Machine Learning Timestep Selection for Numerical Solvers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Iterative numerical solvers for differential equations, such as those used in fluid flow simulations in porous media, face challenges in converging to a solution due to the need for manual selection of timestep sizes, leading to inefficient simulation runtime and resource utilization.
Innovation Solution
A machine learning-based approach using Artificial Neural Networks (ANNs) predicts the optimal timestep size by training on datasets covering a wide range of solver behavior and convergence information, ensuring successful convergence and minimizing execution time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual timestep size selection is used in iterative numerical solvers, then the solver can converge to a solution, but the simulation runtime increases significantly
Solution Approach 1:
The machine learning model is trained in advance on datasets covering a wide range of solver behavior and convergence information. This preliminary training enables the model to predict optimal timestep sizes before the actual simulation runs, eliminating the need for manual timestep selection during execution and significantly reducing simulation runtime while maintaining convergence reliability.
Solution Approach 2:
The manual mechanical process of selecting and adjusting timestep sizes is replaced with an automated machine learning-based prediction system. The ML model substitutes human intervention and iterative manual adjustment with an automated algorithm that predicts optimal timestep sizes based on learned patterns from training data, thereby reducing runtime without sacrificing convergence.
2Reliability
If smaller timestep sizes are used to ensure convergence, then numerical stability improves, but the number of iterations increases and execution time increases
Solution Approach 1:
The machine learning model predicts optimal timestep sizes by learning from training data that encompasses various solver behaviors and convergence patterns. This enables dynamic adjustment of timestep parameters based on the specific problem characteristics and solver state, achieving numerical stability without unnecessarily small timesteps, thus reducing the number of iterations and execution time.
Solution Approach 2:
The system uses feedback from training data that includes solver behavior and convergence information to continuously improve timestep size predictions. The ML model learns from past simulation outcomes and adjusts its predictions to maintain numerical stability while optimizing execution time, creating a closed-loop system that balances stability and efficiency.
3Reliability
If iterative methods are used to solve PDEs, then acceptable solutions can be found, but the process requires repeated attempts with reduced timestep sizes increasing computation time
Solution Approach 1:
The machine learning model performs preliminary learning from training data that includes a wide range of solver behaviors and convergence information before actual simulation. This preliminary action enables the model to predict optimal timestep sizes upfront, avoiding the need for repeated attempts with reduced timesteps during the actual simulation process, thereby reducing computation time while ensuring solution convergence.
Solution Approach 2:
The iterative mechanical process of trial-and-error timestep adjustment is replaced with a machine learning-based prediction system. The ML model substitutes repeated solver attempts with a single prediction step that provides optimal timestep sizes, eliminating the computational overhead of multiple failed convergence attempts while maintaining solution reliability.
Data Source
AI summary
A method for accelerating numerical solution of a differential equation representing fluid flow in porous media associated with hydrocarbon well environments involves obtaining input data associated with a previous timestep of a numerical solver operating on the differential equation, predicting, by a machine learning model, a current timestep size for the numerical solver from the previous timestep to a current timestep immediately following the previous timestep, and executing the numerical solver using the current timestep size on the differential equation to generate a simulation output for the current timestep.