The rapid virtual optimization of industrial processes is at the core of the Industry 4.0 paradigm but it requires the development of computationally efficient methodologies that go beyond a purely simulation driven approach to include machine learning and high-performance computing. This is particularly true in the context of industrial bulk solids handling processes, where the computational expense of high-fidelity simulations is significant. In this work we present an efficient virtual optimization methodology for industrial bulk handling processes, which combines discrete element modelling (DEM), design of experiments, machine learning and optimization methods in the Altair portfolio of tools to significantly reduce the computational expense of optimization relative to a purely simulation driven approach. A key feature of the methodology is the generation of multi-layer perceptron model that constitutes a state-space representation of a dynamic system of interest. Data from a statistically efficient set of high-fidelity discrete element simulations is used to train the model, which can be thought of as Reduced Order Model (ROM) of the system with respect to a set of responses. Because of its high computational efficiency, the ROM can then be used to speed-up system design and optimization or serve as a component of Digital Twins in real-time and control applications. The validity of the methodology is evaluated through the optimization of an industrial scale bin blending system, which is a commonly used unit mixing system in the pharmaceutical and food industries. Batch failures are a common and persistent problem in this type of system, but its physical prototyping is expensive, making it an excellent candidate for virtual optimization. The work focuses on improving the mixing rate in the system by optimizing operational parameters such as the rotational velocity, the level of fill and the bin orientation. The accuracy of the predicted optimal operational parameter set is evaluated in-silico and the advantages and limitations of the methodology are discussed.