Flash Calculation ML Aggregation for Generalizable Fluid Phase Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional flash calculation methods, including classical models and fully-connected neural networks, are computationally expensive and fail to generalize beyond the fluids on which they are trained, leading to inaccurate predictions and compromised collaboration in hydrocarbon mixture simulations.
Innovation Solution
A method using encoder and decoder machine learning models, combined with a set-permutation invariant aggregator function, encodes and aggregates input data to predict the phases and proportions of hydrocarbon mixtures, enabling generalization across varying numbers of components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional cubic equations of state are used with iterative techniques, then measurement precision of flash calculations is improved, but productivity deteriorates due to computational expense and time consumption
Solution Approach 1:
The patent replaces traditional iterative mechanical/mathematical solution methods with a machine learning model that has been trained on flash calculation data. The neural network learns the complex relationships between input parameters (pressure, temperature, composition) and output phases, substituting the iterative numerical solution process with a direct prediction approach that maintains accuracy while dramatically reducing computation time.
Solution Approach 2:
The patent performs preliminary training of the machine learning model using extensive flash calculation data generated from traditional equations of state. This pre-computation phase stores the learned relationships in the model's weights and biases, allowing rapid predictions during actual use without repeating the computationally expensive iterative calculations for each new input scenario.
2Manufacturing precision
If traditional neural networks are trained on fixed chemical compositions, then manufacturing precision for specific fluids is improved, but adaptability deteriorates when component sequences or numbers change
Solution Approach 1:
The patent implements a universal machine learning model architecture that can handle arbitrary numbers of chemical components and any component sequencing. The model uses a composition-aware architecture with attention mechanisms that dynamically weigh the contribution of each component based on its chemical properties and interactions, rather than relying on fixed positional encodings. This allows the single model to generalize across different fluid mixtures, component counts, and ordering arrangements.
Solution Approach 2:
The patent employs dynamic component representation where the model adapts its internal processing based on the specific input mixture composition. The attention mechanisms and gating functions dynamically adjust which components and interactions are most relevant for each prediction, allowing the model to flexibly handle varying numbers of components and their different sequences without requiring retraining or fixed architectural assumptions.
Data Source
AI summary
A method may include obtaining input data including an environmental condition and chemical properties of input components of an input fluid mixture, encoding, by an encoder machine learning model, the input data to obtain encoded input data, and receiving, by an aggregator function and from the encoder machine learning model, the encoded input data ordered in a sequence corresponding to an order of the input components. The method may further include aggregating, by the aggregator function, the encoded input data to obtain aggregated input data. The aggregated input data may be independent of the sequence. The method may further include decoding, by a decoder machine learning model, the aggregated input data to obtain output data including a phase for an output mixture, and presenting the output data.