples of neural networks have mostly been limited to one individual mineral deposit for training. Variation of geologic features among deposits within a type are so large that a single deposit cannot provide proper information to train a neural net to generalize and guide exploration for other deposits. Models trained with only one deposit tend to be academic successes but are not of practical value in exploration for other deposits. This is why it takes much experience examining many deposits to properly train an economic geologist—a neural network is not any different. Two examples of shallow neural networks are used to demonstrate the power of neural networks to possibly locate undiscovered deposits and to provide some suggestions of how to deal with missing data. The training data needs to include information spatially related to known deposits and hopefully information from many different deposits of the type. Lessons learned from these and other examples point to a proposed sampling plan for data that could lead to a generalized neural network for exploration. In this plan, 10 or more well-explored gold-rich porphyry copper deposits from around the world with 100 or more sample sites near and some distance from each deposit would probably capture important variability among such deposits and provide proper data to train and test a shallow neural network to predict locations of undiscovered deposits.
  • 加载中
  • Abedi, M., Norouzi, G. H., 2012. Integration of Various Geophysical Data with Geological and Geochemical Data to Determine Additional Drilling for Copper Exploration. Journal of Applied Geophysics, 83: 35-45. https://doi.org/10.1016/j.jappgeo.2012.05.003 doi:  10.1016/j.jappgeo.2012.05.003
    Cameron, E. M., Hamilton, S. M., Leybourne, M. I., et al., 2004. Finding Deeply Buried Deposits Using Geochemistry. Geochemistry: Exploration, Environment, Analysis, 4(1): 7-32. https://doi.org/10.1144/1467-7873/03-019 doi:  10.1144/1467-7873/03-019
    Cooke, D. R., Wilkinson, J. J., Baker, M., et al., 2015. Using Mineral Chemistry to Detect the Location of Concealed Porphyry Deposits-An Example from Resolution, Arizona. Proceedings of the 27th International Applied Geochemistry Symposium 2015, April 20-24, 2015, Arizona, USA. 1-6
    Cox, D. P., Singer, D. A., 1986. Mineral Deposit Models. U.S. Geological Survey Bulletin, 1693: 379
    Hinton, G. E., Osindero, S., Teh, Y. W., 2006. A Fast Learning Algorithm for Deep Belief Nets. Neural Computation, 18(7): 1527-1554. https://doi.org/10.1162/neco.2006.18.7.1527 doi:  10.1162/neco.2006.18.7.1527
    Masters, T., 1993. Practical Neural Network Recipes in C++. Academic Press, Inc., San Diego, California. 493
    Masters, T., 2013. Assessing and Improving Prediction and Classification. CreateSpace. 560
    Masters, T., 2016. Deep Belief Nets in C++ and CUDA C, Volume Ⅲ: Convolutional Nets. CreateSpace. 207
    Rumelhart, D., McClelland, J., the PDP Research Group, 1986. Parallel Distributed Processing. MIT Press, Cambridge
    Sillitoe, R. H., 2010. Porphyry Copper Systems. Economic Geology, 105(1): 3-41. https://doi.org/10.2113/gsecongeo.105.1.3 doi:  10.2113/gsecongeo.105.1.3
    Singer, D. A., Kouda, R., 1996. Application of a Feedforward Neural Network in the Search for Kuroko Deposits in the Hokuroku District, Japan. Mathematical Geology, 28(8): 1017-1023. https://doi.org/10.1007/bf02068587 doi:  10.1007/bf02068587
    Singer, D. A., Berger, V. I., Moring, B. C., 2008. Porphyry Copper Deposits of the World: Database, Map, and Grade and Tonnage Models, 2008. U.S. Geological Survey Open-File Report 2008-1155. [2020-12-2]. http://pubs.usgs.gov/of/2008/1155/
  • 加载中
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Figures(3)  / Tables(1)

Article Metrics

Article views(4) PDF downloads(4) Cited by()

Related
Proportional views

How Deep Learning Networks could be Designed to Locate Mineral Deposits

doi: 10.1007/s12583-020-1399-2

Abstract: Whether using a shallow neural network with one hidden layer, or a deep network with many hidden layers, the training data must represent subgroups of the deposit type being explored to be useful. Published examples of neural networks have mostly been limited to one individual mineral deposit for training. Variation of geologic features among deposits within a type are so large that a single deposit cannot provide proper information to train a neural net to generalize and guide exploration for other deposits. Models trained with only one deposit tend to be academic successes but are not of practical value in exploration for other deposits. This is why it takes much experience examining many deposits to properly train an economic geologist—a neural network is not any different. Two examples of shallow neural networks are used to demonstrate the power of neural networks to possibly locate undiscovered deposits and to provide some suggestions of how to deal with missing data. The training data needs to include information spatially related to known deposits and hopefully information from many different deposits of the type. Lessons learned from these and other examples point to a proposed sampling plan for data that could lead to a generalized neural network for exploration. In this plan, 10 or more well-explored gold-rich porphyry copper deposits from around the world with 100 or more sample sites near and some distance from each deposit would probably capture important variability among such deposits and provide proper data to train and test a shallow neural network to predict locations of undiscovered deposits.

