AASPI Training

To install AASPI on Windows, you can request the latest version by emailing aaspi@ou.edu (if your company doesn't already have access). Once installed, double-click the AASPI shortcut on your desktop. A directory browser will appear, allowing you to choose the working directory where AASPI output files will be saved.

Accessing AASPI on Linux typically involves assistance from your company's IT staff. They will help define the correct path to the AASPI software in your .bashrc file. You also need to create a .datapath file in your home directory to specify where the binary data will be written. To start AASPI, open an xterm, navigate to the desired directory, and type aaspi_util & to run the program in the background.

The SEGY format, while considered a standard, has variations. AASPI uses a format compatible with the Stanford Exploration Project (SEP) and other research formats. Converting SEGY to AASPI format involves defining the byte locations in the trace headers for x, y coordinates, inline number, and crossline number. The process includes browsing to the SEGY file, viewing the SEGY line header content, entering the appropriate byte locations, specifying a unique project name, and defining the AASPI output file name.

When loading seismic data, especially in the AASPI environment, it's important to give the data a Unique Project Name to avoid overwriting files from other projects. For the output file name, you may name the input data d_mig followed by the Unique Project Name for easy identification. Also, ensure the file name is terminated with the .H extension, indicating an ASCII format data header file (following the SEP convention).

Several common issues can arise when preparing seismic data in SEGY format. These include:

• Incorrect or missing byte locations: Ensure the byte locations for trace headers defining survey geometry (x, y coordinates, scalco) are correctly defined.
• Scalco parameter errors: Check if the scalco parameter has been incorrectly modified during data conversion to floating-point format. If the scalco parameter was not set to 1, then this can result in incorrect bin sizes.
• Corrupted sample interval: Verify that the vertical sample increment in the SEGY file is accurate and compatible with other data volumes.
• Data truncation from 16-bit conversion: If data were previously stored as 16-bit integers, define a level that is numerically zero to mitigate truncation and round-off errors.
• Spikes: Define a threshold as a multiple of RMS amplitude to identify and handle data spikes caused by transcription, transfer, or processing issues.
• Incorrect vertical and horizontal units: Confirm that the vertical (e.g., 's', 'ms', 'm', 'km', 'ft', 'kft') and horizontal units specified in the SEGY file are correct.
• IBM vs. IEEE format: If the data values are ridiculously large, ensure that the proper format is selected (IBM or IEEE). Also, big-endian and little-endian formats may cause problems and should be checked to ensure that the data are correctly formatted.

IBM format refers to the older hexadecimal floating-point format, while IEEE is the more common modern format. Big-endian and little-endian refer to the byte order in which data is stored. These formats are crucial because using the wrong format can lead to data being interpreted incorrectly, resulting in ridiculously large or nonsensical values. SEGY headers should indicate the data format, but it's not always reliable.

The AASPI QC Plotting tab allows you to display your dataset. You can modify the axis limits and increments to view vertical slices or time slices. The software offers statistical data scaling and the ability to plot data between defined minimum and maximum values. You can adjust plotting parameters, color tables, and other settings to effectively visualize and QC your data.

To set up AASPI default parameters, invoke aaspi_util and select the "Set AASPI default parameters" tab. You can define the node list and processors per node for parallel processing. For running on Linux servers, specify the node names and the number of processors on each node. You can also configure parameters for batch submission protocols like LSF, PBS, and SLURM and header byte locations for consistent data handling.

Spectral decomposition breaks down a broadband seismic signal into its constituent spectral components (magnitude and phase) at different frequencies. It's used to analyze thin bed tuning, attenuation, and other interference phenomena in seismic data. AASPI provides different time-frequency decompositions based on matching pursuit, the continuous wavelet transform, maximum entropy, and simple correlation.

The dip3d program computes volumetric dip and azimuth, essential for calculating other geometric attributes and image processing algorithms. It offers two algorithm choices: the Gradient Structure Tensor (GST) and a discrete semblance search algorithm. GST is generally faster and maps dips up to 90 degrees, while the semblance search might be preferred when GST confuses fault planes with stratigraphic reflectors.

