Authors: Riccardo Rossi, Raul Bravo, Sebastian Ares de Parga
In the last decades, the level of accuracy of numerical simulations of physical assets within the manufacturing industry has allowed their throughout study and improvement. It is a well-known fact, for example, the presence of computer aided design and engineering (CAD & CAE) in the first stages of product development, where the high computational cost associated to such simulations is not a major obstacle.
More recently, however, the tendency is for the usage of realistic simulations to extend to all stages of the manufacturing process (industry 4.0), through to the so-called Digital Twins, digital replicas of the physical assets, that optimise their efficiency and extend their life cycles. For Digital Twins, the required simulations are constrained by both computing power and speed of response, which prevents the direct application of traditional numerical simulation techniques. In contrast, surrogate models driven by Reduced Order Models ROMs, should be employed.
ROMs follow the classical machine learning (ML) model of training-inference. During the training step, an expensive campaign of simulations (known as Full Order Models or FOM) should be undertaken to gather the training data to be analyzed. Typical tools for this analysis are large-scale Singular Value Decomposition (SVD) techniques, which extract the most relevant patterns that allow for the construction of the target ROM. During the inference step, the ROM is used to obtain the solution to simulations at a fraction of the time and computational cost.
Within the Pillar I on Manufacturing and Digital Twins of the eFlows4HPC project, the challenges posed by the expensive training stage of ROMs will be addressed by making use of high performance computing HPC. In particular, since the computation of the SVD has been identified as the crucial kernel, a distributed-randomized SVD will be developed employing a task-oriented implementation on top of PyCOMPSs by making use of the dislib library. Moreover, the effective use of supercomputers requires integrating both the training and inference steps within a single complex workflow adaptable to the needs of the specific problem to be addressed. Such a workflow will also be capable of integrating with other Deep Learning (DL) frameworks, to use the ROMs as building blocks for the construction of system-level models.