Migrating Legacy Simulation Models via Hybrid Neural-Analytic Sub-Modules
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy computer simulations, while producing accurate results, are often ad-hoc and difficult to modify or migrate due to their complex, unwieldy code structure, making it challenging to add new features or switch to new hardware/operating systems without significant manual effort and potential loss of accuracy.
Innovation Solution
A method and system that replicate the legacy computation model using a combination of analytic and neural network sub-modules, where neural networks are trained on legacy data to reproduce the computational logic, allowing for a structured and expandable target computation model that maintains high accuracy while simplifying the migration process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If legacy computer simulations use ad-hoc code structure to produce accurate results, then simulation accuracy is maintained, but code complexity and difficulty to modify increases
Solution Approach 1:
The patent divides the legacy computation model into multiple analytic sub-models, each representing a specific functional component or process stage. This segmentation allows the complex ad-hoc code to be broken down into manageable, independently understandable units while preserving the overall simulation accuracy through systematic reconstruction of the computational logic.
Solution Approach 2:
The patent creates a target computation model that replicates the behavior and results of the legacy starting computation model. By copying the input-output relationships and computational logic of the legacy system into a structured format with analytic sub-models, the patent maintains simulation accuracy while eliminating code complexity issues.
2Adaptability or versatility
If legacy simulation code evolves ad-hoc over time to add features, then functionality is extended, but ease of manufacture and maintenance deteriorates
Solution Approach 1:
By segmenting the computation model into distinct analytic sub-models with clearly defined interfaces and responsibilities, the patent enables targeted modifications to specific features without affecting the entire codebase. Each sub-model can be independently updated or extended while maintaining the integrity of the overall system.
Solution Approach 2:
The patent transforms the rigid, ad-hoc code structure into a flexible parameter-based analytic model where features can be modified by changing parameters, equations, or sub-model configurations rather than rewriting code. This allows easy addition of new features through parameter adjustment while maintaining code stability.
3Adaptability or versatility
If manual refactoring of legacy code is performed to migrate to new systems, then code portability is improved, but time consumption and effort increases
Solution Approach 1:
The patent automatically copies the computational logic and behavior of the legacy system into a structured target model format, preserving simulation accuracy while enabling portability. This automated copying process eliminates the need for time-consuming manual refactoring and code translation.
Solution Approach 2:
The patent replaces the manual mechanical process of code refactoring and migration with an automated computational approach. By using algorithmic methods to transform the legacy model into the target analytic model, the patent eliminates the need for manual intervention while maintaining portability and accuracy.
4Loss of information
If analytic descriptions are created for all computational logic, then model understanding is improved, but device complexity increases due to detailed structuring
Solution Approach 1:
The patent segments the computational logic into analytic sub-models that provide clear descriptions of specific functional areas without requiring complete detailed structuring of the entire system. Each sub-model captures the essential analytic relationships for its domain, improving understanding while avoiding unnecessary complexity.
Solution Approach 2:
The patent applies analytic descriptions locally to specific sub-models and computational components rather than uniformly across the entire system. This allows detailed analytic understanding where needed while maintaining simplicity in other areas, balancing clarity with complexity management.
Data Source
Figure 1
Figure 2
AI summary
The invention concerns a method for migrating a starting computation model (3) to a target computation model (4), wherein the target computation model (4) replicates the starting computation model (3), wherein the starting computation model (3) and the target computation model (4) are implemented via respective computer program code and configured to compute a respective vector interface output (6) based on a respective vector interface input (5), wherein the target computation model (4) comprises a plurality of sub-models (7a, b, c) configured to compute a respective output quantity (14a, b, c) from a respective input quantity (15a, b, c), wherein a first set (9) of sub-models (7a, b) of the plurality consists of a respective neural network (8a, b), wherein a second set (11) of sub-models (7c) of the plurality consists of a respective analytic computation model (10), wherein an intermediate quantity (13a, b) is comprised by the output quantity (14a, b) of a sub-model (7a, b) of the first set (9), which intermediate quantity (13a, b) is comprised by the input quantity (15) of an interior sub-model (16), which interior sub-model (16) is comprised by the second set (11), wherein the sub-models (7a, b) of the first set (9) are trained through machine learning based on a plurality of vector training data sets, which plurality of vector training data sets conforms to the vector interface output (6) and the vector interface input (5) of the starting computation model (3). The invention also concerns a corresponding system.