The inline and crossline dip components are fundamental outputs of the dip3d program and are used by almost all subsequent programs that compute geometric attributes. Thus these two volumes are always checked by default.

The similarity3d program computes coherence and energy gradients along structural dip. Coherence measures the similarity of seismic waveforms, while energy gradients reflect amplitude variations. Other available attributes include semblance coherence, coherent energy weighted gradients.

The similarity3d program can compute coherence, gradient, and energy attributes on band-limited versions of the seismic data, which are defined by filter banks. Multispectral coherence involves combining the covariance matrices computed from each of these band-limited versions to create a combined attribute volume. This approach can help to highlight frequency-dependent features and reduce incoherent noise.

IBM format refers to the older hexadecimal floating-point format, while IEEE is the more common modern format. Big-endian and little-endian refer to the byte order in which data is stored. These formats are crucial because using the wrong format can lead to data being interpreted incorrectly, resulting in ridiculously large or nonsensical values. SEGY headers should indicate the data format, but it's not always reliable.

The glcm3d program computes texture attributes based on the gray-level co-occurrence matrix (GLCM). These textures capture statistical variations in seismic amplitudes and are sensitive to features like diagenetic alteration, chaotic deformation, and fractures. While they have limited value in interactive interpretation, they can be useful for seismic facies classification using machine learning techniques.

Alpha blending, opacity, and transparency are interchangeable terms in corendering. It involves assigning varying levels of transparency to different attribute values. For example, in a two-layer alpha blend, the base layer is always opaque, while the second layer's transparency is controlled by the values of a second attribute. By setting low values of an attribute to be opaque and high values to be transparent, the underlying base layer is revealed in areas where the attribute is high. This technique allows interpreters to see through one attribute to the underlying seismic data or other attributes, facilitating correlation.

The Hue, Saturation, Lightness (HSL) color model can be emulated in corendering using alpha blending. The base layer is plotted against a polychromatic color bar (Hue). The second layer, plotted against a monochromatic gray color bar, defines the saturation (S), while the third layer, plotted against a monochrome black or binary black-white color bar, defines the lightness (L). By carefully choosing color bars and transparency settings for each layer, an HSL-like display can be created, allowing for the visualization of three different attribute components simultaneously.

The RGB color addition model assigns different spectral magnitude components to monochrome red, green, and blue color bars. Typically, a lower frequency component is assigned to red, an intermediate frequency component to green, and a higher frequency component to blue, mimicking the visible light spectrum. Combining these color-coded spectral components creates a composite image that highlights variations in spectral content, revealing subtle geological features that might not be apparent in broadband seismic data.

When loading seismic data, especially in the AASPI environment, it's important to give the data a Unique Project Name to avoid overwriting files from other projects. For the output file name, you may name the input data d_mig followed by the Unique Project Name for easy identification. Also, ensure the file name is terminated with the .H extension, indicating an ASCII format data header file (following the SEP convention).

Crossplotting involves plotting the values of two different seismic attributes against each other on a 2D graph. This technique helps interpreters evaluate how attributes are correlated through the underlying geology. The resulting 2D color legend shows how different combinations of attribute values map to different colors, revealing clusters of voxels with similar properties. By corendering the crossplotted volume with seismic amplitude, interpreters can visually correlate the attribute relationships to the original seismic data and identify geological features of interest.

Multiattribute display techniques like opacity blending, the HSL, RGB, and CMY color models, and crossplotting are limited in the number of attributes they can effectively visualize (typically two or three). Crossplotting is limited to defining hexahedral clusters in 3D attribute space. To address higher dimensionality, projection techniques such as principal component analysis (PCA) and independent component analysis (ICA), as well as more general machine learning algorithms, are required. These techniques can reduce the dimensionality of the data while preserving important information, enabling the visualization and analysis of more complex relationships between multiple attributes.

