Uncertainty-aware surrogates for early stage design prototyping

Abstract

Surrogate modelling is a powerful tool that can help to solve both inverse problems as well as acting as a tool for early stage rapid prototyping. It is possible to create surrogate models and visualisation environments that can make it easy for clients, architects and other stakeholders to understand the performance of individual design configurations without necessarily setting up complex models.

With the advent of machine learning based regression and the ready availability of tooling, surrogate modelling has become accessible to practitioners. However, the blocker in the adoption of ML based surrogates is the lack of firm accuracy guarantees. In this talk we develop uncertainty-aware surrogates using Bayesian inference. Our surrogates can range from simple linear and polynomial regression models to more complex neural network-based approaches. In each case, the surrogate returns both a prediction and an associated uncertainty via confidence intervals. The uncertainty can then be compared with a threshold informed by the problem domain to either accept or reject the prediction. Thereby allowing the practitioner to take advantage of the rapid computational advantage without losing accuracy guarantees on unseen data.

Communicating this uncertainty to the practitioner is a key challenge as most engineers are not used to thinking of probabilistic machine learning. We demonstrate how the uncertainty-aware surrogate can be incorporated into a Rhino-based prototyping environment for developing geometries of steel framed buildings.