Back to search

BIA-Brukerstyrt innovasjonsarena

LandSkape - Hybrid Physical-Based Deep Learning for Fast and Reliable Wind Flow Estimation

Alternative title: LandSkape - Hybrid fysisk basert dyplæring for rask og pålitelig vind estimering

Awarded: NOK 4.9 mill.

Project Manager:

Project Number:

327897

Project Period:

2021 - 2023

Funding received from:

Organisation:

With the complexity of modern urban areas, the pedestrian wind environment analysis becomes a critical factor in urban and building planning design, helping to ensure the overall well-being, safety, and comfort in pedestrian zones. The usage of fluid flow simulation enables architects and engineers to predict and optimize the performance of buildings in the early stage of the design process. Especially for cases where fast iteration time is desired, the faster data-driven learning-based surrogate model can represent a reasonable approximation of the simulation. In particular, due to their capabilities in learning complex spatial patterns and dependencies, Deep Neural Network-based architectures represent a fast and alternative solution for efficiently approximating mapping function in high dimensional spaces. The main advantage of using DNN-based architecture as a surrogate model, have proven to be 1) having a model with a high degree of generalization 2) relatively quick training time (different magnitude in comparison to CFD simulation) 3) quick inference with the advantage of interaction during the design process. In summary, traditional CFD methods produce high-accuracy results, but they are computationally expensive and do not work well in the design process of new prototypes in a given domain. To obtain results, it often takes several hours or days, depending on the prototype’s complexity. Nablaflow, together with research partners, have explored deep learning with the objective of creating an interactive tool for testing new designs, even when they are getting computationally hard for physical solvers. In particular, we have defined DL-based architectures that can generate wind flows for arbitrarily shaped buildings in scenario of different complexity (city maps) with the motivation of building a surrogate model that can be used in an interactive tool for smart building assessments.

Outcomes are: • Improved availability of accessible wind assessment tools by making them faster, and less dependent on expert knowledge and skills; • Developed deep-learning-based architectures that can generate wind flows for arbitrarily shaped buildings in scenario of different complexity (city maps); • Strengthened Norwegian industry and science as leaders in smart building assessments, part of smart city developments. Nablaflow has also as a company taken several steps forward as a result of the award, supporting the green shift in a number of use-cases including urban and building planning design and offshore wind development. Incorporating the technology developed as a result of the project into scalable tools is set to underpin growth over the next several years.

With the complexity of modern urban areas, the pedestrian wind environment analysis becomes a critical factor in urban and building planning design, helping to ensure the overall well-being, safety, and comfort in pedestrian zones. The usage of fluid flow simulation enables architects and engineers to predict and optimize the performance of buildings in the early stage of the design process. Especially for cases where fast iteration time is desired, the faster data-driven learning-based surrogate model can represent a reasonable approximation of the simulation. In particular, due to their capabilities in learning complex spatial patterns and dependencies, Deep Neural Network-based architectures represent a fast and alternative solution for efficiently approximating mapping function in high dimensional spaces. The main advantage of using DNN-based architecture as a surrogate model, will be 1) having a model with a high degree of generalization 2) relatively quick training time (different magnitude in comparison to CFD simulation) 3) quick inference with the advantage of interaction during the design process. In summary, traditional CFD methods produce high-accuracy results, but they are computationally expensive and does not work well in the design process of new prototypes in a given domain. To obtain results, it often takes several hours or days, depending on the prototype’s complexity. We aim to explore deep learning with the objective of creating an interactive tool for testing new designs, even when they are getting computationally hard for physical solvers. In particular, we plan to go in a similar direction and define DL-based architectures that can generate wind flows for arbitrarily shaped buildings in scenario of different complexity (city maps) with the motivation of building a surrogate model that can be used in an interactive tool for smart building assessments.

Funding scheme:

BIA-Brukerstyrt innovasjonsarena