Hygrothermal assessment of timber frame walls using a convolutional neural network

Abstract

A correct design of a timbre frame wall's composition is vital to avoid moisture damage. Unfortunately, currently, no general guidelines exist to determine the most optimal wall composition in a specific context. To develop such general guidelines, a comprehensive study is required, taking into account the inherent uncertainty and variability of involved input parameters. Such a probabilistic assessment is typically carried out through a Monte-Carlo approach, which easily becomes computationally inhibitive. This paper thus makes use of a metamodel, which mimics the complex hygrothermal model while being considerably faster. The authors previously developed a convolutional neural network and demonstrated its' capacity to predict the highly non-linear hygrothermal response of a massive masonry wall. In this paper, this network is adapted to predict the hygrothermal response for timber frame walls. A hyper-parameter optimisation is performed, leading to rules-of-thumb on the network architecture. It is shown that the network can accurately predict the hygrothermal time series, and that it can be employed with confidence to estimate the moisture damage risks. Subsequently, the network is used to calculate the hygrothermal response of 96 timber frame wall types, taking into account all influencing uncertainties. The results indicated that timber frame wall compositions should not be recommend based solely on the sd-ratio between vapour and wind barrier. A lower limit for the sd-ratio appears a good criterion to avoid mould growth, if adapted to the climate and cladding type. To avoid condensation, one should ensure either the insulation or the wind barrier can buffer the excess moisture.