Amplitude balancing, typically through Automatic Gain Control (AGC), is generally avoided because it can negatively impact relative amplitude processing that has already been applied to the data. Relative amplitudes are important for attributes and inversions. However, if absolutely necessary, it can be applied to equalize the amplitudes across the entire dataset.

Spectral balancing is a process designed to flatten the amplitude spectrum of seismic data, thereby broadening the bandwidth and improving temporal resolution. The goal is to modify the seismic wavelet so it is uniform across the frequency spectrum. The Continuous Wavelet Transform (CWT) can be used to decompose seismic data into its spectral components. By manipulating these spectral components (e.g., flattening or "blueing" the spectrum) and then reconstructing the data, a spectrally balanced or "blued" data volume can be created. The AASPI software offers several spectral decomposition algorithms using the CWT.

While spectral balancing aims for a flat spectrum, spectral bluing enhances the higher frequencies within the seismic data. The assumption is that the reflectivity spectrum from well logs often behaves as fβ, where 0.0 < β < 0.4. Spectral bluing seeks to emulate this reflectivity response, resulting in a data volume that is considered to be closer to the true reflectivity of the subsurface and improves the relative impedance.

Acquisition footprint refers to patterns in seismic data (often visible on time or depth slices) that are directly related to the acquisition geometry (source and receiver positions). This coherent noise can obscure subtle geological features. Seismic attributes like energy ratio similarity, Sobel filter similarity, and curvedness can be used to enhance and visualize acquisition footprint, making it easier to identify and subsequently suppress. The Sobel filter is particularly useful for highlighting footprint due to its sensitivity to lateral amplitude changes.

Structure-oriented filtering is a technique that filters seismic data along the local structural dip, preserving geological features while suppressing noise that cuts across the structure. This is beneficial because it reduces noise without blurring or distorting the underlying geology, leading to a clearer seismic image and improved attribute calculations. Kuwahara filtering can be useful for well-balanced data, and alpha-trimmed mean filters for data with amplitude spikes.

Interpreters can evaluate the effectiveness of data conditioning processes by examining the data before and after processing. Key areas to focus on are unconformities and faults. Spectral plots should be reviewed to see how the spectrum was altered by the process. Also review the data for ringing due to being too aggressive with the high frequencies. Comparison of seismic attributes, such as coherence, can be used to gauge the impact on noise reduction and structural clarity. Visual inspection of time and depth slices and animation between pre- and post-conditioned data is also crucial for identifying subtle improvements in resolution and signal clarity.

Balancing lower (apparent) spectral components can be important for impedance inversion, which requires accurate low-frequency information. However, these lower frequencies are often contaminated by noise, including migration operator aliasing. The interpreter must decide whether to use the noisy low-frequency components directly or to replace them with a low-frequency model derived from well control. A compromise solution is to apply non-edge-preserving structure-oriented filtering to the lower-frequency components to reduce the impact of steeply dipping noise before spectral balancing, potentially enabling a broader band impedance inversion.

Dip-oriented filters are used to smooth or sharpen attribute volumes, enabling interpreters to focus on specific structural features as they appear on time or horizon slices. The document mentions filters like mean, alpha-trimmed mean, LUM, Laplacian of Gaussian, and MSMTM. An example of an alpha-trimmed mean filter is given to show how faults can either be smeared or enhanced by removing the filtered trend.

Coherence attributes, while useful for identifying edges, often provide "broken" or "fuzzy" images of faults. Additionally, they can highlight stratigraphic edges, unconformities, and condensed sections that are subparallel to stratigraphic dip, as well as stair-step artifacts. The fault enhancement workflow aims to address these issues by producing smoother and more continuous fault images.

The fault enhancement workflow uses dip volumes to filter out coherence anomalies that are subparallel to reflector dip. By setting thresholds based on the angle between the coherence anomaly and the stratigraphic reflector, the workflow retains anomalies that are more likely to represent faults and rejects those that are more likely to represent stratigraphic features.

