Synergies between numerical modelling and machine learning in geotechnical engineering

深度神经网络的训练需要包含数千个时间序列或图像的大型数据集。每个数据集条目必须被标记。例如，要用卷积神经网络从照片中估计土壤颗粒的大小，数据集必须包含数千张具有相应颗粒大小的图像。与生成这些大型数据集相关的成本往往令人望而却步，特别是在岩土工程等专业领域。我们最近在颗粒材料侵蚀的数值模拟和从照片中确定粒度分布（PSD）方面的工作突出了使用有限元或离散元模型创建用于机器学习的大型合成数据集的潜力。我们研究计划的长期目标是创建新的工作流程，将数值建模和深度学习相结合，用于开发岩土工程中的智能测试和传感方法。我们的第一个短期目标是通过更真实的数据和随机化来改善领域转移。领域转移问题是阻碍基于合成数据的网络训练的主要障碍，它是由合成数据的理想化特性造成的。研究计划的第二个目标是使用数值方法生成具有物理意义的次级神经网络输入。我们将用三个实际问题来应用我们的方法。第一个问题是从照片中确定颗粒材料的PSD。准备不同真实感水平的合成图像，用卷积神经网络预测PSD。第二个问题涉及仪器数据的分析。采用有限元法对大坝和隧道建立了与实际破坏机制相对应的合成孔隙压力时间序列。最后一个例子是关于尚普兰粘土沉降的预测。将利用遥感编制住区地图。不同层次的信息（土地利用、地表地质、地下水位）将与一系列简单有限元模型的输出相结合，以预测聚落。领域转移问题对于所有涉及合成数据的应用程序都很重要。选择的三个问题都是重要的岩土工程应用，可以从机器学习解决方案中受益。PSD测定在许多工程领域（例如粉末技术）中很常见，并且仍然经常使用筛子进行。现在生成和存储大量仪器仪表数据相对容易。由于缺乏分析工具，这些数据很少被充分利用。最后，在气候变化的背景下，在敏感粘土上定居是一个重大问题，已经引起了加拿大东部媒体的注意。这个研究项目也将为培养高素质的岩土工程师提供机会。

The training of deep neural networks requires large datasets with thousands of time series or images. Each dataset entry must be labelled. For example, to estimate the size of soil particles from photographs with a convolutional neural network, the dataset must include thousands of images with corresponding particle sizes. The costs associated with the generation of these large datasets are often prohibitive, especially in specialized fields such as geotechnical engineering. Our recent work on the numerical modelling of erosion in granular materials and the determination of particle size distributions (PSD) from photographs has highlighted the potential of using finite-element or discrete-element models for the creation of large synthetic datasets for machine learning. The long-term objective of our research program is to create new workflows that combine numerical modelling and deep learning for the development of intelligent testing and sensing methods in geotechnical engineering. Our first short-term objective is to improve domain transfer through more realistic data and randomization. Domain transfer issues, the main barriers preventing network training based on synthetic data, are caused by the idealized nature of synthetic data. The second objective of the research program is to use numerical methods to generate secondary neural network inputs with physical meaning. Three practical problems will be used to apply our methods. The first problem is the determination of the PSD of granular materials from photographs. Synthetic images with different levels of realism will be prepared to predict the PSD with convolutional neural networks. The second problem concerns the analysis of instrumentation data. Synthetic pore pressure time series corresponding to realistic failure mechanisms will be created with the finite element method for dams and tunnels. The last example concerns the prediction of settlements for Champlain clays. Settlements maps will be prepared using remote sensing. Different layers of information (land use, surface geology, water table depth) will be combined with the output of a series of simple finite element models to predict settlements. Domain transfer problems are important for all applications involving synthetic data. The three problems that were chosen are important geotechnical applications that would benefit from machine learning solutions. PSD determinations are common in many fields of engineering (e.g. powder technology) and still often conducted with sieves. It is now relatively easy to generate and to store large amount of instrumentation data. These data are seldom used to their full potential because of a lack of analysis tools. Finally, in the context of climate changes, settlements in sensitive clays are a significant problem that is already attracting media attention in Eastern Canada. This research program will also provide an opportunity for the training of highly qualified geotechnical engineers.