PetroSIM is our foundational engine. It applies machine learning algorithms (Manifold Learning, K-Means) and stochastic modeling directly to well logs and core data to classify facies and predict petrophysical properties.
Data preparation and transformation
Uncertainty modeling strategy
Manifold learning & clustering
Probability distributions definition
MCMC implementation
Building the scenario space
Convergence of core variables
Decision making & value function
Explora las conexiones entre los algoritmos y los procesos fÃsicos. Arrastra los nodos para ver la fÃsica en acción.
We implement a pipeline of manifold learning (dimensionality reduction via t-SNE) followed by clustering to reveal latent structures in petrophysical logs. This assigns probabilistic facies memberships (soft clustering).
The clustering is powered by algorithms like K-Means, which minimizes the intra-cluster variance, or inertia:
Expert Validation & Metrics: Quality is strictly audited using Silhouette and Trustworthiness scores, alongside our in-house metric: the Structural Consistency Index (SCI). This guarantees stable multi-well partitions that support the identification of gas zones with quantified uncertainty.
Visual representation of high-dimensional log clustering.
Instead of single deterministic values, PetroSIM computes probability density functions (PDFs). We build the scenario space utilizing Markov Chain Monte Carlo (MCMC).
MCMC allows us to sample from complex posterior probability distributions based on Bayes' Theorem:
Where \(P(\theta|D)\) is the posterior probability of the reservoir parameters \(\theta\) given the well log data \(D\). By generating thousands of realizations, we calibrate the results against core variables and map the full spectrum of geological uncertainty (P10, P50, P90).