Important parameters in the fault enhancement workflow include: the input edge attribute (e.g., energy ratio coherence), inline and crossline dip volumes, a total energy volume, the angles relative to stratigraphic dip to accept or reject fault anomalies, and the fault opacity value. These parameters allow for fine-tuning the fault enhancement process.

Fault skeletonization aims to further refine the enhanced fault images by reducing the "thickness" of the fault probability anomalies. It represents the fault with voxels that straddle it or a single voxel where the fault falls directly on a voxel, leading to improved resolution and a clearer visualization of the fault structure.

Corendering techniques allow for the simultaneous display of multiple fault attributes, such as fault dip azimuth, dip magnitude, and fault probability. By using different color bars and opacity settings for each attribute, interpreters can gain a more comprehensive understanding of the fault's geometry and likelihood.

The CNN learns by iteratively comparing its predictions to a baseline of manually labeled fault data. It uses 3D convolutional operators followed by nonlinear activation processes. The coefficients of these convolutional operators are adjusted ("learned") to minimize the difference between the predicted fault locations and the labeled ground truth. Each convolution and activation step is considered a "neuron" whose output can be thought of as a seismic attribute whose physical meaning the user does not have to predefine.

The current AASPI implementation uses Python with the TensorFlow package provided by Google. It can run on either a graphics processing unit (GPU) or on a single node using all available processors. Pre-trained models, like the one by Xinming Wu, are also utilized.

The CNN used was trained on 128x128x128 voxel subcubes, which then determines the size of the predicted output volumes. Overlap between adjacent subcubes (e.g., 32 voxels) is applied and the results summed to minimize edge effects resulting from the processing of each subcube, where a linear taper is applied along the edges of the subcube.

The process outputs two files. One file provides the tapers applied to each 128x128x128 voxel sample. The other file contains the fault probability prediction volume, representing the likelihood of a fault being present at each voxel location.

Skeletonization is a process used to thin and sharpen the fault predictions obtained from the CNN. The CNN fault prediction volume is used as input to the AASPI program fault_enhancement, to estimate fault orientation (dip magnitude and dip azimuth). These estimates are then run through AASPI program skeletonize3d. This helps to refine the fault representations, making them easier to visualize and interpret.

The CNN fault probability volume can be corendered with other attributes like fault dip azimuth and fault dip magnitude, or with the conditioned seismic amplitude data. This combined visualization allows for a better understanding of the relationship between faults and the surrounding geological features.

The Kuwahara filter operates by searching overlapping sub-windows within a larger analysis window. It calculates statistical measures (mean, median, standard deviation, inverse coefficient of variation) within each sub-window. The filter then replaces the central voxel's value with the median (or mean) value from the sub-window that exhibits the highest inverse coefficient of variation. This process effectively smooths the data within homogeneous regions while preserving edges where attribute values change rapidly.

The inverse coefficient of variation (V-1 = μ/σ) is used to determine the "best" analysis window for a given voxel. Unlike using standard deviation alone, the inverse coefficient of variation considers both the mean and the standard deviation of attribute values within a window. This is particularly important for seismic attributes where the mean value can vary significantly across different geologic features. The window with the highest inverse coefficient of variation is considered to be the most homogeneous, and its median (or mean) value is used to replace the central voxel's value.

compute basic statistics (mean, standard deviation, inverse coefficient of variation, and percentiles) for the input attribute volume. Second, the kuwahara3d program uses the statistics generated by stat3d to apply the Kuwahara filter, selecting the appropriate median or mean value based on the inverse coefficient of variation. Both programs are applied along structural dip.

The size of the analysis window controls the degree of smoothing and noise suppression. A larger window results in a smoother, but potentially blockier, image. A smaller window reduces blockiness but provides less smoothing. Cascading multiple Kuwahara filters with smaller window sizes can provide superior results (better smoothing and edge preservation) compared to using a single filter with a large window.

