Machine Learning, Geoscience, and Black Boxes

When a plane crashes, the black box is considered the holy grail by investigators, containing most of the important and relevant information regarding the event. However, in the context of Deep Learning & Machine Learning, black boxes are the opposite. Tightly locked, opaque, and somehow all-knowing, these types of black boxes regularly generate results that are used by the energy industry to make very expensive decisions, often without any sort of quantitative quality control (QC) of results or complete understanding of how exactly the black box generated them.

It doesn’t take an advanced degree to understand that applying results from these “magic” processes to actual drilling decisions is insanely dangerous for several reasons. The most obvious being that the network used in the black box was likely NOT trained on the data you are feeding it. Basic parameterization exposed to geoscientists in these tools give the illusion of network control without really ceding any significant control at all. Let’s not forget processing times can be on the order of days or weeks before you can supply feedback to the network…at which point you will need to run it again from scratch. Many of these black box tools and processes require you to generate a full result before modifying weights, biases, or the network training set, making them inefficient for long-term adaptation. Thus, we introduce the concept of Interactive Machine Learning (Interactive ML), wherein the geoscientist remains in the figurative driver’s seat and steers the network as it learns - the adjustment and optimization of training set information provides immediate feedback to the network, which adjusts weights and biases accordingly in real-time. Essentially, we remove the shrouding veil of the black box and expose the Deep Learning tools to the user up-front in a familiar, geoscientific context.

Fig. Conceptual representation of interactive Machine Learning. (istockphoto.com)

There are two types of approaches to deep learning: the overwhelmingly popular approach right now is what I refer to as the “black box”. The black box is a Deep Learning network that has been trained on unknown data and is offered as a sort of processing step - essentially telling the geoscientist to “put your seismic data in here, trust us, and we will create an answer for you”. I emphasize the “trust us” because that’s exactly what you’re doing with this approach - trusting a network that has been trained on data that you have not seen and therefore have no control over, created with a set of unknown biases - to work as you intend on your data. Given the inherent variability in seismic data quality due to vintage, geologic environment, or both, it is unreasonable to assume a network trained in one part (or parts) of the world by others with varying levels of skills and expertise will work where you need it to. These datasets may contain faulty labels which could translate to erroneous predictions within your area of interest. In fact, it can be quite dangerous to blindly trust the outputs of tools using a black box approach because the learning process is hidden from the expert (you).

The second type of approach is the Interactive ML approach. This approach, as mentioned earlier, removes the black box effect, and exposes the major controls for the deep learning network to the user. The interactive approach is a regional, data-specific approach that focuses on creating the best learning model for the data you are working with. The idea behind the interactive deep learning approach is to start with a clean, blank network and train it while labeling geologic features on seismic data.

Fig. Two very different approaches to machine learning within the geosciences.

This labeling process differs from the interpretation process that we’re used to in one major way - labels should only be placed precisely where the features of interest are visible, and never where our geoscience minds want to extend or propagate them. Take for example the image below, which perfectly highlights the difference between a label set and an interpretation using an obvious fault interrupted by a thick shale sequence (NW Shelf Australia; data courtesy of Australian Government). While the interpretation of this feature would include the shale section itself, the correct labeling strategy for adding meaningful information to the training set would be to label the fault where it is clearly imaged and omit the shale section at the beginning of the training cycle. The reason for this is to keep our initial training set as clean and uncorrupted as possible - the network will predict the presence of the fault through this area, after which the prediction may (or may not) be reinforced by the geoscientist placing a label where the inference is.

Fig. The difference between labeling a fault for machine learning (left) and providing a fault interpretation (right). (Bonaventure 3D, NW Shelf Australia; courtesy of Australian Government and Bluware Corp.)

It is this simple application of live, iterative reinforcement learning that makes the Interactive ML approach unique. Interpreters and oilfinders are no longer restricted to what exists inside the black box. By exposing the figurative (and literal) knobs and buttons of the network to the geoscientist, an interactive deep learning product can be generated in a fraction of the time it takes more traditional black box approaches days or even weeks to process. Creating the training set interactively allows the geoscientist to exert control over what the network is looking for and where in the 3D seismic volume it looks for it. By modifying and building on the initial label set, the geoscientist is simultaneously perfecting the input information and generating a mental image of what the earth model will look like. Generating output objects from the trained network brings you to the doorstep of static geomodelling; because the interactive approach makes deep learning tools accessible to geoscientists in a context that they are familiar with. The time it takes to go from an uninterpreted 3D seismic volume to a working, coherent earth model has been cut from months to days…or in some case hours. If you wouldn’t trust a black box to put a spud location on a map, you shouldn’t trust one to provide “interpretations” that will ultimately carry heavy financial and operational risks. AI Won’t replace seismic interpreters, but seismic interpreters who use AI will replace those who do not. Take control of your deep learning tools by applying the interactive approach and stop putting all your faith in black boxes - we’re scientists, not airplanes.