Application of transfer learning for the prediction of blast impulse

Abstract

Transfer learning offers the potential to increase the utility of obtained data and improve predictive model performance in a new domain, particularly useful in an environment where data is expensive to obtain such as in a blast engineering context. A successful application in this respect will improve existing surrogate modelling approaches to allow for holistic and efficient strategies to protect people and structures subjected to the effects of an explosion. This paper presents a novel application of transfer learning for the prediction of peak specific impulse where we demonstrate that previous knowledge learned when modelling spherical charges can be transferred to provide a performance benefit when modelling cylindrical charges. To evaluate the influence of transfer learning, two artificial neural network architectures were stress tested for three levels of random data removal: the first model (NN) did not implement transfer learning whilst the second model (TNN) did by including a bolt-on network to a previously published NN model trained on the spherical dataset. It is shown the TNN consistently outperforms the NN, with this out-performance increasing as the proportion of data removed increases and showing statistically significant results for the low and high threshold with less variability in all cases. This paper indicates transfer learning applications can be used successfully with considerable benefit with respect to surrogate modelling in a blast engineering context.

Blast wave interaction with structures – An overview

“…Nevertheless, the ANN was shown to be able to predict the roof loads on the building within a fraction of a second, making it a promising technique for future fast-running codes. Recently, the use of ANNs for confined loading (Dennis et al, 2020) and for near-field loading (Pannell et al, 2021(Pannell et al, , 2022a(Pannell et al, , 2022b were also successfully demonstrated.…”

A Review of Blast Loading in the Urban Environment

“…Likewise, Ruscade et al [111] use data-rich experimental work (30 pressure vs. time data sets per test) for the training of a numerical model, although no further detail is provided about the model. Other blast-related applications of data-driven [112,113] and neural network-based predictive approaches [114][115][116] have been shown to have highly favourable computa-tional times compared to CFD simulations. Flood et al [117] showcased a neural network that was capable of running as much as six orders of magnitude faster than a CFD simulation for a 3D scenario, with similar results reported by Kang and Park [118].…”

Mitigation of blast loading through blast–obstacle interaction

“…Recently, Dennis et al (2021) successfully developed a rapid artificial neural network (ANN) model to predict the blast loading in a confined internal environment. Pannell et al (2022a; 2022b) utilised ANNs to explore physics-guided regularisation and transfer learning respectively to improve the ability of an ANN to predict near-field blast loading. Another ANN methodology to model non-linear blast loading behind a blast barrier protective wall was developed by Bewick et al (2011).…”

“…This interaction has been investigated in the literature along two different lines: (i) the initial interaction between an obstacle and the blast wave to predict the frontal load, and (ii) to understand the consequences of the wave interaction in the downstream region. To predict such a load, ANN methods have successfully demonstrated the ability to estimate complex blast loads in scenarios such as a confined environment (Dennis et al, 2021), near-field blast loading (Pannell et al, 2022a; 2022b; Holgado et al, 2022) and for a target protected by a solid wall (Bewick et al, 2011). However, there is no deep insight into the characterization of blast wave interaction with a cylindrical obstacle and its role in mitigating the load acting on shadowed structural rigid targets.…”