Integrating physics-informed machine learning with resonance effect for structural dynamic performance modeling

Jiaxin Zhang, Xiaoming Lei, Pak wai Chan, You Dong

Research output: Journal article publicationJournal articleAcademic researchpeer-review


Accurate prediction the dynamic response of structures under various loads is a complex and challenging task due to the uncertainties and intricacies of loading scenarios and analysis methods. Traditional approaches to structural analysis and modeling often rely on physics-based simulations, which can be computationally expensive and time-consuming. Over the years, Machine Learning (ML) methods have proven to be a formidable tool for data-driven modeling, based on in-situ monitoring data. However, these ML models are generally purely data-driven, lacking consideration for relevant physical information such as dynamic characteristics. Hence, this paper proposes a Physics-Informed Machine Learning (PIML) approach for modeling structural dynamic performance. This method embeds resonance effect, a crucial piece of physical information in real-world engineering, into the design of the multilayer perceptron (MLP) loss function. The model is validated through numerical simulations of forced vibrations in a single-degree-of-freedom system and a vehicle-bridge coupled system. The results indicate that the proposed PIML can model the system dynamic response effectively and simulate resonance effect at key frequency-domain locations accurately. Subsequently, features are extracted from the real-world monitoring data of a high-rise building dynamic response during Typhoon Hato, and PIML is applied for training. The results reveal that PIML significantly outperforms pure ML algorithms, even with a small dataset, and can effectively capture resonance effect in wind-induced vibrations. Using a well-trained PIML, a parameter analysis is conducted, comparing the dynamic performance under different wind speeds and frequency conditions. Lastly, considering the variation in wind parameters due to climate change, PIML is employed to analyze long-term structural fragility. The findings indicate significant changes in fragility from 1985 to 2045. In conclusion, the proposed PIML algorithm offers several advantages over conventional methods, such as reduced computational costs, improved accuracy, and the ability to capture complex nonlinear behaviors and resonance effect of interest, and efficiently serve in analyses of structural reliability and fragility.

Original languageEnglish
Article number108627
JournalJournal of Building Engineering
Publication statusPublished - 1 May 2024


  • Climate change
  • Fragility analysis
  • Physics-informed machine learning (PIML)
  • Resonance effect
  • Response prediction
  • Structural health monitoring

ASJC Scopus subject areas

  • Civil and Structural Engineering
  • Architecture
  • Building and Construction
  • Safety, Risk, Reliability and Quality
  • Mechanics of Materials


Dive into the research topics of 'Integrating physics-informed machine learning with resonance effect for structural dynamic performance modeling'. Together they form a unique fingerprint.

Cite this