The increasing demand for small batch production and high productivity is the driving force for the ongoing automation and flexibilization of manufacturing processes. Substituting conventional machine tools by industrial robots constitutes a significant step towards an adaptable production. Nowadays, industrial robots are already widely used in industry to automate pick-and-place, assembly or welding tasks. Advantages that contribute to the application of industrial robots are the large available workspace, high feed rates, six or more degrees of freedom and a generalized end-effector control. Such beneficial attributes motivate research and industry to widen the application range of industrial robots to machining operations. However, the accumulation of uncertainties along the serial-kinematic chain of industrial robots results in a low absolute positional accuracy of the tool center point. To overcome this downside, parameterized models of the industrial robot are derived, which consider the geometric, static and thermal behavior. Depending on the modeled effects, up to 100 unknown parameters have to be identified by suitable optimization methods. An increase in model complexity to capture nonlinearities and hysteresis behavior is limited due to the identifiability of the parameters and the requirement for real-time capability of the model. Therefore, the use of data-based regression models in the sense of machine learning is increasingly explored. For an efficient and safe use of these models, the prediction of probability distributions as well as a real-time parallel training is essential. Furthermore, it remains an open research question whether data-driven models alone represent the key technology for the substitution of machine tools by industrial robots or a combination with conventional approaches is beneficial.

The dynamo theory is on magnetic field generation by flows of an electrically conducting fluid. In this talk, I will present the first steps in the framework of an ongoing project that involves applications of the machine learning techniques to the problem of magnetic field generation. Physicists ask the following question: why does one flow of a conducting fluid generate a magnetic field better than the other? The mathematical answer to this question is because the real part of the dominant eigenvalue of a certain elliptic operator (where the flow enters as a parameter) is larger for one flow and smaller for the other. Unfortunately, the physicists find this answer unsatisfactory. In this project, we plan to find an alternative answer for this question. For a certain class of randomly generated steady flows, we construct a convolution neural network (CNN) trained to solve the kinematic dynamo problem -- the eigenvalue problem for the operator of magnetic induction. The data set for training and validation is generated by solving the eigenvalue problem on high performance computers using the pseudospectral methods. Our first goal is to find out if this problem can be solved by CNN with a reasonable precision. The second goal is, applying the 3D-Grad-CAM technique (producing "visual explanations" for decisions), identify and analyze the regions of the flow responsible for the decision. Time permitting, we will also check if the pre-trained CNNs implemented in Vertex AI (Google Cloud) are suitable for this problem. Some of the computations for this ongoing project are performed on the Oblivion supercomputer (hosted by the University of Evora, Portugal) in the framework of the computational projects 2021.09815.CPCA and 2022.15706.CPCA.A2 financed by the Foundation for Science and Technology (FCT), Portugal and supported by Google and FCT (ref. CPCA-IAC/AF/475871/2022) in the Call on Advanced Computing Projects: Artificial Intelligence in Cloud (1st edition).
Authors: S. Ranjith, R. Chertovskih and R.J.P. Gonçalves

Renal diseases affect thousands of patients who, to survive, must incur in dialysis -- a costly treatment with many negative implications in their quality of life. As an alternative, patients may enter a waiting list for receiving a kidney from a deceased donor; however, waiting times are typically very long. For reducing the waiting time, another alternative in some countries is to find a healthy living donor -- usually, a relative of a person emotionally connected -- who volunteers to cede one of their kidneys. However, in some situations transplantation is not possible due to blood, or tissue-level incompatibility. In these cases, a donor-patient pair may enter a pool of pairs in the same situation, in the hope of finding compatibility in crossed transplants. The problem has been studied under different perspectives, but the most commonly used objective is maximizing the number of patients in the pool for which a crossed transplant is possible. We propose to change this objective by that of maximizing the cumulative patient survival times. This model departs from the previous deterministic setting, putting into play a method for predicting survival time based on historical data.

The Industrie 4.0 era enables the digitalization of existing processes, where processes and assets are represented in the digital domain in ways not possible until now, e.g. through end-to-end data availability. At the same time, advances and ease of use in the field of artificial intelligence and machine learning, allow for the development and implementation of appropriate solutions to optimize existing manufacturing processes along the entire value chain. Holistic approaches for such optimizations require a reference architecture that includes data acquisition on the sensor level, edge computing and digital twin applications. For that purpose, condition-related information, process monitoring and digital assistance systems are integrated to develop application-specific digital twins based on the proposed architecture, integrating heterogenous data sources in order to enhance the accuracy of the machine learning models.

After a brief explanation of what point vortices and passive particles are — on the plane and on the sphere — and how they can mimic real world flows, we will discuss some optimization issues for passive particles advected by point vortices and methods for estimating point vortice circulation and position in the presence of Gaussian noise using passive particle trajectory data.