Donald A. Singer. How Deep Learning Networks could be Designed to Locate Mineral Deposits. Journal of Earth Science, 2021, 32(2): 288-292. doi: 10.1007/s12583-020-1399-2
Citation: Donald A. Singer. How Deep Learning Networks could be Designed to Locate Mineral Deposits. Journal of Earth Science, 2021, 32(2): 288-292. doi: 10.1007/s12583-020-1399-2
  • Successful economic geologists commonly require many years of experience examining many different deposits. This is due to the wide variability of features of mineral deposits even of the same type. Having a tool such as a neural network that could complement the economic geologist would be welcome. A fundamental problem is to determine how to properly train a neural network for this task so that it could learn the patterns like experienced economic geologists.

    Zoning of geochemistry, geophysics, alteration and associated mineral deposits have been well documented for individual deposits and in generalizations in mineral deposit models (Cox and Singer, 1986). It is often difficult to use this information in exploration due to the fact that the spatial information typically is specific for individual deposits and even deposits of a specific type vary considerably in details of geologic ages, rock types and local information available. Local host rocks often affect alteration patterns, spatially associated deposits, and geo chemical zoning. An additional problem is that much exploration interest is for partially or completely concealed mineral deposits. In practical terms there are two key problems, (1) it is not clear how to integrate diverse complex information when some of the information is missing, and (2) typically there are too few well-studied deposits in the region of interest.

    Here some possible answers to these two problems are presented. First, an overview is presented of recent advances in deep learning and neural networks that have demonstrated remarkable abilities to recognize patterns with large quantities of data. Following this, two simple examples of neural networks are used to illustrate some of the possibilities. Next, some ways are presented in which appropriate data could be collected and properly prepared for learning to address the key problems.

  • Geologists and prospectors have long recognized patterns that helped them discover mineral deposits. Successful geologists typically had extensive experience by seeing many deposits which helped them recognize many of the patterns. Such experience has been difficult to obtain even over many years. Modern exploration frequently must now deal with concealed deposits for which patterns recognized in exposed deposits might not be observable. Where the patterns have been observed, statistical tools such as discriminate analysis, logistic regression, and Bayesian methods have extended the ability of geologists in their searches.

    In 1986 a publication by Rumelhart and others opened up the new field of neural networks. The most widely used was called a multiple-layer feedforward network. Simply put, a neural network is a set of input information that is passed into an internal processing system which produces a set out output values. The most recognized form of network can be shown as a set of inputs (the input layer), a hidden layer with numerous neurons (processing units) and an output layer with one or more output neurons (Fig. 1). Neurons in every layer feed into every neuron in next layer above. Networks learn weights to the connections between neurons that are adjusted to meet a measure of error for the output. The shallow network shown on the left of Fig. 1 has been found to be quite powerful at integrating information for classification, noise reduction, and prediction (Masters, 1993).

    Figure 1.  Subset of connections in a shallow network and deep network. Neurons in every layer feed into every neuron in next layer above (after Masters, 2016).

    The successes of and the difficulty of training networks with more than two hidden layers led to widespread use of shallow neural networks. As models of the power of the human brain, they only failed with complex images and patterns that humans excelled at. They also failed to mimic the human brain in the computational structure of deep layers. A major breakthrough was made in 2006 when Geoffrey Hinton and others published a paper showing how multiple hidden layers with few hidden neurons per layer (Fig. 1) could be solved. This seminal paper provided the basis for what is now called deep networks. The power of these networks led to breakthroughs in image and speech recognition and are now widely used in many disciplines (Masters, 2016). Two examples of the use of shallow single hidden layer networks are provided in the next section.

  • Examples of the use of shallow neural networks are used here to demonstrate the power of shallow neural networks and to provide some indications of what needs to be considered in successful applications of any neural network for exploration. The first example is from an attempt to test the power of neural networks to locate deposits whereas the second example represents a test of what might be possible with the correct kinds of information.

