NATO STO

High-fidelity computer simulations are having an increased importance in design decisions for marine, ground, and air vehicles. As the sophistication of the models grows, so does the need for experimental data to develop new models and validate existing ones. Therefore, modern experiments have to monitor multiple measurands simultaneously and in a distributed manner. Furthermore, next generation computational models are being developed based on data assimilation instead of the traditional regression approach and require big data to train the algorithms. With the advancement of low-cost sensors and multi-sensor data fusion algorithms developed over the years by military (e.g. for situational awareness on the battlefield) and consumer industry (e.g. self-driving cars), one is on the verge of a paradigm shift in scientific instrumentation and computations. Traditionally, laboratory scales measurements rely on a single technique and, therefore, single physical process (temperature, velocity, concentration, etc.), resulting in idealized canonical experiments where real world effects are not present with very specialized instruments (i.e. expensive and complex because produced in small quantities). Network of multiple inexpensive sensors integrated with advanced algorithms could enable to bridge several gaps in traditional scientific instrumentation and enable data mining. The recent development and application of machine learning, surrogate modelling, and optimization methods for a variety of problems in scientific modelling and computing (including fusion of multi-fidelity models, solvers, data) offer a unique opportunity for a further advancement of the effective use of low-cost instrumentation and multi-sensor data. Novel paradigms for the optimal choice of instrumentation and sensor deployment may be considered, based on computer simulations and available experimental data, offering the possibility of pursuing both static/a priori and time-dependent adaptive approaches, depending on the application. Machine learning and optimization procedures may be integrated in a virtual/surrogate sensing framework with the multiple aim of reducing costs, improving accuracy, and enhancing robustness through sensor redundancy.