Documentation

Welcome to the Mineflow documentation!

Overview

Mineflow uses your exploration data to train a neural network to predict the shape and location of mineral deposits. The custom neural networks that Mineflow develops require minimal updates or feedback from the user. Use of the Mineflow platform does not require any background in machine learning on your part.

Getting Started

Start by uploading any exploration data from your sites.

Data Uploads

1. Go to the datasets section of the Portal, press 'New Dataset,' and select the files you want to upload.

What data should I upload?

Great question. The short answer: upload any shapefile or Excel file related to the exploration site you're focusing on.

A more detailed answer: It’s easy to understand why a shapefile containing outcrops of rocks in the area is valuable for training a predictive model. But what about a shapefile with polygons representing all the houses in the area where you’re drilling? Is that useful? Absolutely. Mineflow makes no assumptions about the data distribution and will supplement your uploads with additional data it finds online. So, a file identifying house locations can help Mineflow correlate data samples. For instance, Mineflow often uses satellite data to understand a site, and when combined with your house shapefile, Mineflow can learn to distinguish houses from significant geological features visible in aerial imagery.

Non-drill data

You can upload zip files (*.zip) that contain shapefiles or shapefiles themselves (*.sbx, *.dbf, *.shp, *.shx, *.xml, *.prj, *.sbn, *.cpg) for the non-drill data.

Drill data

For your drill data, upload Excel files (.xlsx). Ideally, your files should have columns that give the starting depth and the ending depth of each segment as well as some unique identifier for each drill hole segment (e.g. hole name).

Data Designations

After you upload your data, you'll notice that each file is assigned a "designation."

A designation is Mineflow's guess as to what type of data you uploaded. There are eight possible designations:

1. Rock outcropping/expert labeled rock body: Polygons that represent rock bodies that are either outcroppings or generally agreed on rock structures underground
2. Non-rock visible object: Any object that is visible on the ground (e.g. building, river, road, tree) that has not been labeled as a rock
3. AOI: Any general area of interest, usually some sort of polygon
4. Aerial survey: All contour-line based data types. Any overhead surveys or imagery or LiDAR is going to fall into this group
5. Rock/soil/geochemical samples: Points where rocks or soil were sampled (e.g. descriptions of the rock sampled, PPMs of elements in the soil)
6. Drill collar locations and names: Data about the location of the collars
7. Drill hole traces: Files that contain the depth, dips and azimuths of your drill segments
8. Drill data: Drill assays or lithologies (or any other drill data)

Updating designation

You'll notice that you can change the designation for a file. This is useful for Mineflow when it starts to train the predictive model on your dataset.

Data Classes

After you upload your data, you'll notice that it is also assigned a "class."

A class is Mineflow's guess as to how many different types of data are represented in that file. The word class here comes from machine learning terminology where a class is a type of label that a piece of data falls into. A file can be one of two things:

1. Mono-class: Mono-class files are files that only contain one type of an object. A few examples of this: a shapefile that only contains granite rock outcroppings, a shapefile that contains lines that represent waterways in the area, a shapefile that contains the general structure (polygon) of a belt of pegmatites running through the area, a shapefile that is a set of points where you found rocks containing mica on the ground, a shapefile that has polygons that are all the houses in the area (yes, this data can be useful to upload!).
2. Multi-class: Multi-class files are files that contain multiple types of an object. A few examples of this: a shapefile that contains rock outcroppings that you have labeled as having different lithologies (e.g. granite, marble, etc.).

Mineflow creates your predictive model

After creating your dataset, the Mineflow team will "train" a model on that dataset. That means that we create a statistical model with the objective of generalizing from its experience. We will train two models for you:

1. Two-dimensional predictions: also known as prospectivity maps, these are 2D maps that represent the general shape of rock and mineral structures from a bird's eye view.
2. Three-dimensional predictions: also known as block models, these are 3D grids that represent the detailed shape and location of mineral structures underground.

The time it takes to train a predictive model depends on the size of the dataset, the size of the area and whether or not the model is 2D or 3D.

• Small datasets (less than 25 megabytes of data) for 2D will generally take 1 hour per 1 sq. mile of area. So, if I have a dataset that is a few megabytes and I've collected data over a 2 mile by 2 mile area, then the training of the 2D model will take about 4 hours.
• Large datasets (greater than 25 megabytes of data) for 2D will generally take 4 hours per 1 sq. mile of area. So, if I have a dataset that is 100 megabytes in size and I've collected data over a 2 mile by 2 mile area, then the training of the 2D model will take around 16 hours.
• 3D models generally take twice as long as 2D models, so for small datasets on a 2x2 mile area, training will take 8 hours and for large datasets on a 2x2 mile area, training will take around 32 hours.