While mean and median filters can smooth data, they often blur edges and may not effectively remove "salt and pepper" noise. The Kuwahara filter is designed to preserve edges while smoothing homogeneous regions. It selects the most homogeneous sub-window based on the inverse coefficient of variation, allowing it to better maintain the sharpness of boundaries between different geologic features.

The Kuwahara filter preconditions seismic attribute volumes by smoothing intra-facies variations and sharpening inter-facies boundaries. This makes the data more suitable for subsequent seismic facies classification using machine learning algorithms. By reducing noise and enhancing the contrast between different facies, the Kuwahara filter helps improve the accuracy and reliability of machine learning-based facies classification.

PCA decomposes the data into principal components that represent the variability of the data. It identifies eigenvectors that define orientations in multidimensional attribute space, capturing the data's variance. The first few principal components capture the majority of the signal, while later components often represent random noise. PCA assumes that most of the useful information is contained within these first few components, allowing the interpreter to ignore the later principal components that contribute little to the data's variability.

The "curse of dimensionality" refers to the phenomenon where the volume of N-dimensional space increases so rapidly that the data becomes sparse. This can limit the usefulness of any statistical analysis because the patterns become harder to discern. In seismic analysis, as more attributes are added, the risk of overfitting and misinterpretation increases, making dimensionality reduction techniques crucial.

The AASPI PCA workflow includes:

1. Generate regularized training data by selecting seismic attributes and defining a temporal operation window;
2. Analyze unsupervised ML models (PCA) to test different parameters and select optimal modeling parameters;
3. Create an unsupervised model using the selected parameters; and
4. Perform unsupervised ML analysis by applying the model to the input attribute volumes and projecting each voxel, resulting in principal components.

Unlike PCA, which focuses on representing the data's variability through uncorrelated components, ICA aims to separate a multivariate dataset into independent components using higher-order statistics. ICA identifies components that are statistically independent, allowing it to distinguish between signal and noise based on the features in the data volume. The interpreter then judges which independent components represent meaningful geologic features versus noise.

The AASPI toolbox allows users to define an initial training dataset and then apply both PCA and ICA algorithms using the same inputs. This enables an unbiased comparison between the algorithms, allowing users to assess which technique is more effective for a specific dataset and geologic problem.

The resulting principal components (from PCA) or independent components (from ICA) can be viewed individually using the AASPI 2D Plot program. Crossplots of any two components can also be generated to visualize the relationships between them. These crossplots and component displays, combined with seismic geomorphology concepts and vertical slices through the seismic amplitude data, aid in interpreting geologic features such as channels and fans. The order of the independent components is undefined and should be sorted by visual inspection based on geological insight.

K-means clustering groups data points (voxels in a seismic volume) into k clusters based on their proximity in attribute space. The algorithm iteratively assigns each voxel to the nearest cluster center and then recalculates the cluster centers based on the assigned voxels. This process repeats until the cluster assignments stabilize or a defined tolerance is met. The provided workflow is as follows: 1. Generate Training Data, 2. Analyze unsupervised ML Models, 3. Create a k-means clustering model, 4. Apply the k-means model to the entire input volumes, 5. Display the results using crossplot or corendering.

The Mahalanobis distance is a measure of distance between a point and a distribution, accounting for the covariance between different attributes. This is crucial when working with seismic attributes, as they often have different units, value ranges, and correlations. Using the Mahalanobis distance, which incorporates the inverse of the covariance matrix, provides a more accurate measure of similarity than simpler metrics like Euclidean distance.

Key parameters include:

Minimum and maximum number of clusters: This defines the range of cluster numbers the algorithm will test to determine the optimal number of clusters.

Random seed: This ensures the repeatability of the results by controlling the random number generation used in initializing cluster centers.

Number of principal axes for cluster projections: This specifies how many principal components the cluster centers are projected onto for visualization and analysis.

