An Assessment Of Tool Interoperability And Its Effect On Technological Uptake For Urban Microclimate Prediction With Deep Learning Models

Nariddh Khean Bauhaus-University Weimar; City Intelligence Lab, Austrian Institute of Technology
Serjoscha Düring Bauhaus-University Weimar; City Intelligence Lab, Austrian Institute of Technology
Angelos Chronis City Intelligence Lab, Austrian Institute of Technology
Reinhard König Bauhaus-University Weimar; City Intelligence Lab, Austrian Institute of Technology
Matthias Hank Haeusler University of New South Wales

There is a high cost associated with training deep learning models: data acquisition and engineering, computation time and resources for training, and continual model assessment and retraining upon conceptual drift. However, the benefits of the resultant models can outweigh these costs. Deep learning (DL) models rose to dominate in the fields of computer vision and natural language processing due their unique characteristics: they are fast and possess a heightened ability to generalise, but often do so at the cost of accuracy. These attributes make DL models uniquely equipped to tackle computationally intensive simulation problems. The focus of this research will be on the application of DL models in urban microclimate simulations based on computational fluid dynamics (CFD). These models have the potential to offer great strides toward the 11th Sustainable Development Goal. Due to the speed of these models, their use for assessing the microclimatic impact of urban plans can shift from an afterthought to earlier stages in the design process, facilitating “inclusive and sustainable urbanisation and [bolstering the] capacity for participatory, integrated and sustainable human settlement planning and management” (UN General Assembly, 2015). Yet, despite the work that has been done in this area and the proven benefits provided by the DL approach, adoption of these models are still low. The City Intelligence Lab’s (CIL) Intelligent Framework for Resilient Design (InFraReD) (Chronis, A., et al, 2020), arguably the CFD DL models that have seen the most progress put toward its dissemination, have wrapped their models with a web server stack and offers its capabilities through a web app. Since its closed alpha release, the CIL has forged partnerships with several industry partners, as well as conducted workshops around the use of InFraReD (DigitalFUTURES, 2021). However, we suspect that the web app alone will fall short toward the goal of long-term uptake and adoption by practitioners within industry. There is a high transition cost associated with new tools. And if the tool fails to fit within pre-existing workflows and processes, a cost-benefit analysis may prove too great a resistance to its integration. Attempting to control the uncertainty of industry uptake is often beyond the capability of the technology’s creators, however, the “benefits and costs can be influenced” (Hall, B. H., & Khan, B., 2003). Thus, this research intends to explore the impact on the adoption of new tools depending on changes that the tool’s creators can control, such as the avenues by which the models can be accessed. As it stands, the CFD DL models behind InFraReD can be accessed through a web app, whereby the Unity-based front-end sends GraphQL requests to InFraReD’s API. This research proposes a Grasshopper plugin that performs the functional equivalent of the web app, but in a software environment more familiar to architects and urban planners. Both methods will be offered in a workshop with participants ranging from students to industry practitioners, and their behaviours and preferences will be qualitatively assessed. The outcomes of this research will make steps toward understanding how the methods to access new tools affects the resistance to its uptake, and by extension, also ascertain some notion of the value proposition of deep learning-based microclimate predictions.

Keywords: Sgd11, Deep Learning, Technological Adoption, Computational Fluid Dynamics, Urban Microclimate Simulation, Grasshopper

View Full Paper