Once the Mineflow team has trained your custom model (see training ETA just below), you'll get an email from notifications@mineflow.ai, saying:

Hi {name},

{model name} is done training. Click here to start making predictions with it!

Here is an example of the email:

When you return to the Portal, your model will appear there, marked as successfully trained:

Generating predictions

Once we have created your model, the Mineflow team will create a project where you can see an image with location, shape and size of your model's predictions.

Navigate to the Portal to see your new project.

Clicking on your project will take you to a screen with a map on it that shows all of the data you uploaded:

Then, click the predict tab and choose the model that we just trained:

Then, select the area you're interested in:

Press "generate predictions":

A few minutes (or hours depending on the size of the area) later, you'll get an email saying that your predictions are ready and when you return to the project, it will have generated both a 2D prospectivity map for that area as well as a 3D block model for that area.

Here is an example of what the 2D prospectivity map might look like:

Here is an example of what a 3D block model might look like for a copper deposit. You can manipulate the 3D shape and adjust the visibility and transparency of each of the grades of copper.

The model can generate predictions for any mineral you want.

For instance, here is the software generating a block model for a gold deposit:

Video demo

A video showing the site in use can be found here.

Best practices

1. Generally, the more data you upload, the easier it is for the model to learn to predict your deposit.
2. The more accurate your data is, the easier it is for the model to learn to predict your deposit. The more missing data and improperly labeled rocks you have in your dataset, the more likely it is that your model takes longer to train and yields less reliable results.
3. When choosing an area where you wish to generate predictions, the closer your location is to where your dataset is clustered, the better your predictions will be. "Close" in this context means within 5 miles. It is much harder for the model to make an inference about a location in the middle of the Pacific Ocean if your dataset was collected in South Africa!
4. One area where Mineflow can be particularly useful to a veteran geo is in remembering that single data point from that site you visited 15 years ago and inferring that the exploration site you're looking at now actually has a lot in common with it, despite being halfway across the globe.

FAQ

How does Mineflow compare to existing solutions?

Traditional geological modeling software usually supports estimators like Radial Basis Function (RBF), Inverse Distance estimator and techniques like wireframing and kriging. These can be used to take drill data and make guesses about what the shape of the deposit might look like.

Here's how Mineflow differs:

1. Comprehensive data integration: Our model incorporates all available data points from the site, not just drilling information, ensuring a more complete analysis. RBF/Inverse distance/others cannot support the complex, multimodal data you have collected on site (like individual rock samples and geochemical samples). All data types are supported by Mineflow. Mineflow generates predictions based on all of the data you have collected as long as it is georeferenced.

2. Cross-site learning: Mineflow allows you to leverage data from one site to inform exploration at another. If there is any similarity at all in the data collected at the two sites, Mineflow's neural net will find that correlation. RBF/Inverse distance/others are not able to learn how to make predictions based on multiple-sites worth of data.

3. Advanced pattern recognition: Mineflow uses deep learning-based AI and sophisticated machine learning algorithms to understand complex, non-linear relationships between multiple samples even if they are not close to each other on a map. RBF/Inverse distance/others are just functions that find lines between points, so they are not able to do this.

4. Identify missed deposits: Because RBF/Inverse distance/others just find lines between data points, if there is a second deposit that is close by that you haven't directly intersected with, RBF and Inverse Distance will never find it. Mineflow will likely find it.

Why is Mineflow better than joint inversion models?

1. Parametrization: In joint inversion models, you need to parametrize the shape of your deposit, which means that you need to make assumptions about the data and the way rock structures form. In the case where your assumptions about how to parametrize a deposit are correct, great! Your joint inversion model will probably take longer to train than a Mineflow model and may achieve results which are nearly as high in accuracy as Mineflow. In the case where your assumptions about how to parametrize a deposit are incorrect, your joint inversion model will be misleading and you will end up wasting exploration resources by drilling in suboptimal locations.

2. Cross-site learning: Same as point 2 in the first question.

Why are few assumptions good?

Humans have developed an strong understanding of geology and with it, powerful abstractions for how rock structures tend to form and correlate with one another. This leads to new mineral discoveries and scientific achievements regularly.

That said, when you consider any field from a computational perspective, trying to program human intuition into your model of how systems work is difficult and often misguided. When we look at the history of computational models, we find that those that excel are big and generalizable--that is, they take in a massive amount of data and they have few assumptions hard-coded into how they handle that data. The field of mineral exploration is well-poised to take advantage of the explosion of publicly available data on the natural world. From academic datasets to satellite APIs and published 43-101s, the internet is full of data that we combine with the power of AI to better predict what is underground.