Data clipping percentile: This mitigates the impact of extreme values on histograms and partitioning maps.

Clustering tolerance: This determines the threshold at which the iterative clustering process stops based on the fraction of data points changing clusters.

Use L1-norm instead of L2-norm distance: Chooses between using true distance (L1) or squared distance (L2) for clustering, with L1 being more precise but slower.

The "best" number of clusters is not simply the one with the lowest total distance to cluster centers. Instead, it's determined by analyzing the total distance curve. The AASPI software looks for an "elbow" in the curve – a point beyond which adding more clusters doesn't significantly reduce the total distance. Mathematically, this is approximated as the point with the lowest third derivative of the total distance curve.

Projecting cluster centers onto principal axes allows for the visualization and analysis of the relationships between clusters. Instead of simply outputting cluster numbers, the AASPI software provides the projections of the k centroids onto the principal axes. Clusters that are close together in the projected space are likely to represent similar seismic facies, which aids in geological interpretation. These projections can be visualized using crossplots or RGB corendering, enabling interpreters to understand the relationships between different clusters and their geological significance.

The primary method for interpreting K-means results is through corendering the output volumes corresponding to the principal axes. This involves assigning each principal axis projection to a color channel (Red, Green, Blue) and blending them. By examining the resulting RGB image, interpreters can visually identify different seismic facies based on their color signatures. It is important to use "All Positive" statistical scaling for RGB blending to avoid artifacts in the results. Crossplots of different principal axis projections are another way to visualize the clustering and identify relationships between attributes and clusters.

The AASPI implementation of SOM and GTM limits itself to 2D latent spaces and 2D manifolds. These algorithms initialize with a planar manifold defined by the first two eigenvectors of the N-dimensional attribute data. This manifold is mapped to a 2D latent space using the first two principal components, providing a local coordinate system. Through iteration, the 2D manifold better represents the N-dimensional data. After training, each N-dimensional data vector is projected onto the 2D manifold and mapped to the corresponding 2D latent space, which can be visualized and crossplotted.

Key parameters in som3d include:

• Input seismic data files: The suite of attributes to be used in the classification.
• Unique project name: A name to identify the output files.
• Number of input attributes: Automatically updated when files are selected.
• Maximum number of classes: An over-defined number of classes to start with.
• Number of standard deviation values to span the latent space: Defines the extent of the 2D latent space, with 3 standard deviations representing 97% of the data.
• Scale of the initial neighborhood radius: Defines a circle within which neighboring prototype vectors are updated.
• Maximum number of data training iterations: The number of times the algorithm will iterate through the training data.
• Decimation rate of input data: Used to generate smaller volumes for training.
• Grid spacing for the cluster projection space: Defines the resolution of the latent space grid.

Additionally, you can define a mask for non-uniform training data and use data-adaptive weights for input attributes.

The som3d program allows the user to define a mask file specifying locations for training data, instead of uniformly sampling the input attribute volumes. This is useful for focusing the SOM analysis on a smaller region of interest or emphasizing less common but more interpretationally important parts of the data. Such a mask file can be generated using the Mask Generator utility. By defining the training data to be more sensitive to specific target facies, a level of supervision can be added to an otherwise unsupervised learning algorithm.

The SOM results are best displayed by crossplotting the two SOM components (SOM axis 1 and SOM axis 2). The crossplot program in AASPI can be used to generate a 2D color bar, a 2D histogram showing the probability density of SOM1 vs SOM2, the color bar corendered with the histogram, and a crossplotted data volume. This last volume can be converted into SEGY for loading into interpretation software. Interpreters can define polygons in the crossplot space to construct geobodies.

While both SOM and GTM map high-dimensional data to a lower-dimensional latent space, GTM is a probabilistic reformulation of SOM that addresses some of its shortcomings. GTM does not require manual selection of the training radius and learning rate, and it provides a measure of "confidence" in the final clustering results through a defined cost function. GTM uses Gaussian functions in latent space, with grid points sharing a "responsibility" in representing each data vector, allowing for different ways to visualize the relationship.

