Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Indoor airflow field reconstruction using physics-informed neural network
https://orcid.org/0000-0003-4441-9981
Abstract Obtaining a detailed indoor airflow field is important for the accurate and efficient control of indoor environmental comfort. Traditional computational fluid dynamics (CFD) methods and CFD-based surrogate models are time-consuming and sometimes produce inaccurate results because of difficulties in reproducing accurate inlet boundary conditions. Artificial neural networks (ANN) can be utilized to reconstruct indoor airflow fields directly from measurement data without building a large inaccurate and time-consuming CFD database. However, as a purely data-driven method, a normal ANN can yield unphysical results. A physics-informed neural network (PINN) is one possible solution. In this study, a PINN was introduced to reconstruct an indoor airflow field basing on measurement data (without inlet boundary conditions), and compared with ANN. The results show that the PINN produced more physical results than the ANN and is more tolerant to a reduction in the number of measurement points. In specific cases, the mean errors of the PINN results for the 98-, 32, and 16 point cases were 89%, 79%, and 70% of those of the ANN results, respectively. The PINN showed practical application potential in cases where the amount of measured data was relatively small. Comparing to traditional CFD, PINN can reconstruct the detailed airflow field directly from measurement data, avoiding inaccurate simulation conditions. Meanwhile, PINN saved 42% calculation time, comparing to CFD. Moreover, there is a potential of PINN in using less time to apply a trained PINN to a new case by transfer learning, where however CFD needs to recalculate a new case.
Highlights Physics-informed neural network (PINN) reconstructs an indoor airflow field. PINN reconstructs well from sensor data without CFD data assistance. PINN performs well while accurate inlet and outlet conditions is unknown. PINN provides more physical results than interpolation methods and data-driven ANN. PINN result for a 16-sensor case is better than ANN result for a 106-sensor case.
Indoor airflow field reconstruction using physics-informed neural network
https://orcid.org/0000-0003-4441-9981
Abstract Obtaining a detailed indoor airflow field is important for the accurate and efficient control of indoor environmental comfort. Traditional computational fluid dynamics (CFD) methods and CFD-based surrogate models are time-consuming and sometimes produce inaccurate results because of difficulties in reproducing accurate inlet boundary conditions. Artificial neural networks (ANN) can be utilized to reconstruct indoor airflow fields directly from measurement data without building a large inaccurate and time-consuming CFD database. However, as a purely data-driven method, a normal ANN can yield unphysical results. A physics-informed neural network (PINN) is one possible solution. In this study, a PINN was introduced to reconstruct an indoor airflow field basing on measurement data (without inlet boundary conditions), and compared with ANN. The results show that the PINN produced more physical results than the ANN and is more tolerant to a reduction in the number of measurement points. In specific cases, the mean errors of the PINN results for the 98-, 32, and 16 point cases were 89%, 79%, and 70% of those of the ANN results, respectively. The PINN showed practical application potential in cases where the amount of measured data was relatively small. Comparing to traditional CFD, PINN can reconstruct the detailed airflow field directly from measurement data, avoiding inaccurate simulation conditions. Meanwhile, PINN saved 42% calculation time, comparing to CFD. Moreover, there is a potential of PINN in using less time to apply a trained PINN to a new case by transfer learning, where however CFD needs to recalculate a new case.
Highlights Physics-informed neural network (PINN) reconstructs an indoor airflow field. PINN reconstructs well from sensor data without CFD data assistance. PINN performs well while accurate inlet and outlet conditions is unknown. PINN provides more physical results than interpolation methods and data-driven ANN. PINN result for a 16-sensor case is better than ANN result for a 106-sensor case.
Indoor airflow field reconstruction using physics-informed neural network
https://orcid.org/0000-0003-4441-9981
Abstract Obtaining a detailed indoor airflow field is important for the accurate and efficient control of indoor environmental comfort. Traditional computational fluid dynamics (CFD) methods and CFD-based surrogate models are time-consuming and sometimes produce inaccurate results because of difficulties in reproducing accurate inlet boundary conditions. Artificial neural networks (ANN) can be utilized to reconstruct indoor airflow fields directly from measurement data without building a large inaccurate and time-consuming CFD database. However, as a purely data-driven method, a normal ANN can yield unphysical results. A physics-informed neural network (PINN) is one possible solution. In this study, a PINN was introduced to reconstruct an indoor airflow field basing on measurement data (without inlet boundary conditions), and compared with ANN. The results show that the PINN produced more physical results than the ANN and is more tolerant to a reduction in the number of measurement points. In specific cases, the mean errors of the PINN results for the 98-, 32, and 16 point cases were 89%, 79%, and 70% of those of the ANN results, respectively. The PINN showed practical application potential in cases where the amount of measured data was relatively small. Comparing to traditional CFD, PINN can reconstruct the detailed airflow field directly from measurement data, avoiding inaccurate simulation conditions. Meanwhile, PINN saved 42% calculation time, comparing to CFD. Moreover, there is a potential of PINN in using less time to apply a trained PINN to a new case by transfer learning, where however CFD needs to recalculate a new case.
Highlights Physics-informed neural network (PINN) reconstructs an indoor airflow field. PINN reconstructs well from sensor data without CFD data assistance. PINN performs well while accurate inlet and outlet conditions is unknown. PINN provides more physical results than interpolation methods and data-driven ANN. PINN result for a 16-sensor case is better than ANN result for a 106-sensor case.
Indoor airflow field reconstruction using physics-informed neural network
Wei, Chenghao (Autor:in) / Ooka, Ryozo (Autor:in)
Building and Environment ; 242
24.06.2023
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
Comparison of different deep neural network architectures for isothermal indoor airflow prediction
| Springer Verlag | 2020