Multivariate engineering formulas discovery with knowledge‐based neural network

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 26, 2025

Abstract Multivariate engineering formulas are the foundation of various standards worldwide for constructing complex systems. Traditional formula discovery methods suffer from low efficiency, curse dimensionality, and physical interpretability. To address these limitations, this study proposes a knowledge‐based method efficiently generating multivariate directly data. The consists four components: (1) deep generative model considering dimensional homogeneity, (2) physics‐adaptive normalization multiple variables with different units, (3) feature merging algorithm grounded in dimensionality theory, (4) machine learning‐based data segmentation piecewise formulas. Experiments on two ground‐truth datasets demonstrate that our proposed improves accuracy generated by 35.6% (measured mean absolute error), compared to Eureqa program. Additionally, it enhances mechanistic interpretability results, both emerging physics‐informed neural network‐based equation methods. successfully capture implicit mechanisms experimental data, consistent theoretical analysis. Overall, holds great promise improving efficiency discovering interpretable generalizable formulas, facilitating transformation new techniques testing applications.

Language: Английский

Towards efficient structural inverse analysis based on AI-driven differentiable optimization method

Mechanical Systems and Signal Processing, Journal Year: 2025, Volume and Issue: 230, P. 112618 - 112618

Published: March 25, 2025

Language: Английский