Even though GTM is generally an unsupervised classification algorithm, gtm3d allows for a posteriori supervision, allowing users to define seismic facies on the input data volume. The program then projects the attributes onto the previously defined manifold and compares this projection to the projections of the data vectors at each voxel using the Bhattacharyya distance. This analysis measures the similarity between the data sample and the defined facies. This supervision is useful for refining and validating the GTM classification based on interpreter knowledge and geological context.

PNNs are feedforward neural networks that use Bayes's criteria and Parzen windows to estimate the probability density function from random samples. In seismic facies classification, PNNs classify unknown seismic data points into specific facies based on their attributes. They calculate the distance between validation attributes and training attributes, apply a Gaussian activation function, compute the average estimated density function for each class, and assign the unknown sample to the class with the highest probability.

The "exhaustive PNN" approach combines Probabilistic Neural Networks with an exhaustive search algorithm. This method determines the best combination of seismic attributes to distinguish between different seismic facies (e.g., salt vs. nonsalt). It addresses the limitations of stepwise linear regression by testing all possible attribute combinations and identifying non-linear relationships between them, ultimately leading to a more accurate classification.

The key steps include:

• Choosing a suite of candidate seismic attributes based on geological insight.
• Applying a 3D Kuwahara median filter to smooth and block the attributes.
• Defining polygons for each facies to create training and validation datasets.
• Applying a robust scaling scheme to avoid bias due to different attribute units.
• Using the exhaustive PNN algorithm to test different attribute combinations and smoothing parameters (r) to find the combination with the lowest validation error.
• Optionally, applying a first-order gradient optimization technique (Adam) to further minimize the validation error.
• Performing supervised seismic facies classification using the selected attribute subset and computing the probability of each class to measure classification confidence.

Interpreters pick polygons on 2D displays (inline, crossline, or time slice) and assign them to specific facies of interest. These polygons are then converted to point sets, where each point is defined by its inline, crossline, and time/depth coordinates. This process essentially labels specific regions of the seismic data with the facies they represent, providing the training data for the supervised learning algorithm. It is recommended to co-render two to three attributes together to better delineate different facies when picking polygons.

Converting polygons to point sets translates the interpreter-defined regions into a collection of discrete points (voxels) within the 3D seismic volume. Decimation, which involves skipping points at specific intervals (e.g., every 5 inlines x 5 crosslines x 5 vertical samples), reduces the computational cost of the machine learning algorithm, particularly for computationally intensive methods like PNN.

The steps in analyzing supervised ML models in AASPI are:

• Load and check training data consistency: Loads the extracted training data files and checks for missing files or inconsistencies in the number of samples. 
• Manage Attributes: Views, adds, removes and applys a scaling algorithm to attributes used.
• Manage Point Sets: Define training and validation point sets. Supports user-defined and one-point-set cross-validation methods.
• Manage Facies: Facilitates viewing and editing the facies list, but should be avoided unless merging datasets with different lists.
• Set Runtime Parameters: Defines the unique project name and suffix for the analysis.
• Parallelization Parameters: Sets MPI parameters for parallel processing to speed up computations.
• Select and run the desired algorithm.

The final steps involve:

• Browsing for an AASPI-formatted model file.
• Verifying/modifying the attribute list, if needed, ensuring consistent geometric configuration across attributes.
• Defining vertical limits to constrain the analysis to a specific region of interest. 
• Setting up parallel processing parameters.
• Executing the classification.

The generated output files include:

• PNN facies prediction volume: Classifies each voxel into one of the defined facies.
• PNN facies probability volumes: Provides the probability of each facies for each voxel, indicating the confidence in the classification.
• PNN facies maximum probability volume: Shows the maximum probability value for each voxel.
• PNN facies modulated probability volume: Shows the probability value multiplied by other parameters, to enhance the prediction value