Data-driven modelling of blast waves for safer cities

目前，防爆工程师没有足够或适当的工具来研究复杂城市环境中的整体爆炸行为。因此，防爆解决方案仅在局部层面实施，即：“如果这种大小的炸弹被放置在这个位置，这个特殊的结构元件会发生什么？”该项目旨在开发一种新的分析方法，可以快速模拟爆炸波在城市景观或城市环境中的传播，以便防爆工程师和城市规划者将防爆解决方案构建到城市和公共空间的结构中，而不是可选的“螺栓”。为此，我们将利用最先进的机器学习方法，开发一个数据驱动的爆炸传播模拟器，该模拟器基于对物理行为的近似，通过对世界领先的谢菲尔德大学爆炸和冲击实验室收集的先进数值建模和详细实验数据的询问来学习。高保真的基于物理的解决方案可以为非常具体的情况提供准确的答案。然而，即使只考虑少量的潜在情况，这些方法也会变得笨拙和不合适。另一方面，近似设计工具可以在几秒钟内评估数十万种排列，但目前可用的方法缺乏复杂性，不能准确地模拟爆炸在复杂环境中的行为，当爆炸波可能合并，衍射和反射时。基于机器学习（ML）的数据驱动建模提供了在最先进的爆炸载荷量化方面结合两种方法的潜力：快速运行的求解器可以模拟具有许多潜在排列的事件，但具有物理上有效且经过验证的假设来支撑分析。首先将研究高不确定性下数据驱动建模的研究，然后是ML建模结构中的物理约束监督学习。机器学习在大型数据集的模式识别方面显示出巨大的潜力，最近已被用于预测海浪状况[James等人（2017）预测海浪状况的机器学习框架]。海岸工程学报[j]。在这里，开发的ML模型能够在1/1000的物理模型计算时间内预测海浪。将这种方法推广到爆炸波传播的预测中，将提供一个更加鲁棒和准确的爆炸载荷预测，其精度与有限元分析相当。ML需要大量的数据输入。最初，建模结构将使用土木和结构工程城市流量实验室测量的几何3D点云进行训练，随后能够快速重建任何城市景观，因此在该城市景观中产生爆炸效果。选择数值分析将根据导入的几何图形提供训练数据（3D空间中的点网格为每次运行提供4个维度的训练数据），目的是开发更通用的预测方法。最后，开发的模型将在世界领先的谢菲尔德大学爆炸和冲击实验室进行一系列控制良好的实验测试。在这里，复杂几何形状的爆炸压力将使用最先进的高震级压力表进行测量，采样频率为~5 MHz，以确保收集到丰富的数据。

Currently, blast protection engineers are not equipped with adequate or appropriate tools to study holistic blast behaviour in complex urban environments. As such, blast protection solutions are implemented at a local level only, i.e. "what happens to this particular structural element if a bomb of this size is placed in this exact location". This project aims to develop a new analysis method that can rapidly model the propagation of a blast wave in a cityscape or urban environment, in order for blast protection engineers and city planners to build blast protection solutions into the fabric of cities and public spaces, as opposed to being optional 'bolt-ons'. To do this, we will take advantage of state-of-the-art Machine Learning methodologies and develop a data-driven blast propagation emulator, based on approximations to physical behaviour learnt through interrogation of both advanced numerical modelling and detailed experimental data gathered at the world-leading University of Sheffield Blast and Impact Laboratory. High-fidelity physics-based solvers can provide accurate answers to very specific situations. These approaches, however, become unwieldy and unsuitable even when considering only a small number of potential situations. On the other hand, approximate design tools can evaluate hundreds of thousands of permutations within seconds, but currently the available approaches are lacking in sophistication and cannot accurately model blast behaviour in complex environments, when a blast wave may coalesce, diffract, and reflect. Data-driven modelling based on Machine Learning (ML) offers the potential to combine the best of both worlds with respect to the state-of-the art in blast load quantification: quick-running solvers that can simulate events with many, many potential permutations, but with physically valid and verified assumptions underpinning the analyses. Research in data-driven modelling under high uncertainty will initially be investigated, followed by physics-constraint supervised learning in ML modelling structures.Machine learning has shown enormous potential for pattern recognition in large data sets, and has recently been used to forecast ocean wave conditions [James et al. (2017) A Machine Learning Framework to Forecast Wave Conditions. Journal of Coastal Engineering]. Here, the developed ML model was able to predict ocean waves in 1/1000th the computation time of a physics-based model. Such an approach extended to the prediction of blast wave propagation will provide a more robust and accurate blast load predictor with accuracy commensurate with FEA.ML requires large data input. Initially, the modelling structure will be trained using geometric 3D point clouds measured from the Civil and Structural Engineering Urban Flows Laboratory, with the intention of subsequently being able to rapidly recreate any cityscape and therefore blast effects within that cityscape. Select numerical analyses will provide training data (a grid of points in 3D space provides 4 dimensions of training data for each run) based on the imported geometries, with the intention of developing a more generalised predictive approach.Finally, the developed model will be validated against a series of well-controlled experimental tests conducted at the world-leading University of Sheffield Blast and Impact Laboratory. Here, blast pressures in complex geometries will be measured using state-of-the-art high magnitude pressure gauges sampled at ~5 MHz to ensure a rich suite of data is gathered.