Multiscale Mesh Simulation via Deterministic Sampling and Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques are unsuitable for simulating complex physical phenomena, such as fluid flow, on unstructured data representing physical geometries, and fail to reduce computing time and energy consumption effectively, limiting adaptability to available computing resources.
Innovation Solution
A method involving deterministic sampling to generate multiscale meshes from an initial mesh, combined with a machine learning algorithm like graph convolutional neural networks, to perform numerical simulations efficiently, while preserving physical coherence and reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning techniques are used to simulate complex physical phenomena on unstructured data, then computing time and energy consumption are reduced, but the techniques are unsuitable for preserving physical coherence on unstructured data representing physical geometries
Solution Approach 1:
The method segments the unstructured mesh into multiple levels of detail, creating a hierarchical representation where coarse regions capture global physical behavior and fine regions preserve local details. This segmentation allows machine learning models to operate efficiently on coarser representations while maintaining physical coherence through controlled refinement in critical areas.
Solution Approach 2:
The invention introduces a new dimension of abstraction by representing the unstructured mesh as a hierarchical tree structure with multiple levels. This dimensional transformation enables the system to navigate between different scales of representation, allowing efficient computation at coarser levels while preserving physical coherence through selective refinement at finer levels where needed.
2Use of energy by moving object
If machine learning techniques are used to simulate complex physical phenomena on unstructured data, then energy consumption is reduced, but the techniques are unsuitable for preserving physical coherence on unstructured data representing physical geometries
Solution Approach 1:
The method segments the computational domain into hierarchical levels, allowing the machine learning model to process information at multiple scales. Coarse segments handle global energy-intensive computations efficiently, while fine segments preserve local physical coherence with reduced computational overhead, overall reducing total energy consumption while maintaining reliability.
Solution Approach 2:
By introducing hierarchical levels as an additional dimension, the system can perform energy-efficient computations at coarser levels and only refine to finer levels when physical coherence requirements demand it, thus optimizing the energy-reliability tradeoff.
3Adaptability or versatility
If machine learning techniques are used to simulate complex physical phenomena, then adaptability to available computing resources is improved, but the techniques fail to reduce computing time and energy consumption effectively
Solution Approach 1:
The hierarchical mesh structure enables dynamic adaptation of computational resources. The system can adjust the level of refinement in different regions based on available computing resources, physical coherence requirements, and problem complexity, allowing efficient scaling from small to large computational budgets while maintaining effectiveness.
Solution Approach 2:
By segmenting the mesh into hierarchical levels, the system can selectively activate computation at different levels based on available resources. Coarse levels provide quick results with minimal resources, while finer levels can be engaged when more computing time and energy are available, achieving effective adaptability.
Data Source
AI summary
A method for numerical simulation of a flow of a fluid in a space around a geometry, implemented by computer, including: a so-called deterministic sampling step, configured to process the initial mesh M0 as an input so as to obtain a set of so-called simulation multiscale meshes, the set of simulation messages including a number Z≥1 of subsampled meshes Mi; a step of generating a simulation result from at least one machine learning algorithm of the neural network type, previously trained from a database including a plurality of sets of so-called training multiscale meshes each associated with a numerical simulation, to provide a set of simulation data for all or some of the nodes of a mesh Mi; in the set of multiscale meshes obtained during the deterministic sampling step.