    An early study of neural networks in exploration used a multiple-layer feedforward network to predict possible locations of undiscovered mineral deposits (Singer and Kouda, 1996). The goal was to locate possible kuroko massive sulfide deposits using the data from 20 years of drilling in the largely concealed Hokuroku District of Japan (Fig. 2). The data used were from X-ray analyses of gypsum, sericite and pyrite in 152 drill holes in the approximately 40×40 km2 district. Average amounts per hole of pyrite, sericite, and gypsum plus anhydrite as measured by X-rays in 69 drill holes were used to train the net. Mainly drill holes near and between the Fukazawa, Furutobe, and Shakanai mines were used to train the network with the output neuron being the logarithm of distance to a deposit. Five neurons were used in the hidden layer. The training data were selected carefully to represent well-explored areas where confidence of the distance to ore was assured. The validation consisted of testing the trained network on the total available dataset, including the 69 drill holes used for training. The network succeeded in identifying all known deposits and pointed to an area in the northeast part of the District where an independently discovered a kuroko deposit had been found (Fig. 2). Such drill hole data is not commonly available in exploration for mineral deposits in concealed settings, so the nature of the data limited the network's applicability elsewhere. In addition, the training of the network was mostly based on one deposit which limits their application for other deposits even of the same type.

    Figure 2.  Neural network output in the Hokuroku district, Japan. Based on X-ray data of gypsum, sericite, and pyrite in drill holes. One hidden layer, 5 neurons (from Singer and Kouda, 1996).

    Commonly in exploration for undiscovered deposits, the information that might be useful is widely scattered and often some samples contain missing data for some of the variables. To test one way to deal with missing information, an artificial example of 800 points was generated with three kinds of alteration and a geophysical variable. The alteration was assumed to be zoned around the deposit but with some irregularities. The alteration variables were coded as 1 if present in the sample, -1 if not present, and 0 if no information was available on alteration such as when the sample was located where possible altered rocks were covered by younger sediments. The geophysical variable was available in all samples and had higher values in an annulus around the deposit. All simulated data were located in X-Y space in kilometers from the center of the deposit. A probabilistic neural network (Masters, 1993) was trained on a randomly selected subset of 360 points into one of four distance classes of 0–1 km, 1–3, 3–5 and 5–10+ km (Table 1). The rows in Table 1 that have 0 for all three alteration variables represent samples located where rocks that could have been altered were covered—that is no alteration could have been observed even if it existed.

    PNN INPUT PNN OUTPUT PROB. PNN PNN
    X Y Dist. Group Alt1 Alt2 Alt3 Geoph. Gp1 Gp2 Gp3 Gp4 Group Distance
    0.88 0.32 0.94 1 1 -1 -1 0.305 1.00 0.00 0.00 0.00 1 0.5
    2.45 1.63 2.95 2 -1 1 -1 0.689 0.00 1.00 0.00 0.00 2 2
    2.18 -0.96 2.38 2 0 0 0 0.727 0.21 0.35 0.08 0.35 4 8
    -0.94 3.06 3.20 3 -1 -1 1 1.129 0.00 0.00 1.00 0.00 3 4
    2.38 2.85 3.72 3 0 0 0 1.495 0.00 0.54 0.41 0.05 2 2
    -1.68 4.63 4.93 3 -1 -1 1 1.402 0.00 0.00 1.00 0.00 3 4
    4.94 0.72 4.99 3 0 0 0 1.902 0.00 0.14 0.85 0.00 3 4
    6.82 1.55 6.99 4 0 0 0 0.865 0.10 0.47 0.13 0.30 2 2
    2.82 4.42 5.25 4 -1 -1 -1 0.675 0.00 0.00 0.00 1.00 4 8

    Table 1.  Example input and output of three alteration and one geophysical variables. X and Y are distances of sample in km from deposit, Group is distance group, PNN INPUT is the input alteration and geophysical variables. PNN OUTPUT are the probabilistic neural network probabilities and the median distances

    Figure 3 shows roughly contoured distances of the output of the probabilistic neural network using both training and test samples. The plot demonstrates that samples with some missing variables can be coded so that they can be used to point to the location of a mineral deposit. With only 360 training points this shallow neural network was powerful enough to learn reasonable estimates of distance to a deposit.

    Figure 3.  Contours of maximum estimated probability distance classes to center of deposit from 800 artificial samples of geophysics and alteration around a porphyry copper deposit.

    These two examples show both the power of even shallow neural networks and point to some significant issues that need to be addressed in order for these networks to successfully generalize for discovering new mineral deposits.

  • The wide variability of mineral deposits even of the same type is unlikely to be recognized and learned by a machine any more than it is by an economic geologist studying only one or two deposits. For one example, gold-rich porphyry copper deposits tend to be spatially associated with certain kinds of gold deposits which are not found near molybdenum-rich porphyry copper deposits. Yet published studies of neural networks applications on porphyry copper deposits have typically been based only on one deposit.

    So, the most important key to generalizing a neural network system to guide exploration for any deposit type would be to train it with a variety of deposits from different areas of the type of interest. A neural network for exploring for porphyry copper deposits can serve as an example of the how training and testing data could be gathered so that the resulting model would be generalized.

    A suggested data set of at least 10 porphyry copper deposits from around the world with perhaps 100 or more sample sites near and some distance away from each deposit would probably capture most of the variability among such deposits. Because gold-rich porphyry copper deposits may have different geophysical and other features than molybdenum-rich deposits, the tests should probably be done on the more common gold- to intermediate- porphyry copper deposits in order to reduce the learning complexities. The samples would represent information already gathered and recorded as distance from the center of the deposit. Clearly the measures need to be of the sort commonly recorded in public reports. Previous studies demonstrate that there are numerous variables capture the geochemical zoning patterns such as around these deposits for example (Cooke and others, 2015; Cameron and others, 2004). As suggested in the above example, potassic, sericite, advance argillic alteration have spatial patterns around porphyry copper deposits with the pyrite-bearing alteration typically widespread (Sillitoe, 2010; Singer and others, 2008). Magnetic anomalies vary by subtype of porphyry copper deposit and could be useful for covered areas (Abedi and Norouzi, 2012).

    An important variable that seems to not have been used is the spatial patterns of associated kinds of mineral deposits and prospects. For example, in settings where carbonate rocks are present, Cu-Au skarns are typically within 2 km of porphyry copper deposits and replacement bodies are more distant at 2–7 km such as near Bingham, United States where sediment-hosted and placer gold were also present. Near gold-rich porphyry copper deposits high-sulfidation epithermal gold deposits can occur within about 2 km and low-sulfidation epithermal gold deposits more distant such as near Far Southeast, Philippines. Where carbonate rocks are not present, polymetallic veins are often peripheral to porphyry copper deposits such as near Dexing, China (Sillitoe, 2010; Singer and others, 2008).

    There is widespread evidence that, at least for porphyry copper deposits, many variables are known to have strong spatial patterns that would be helpful for use in neural networks. Spatial information gathered from at least 10 porphyry copper deposits should provide a powerful set of data to train and test a neural network. It is possible that a shallow network would be all that is needed for satisfactory predictions. But if more power is needed, a deep network would certainly integrate this information into a successful predictor of the possible locations of undiscovered porphyry copper deposits even in concealed settings.

  • Deep learning neural networks have shown their power in their ability to deal with complex systems to make predictions. Even shallow networks have a history of successful integration of diverse information for classification and predictions. Attempts to use these methods for mineral exploration have shown their power but have not been shown to be able to generalize to predict locations of undiscovered deposits primarily due to training being limited to one deposit only. The considerable variability of geologic patterns around deposits of the same deposit type is the reason it takes economic geologists years of study to be competent and why neural networks need to be trained differently than they have been in the past. For neural networks to be successful mineral exploration tools they must learn from the same breadth of deposits that train successful economic geologists.

    There are two key problems, (1) it is not clear how to integrate diverse complex information when some of the information is missing, and (2) typically there are too few well-studied deposits of the type of interest in the region of interest. For presence-absence variables such as alteration types or spatially associated deposits, a simple coding of 1 for present, -1 for absent, and 0 for not measured seems to work in neural networks. For continuous variables the coding for missing observation needs to provide a neutral response for missing data.

    For the second issue of too few well-studied deposits available to represent the variety of geologic attributes that exist within a deposit type, it is necessary to train the network with many deposits from around the world. The network must learn from the same breadth of deposits that train successful economic geologists. A proposal is made here to gather a data set of 10 or more porphyry copper deposits from around the world with perhaps 100 or more sample sites near and away from each deposit would probably capture most of the variability among such deposits and provide the proper data to train a shallow neural network. The samples would represent information already gathered and recorded as distance from the center of the deposit. Clearly, the measures need to be of the sort commonly recorded in public reports. The proposed spatial information gathered from at least 10 gold-rich porphyry copper deposits should provide a powerful set of data to train and independently test a neural network. Based on the examples shown here, powerful deep neural networks that typically use thousands of data points to train to recognize patterns in images are probably not needed to the solve these exploration problems. Shallow neural networks with a relatively low number of training points and proper data preparation are likely to have enough predictive power for these problems.

Reference (12)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return