Search

Presentation of AI-assisted strategiesfor high-dimensional optimisation in turbomachinery at the CM3 Conference in Barcelona (by Diego Lopez)

The increased need to design higher performing aerodynamic shapes has led to design optimization cycles requiring high-fidelity CFD models and high-dimensional parametrizationschemes. The computational cost of employing global search algorithms on such scenarioshas typically been prohibitive for most academic and industrial environments. In this paper,a novel strategy is presented that leverages the capabilities of artificial neural networks forregressing complex unstructured data, while coupling them with dimensionality reductionalgorithms. This approach enables employing global-based optimization methods on high-dimensional applications through a reduced computational cost. This methodologyis demonstrated on the efficiency optimization of a modern jet engine fan blade with con- strained pressure ratio. The outcome is compared against a state-of-the-art adjoint-based approach. Results indicate that the strategy proposed achieves comparable improvements to its adjoint counterpart with a reduced computational cost and can scale better to multi-objective optimization applications.

Wildland fire models and simulators developed in the last two decades are increasingly applied in different ecosystems and countries of the world to predict fire behaviour and effects. Fire models range from empirical formulas, such as the ones defined by Rothermel [1] and applied in spatially and temporally explicit fire simulators (i.e. Farsite), to complex three-dimensional CFD approaches solving the partial differential equation of continuity, momentum and energy. The wide range of length and time scales governing wildland fire (from the millimeter scale of combustion processes to the hundreds of meters scale of synoptic wind flow) complicates the use of a full-3D numerical approach, at least for operational forecasting purposes. At the same time, there is an important two-way influence between weather and fire: wind determines fire propagation and, conversely, the buoyancy effects generated by fire heat modify the local wind field, -creating their own weather?. A possible solution to this problem is the use of a simpler (and thus computationally cheaper) model to describe fire propagation, while maintaining a CFD approach to model wind behaviour and, more importantly, the two-way interaction due to fire heat release. A number of studies applied this approach in the last few years. This work describes the initial steps in the development of a model for fire-front propagation based on a level-set methodology and its integration into a CFD model.

A number of studies have reported on the application of computational fluid dynamics (CFD) models to produce high resolution data of wind flow. CFD models are based on the numerical solution of a number of partial differential equations expressing conservation of mass, momentum and energy (when buoyancy effects are not negligible) and make use of a number of simplifications (turbulence models) to reduce the range of length scales which are directly resolved. Most of these works validate the proposed approaches in simplified domains, where separated flows and recirculation regions are limited. In addition, application of CFD models requires a precise specification of both simulation domain characteristics and boundary conditions (in terms of flow direction and intensity); unfortunately, especially in areas characterised by complex terrains, measured data of wind intensity and direction collected close to the limits of the computational domain (fundamental to prescribe appropriate boundary conditions) are often not available. The aims of this work were (1) to develop a methodology for simulating wind flow in domains with complex orography, by matching data within the domain, (2) to compare wind fields predicted by different commercial or open-source software and (3) to validate the simulation results on a set of experimental data collected in different areas of the computational domain. More specifically, CFD simulations were linked to an automatic optimisation system developed by the authors based on a pattern search algorithm, which attempts to minimise two functions (deviation of wind speed and direction) at a number of prescribed locations within the domain where experimental data are available. A kriging approximation was used to accelerate convergence to a minimum. The study demonstrates the capabilities of the proposed approach in simulating wind flow in areas characterised by a complex topography and uncertainty in boundary conditions.

According to recent projections on future climate, an increase in risk of summer droughts in both Central and Southern Europe is expected. The most pronounced warming is likely to take place in the Mediterranean area in summer when the highest temperatures occur. In addition, annual precipitation and the annual number of rainy days are very likely to decrease in most of the Mediterranean Basin. More prolonged drought seasons induced by climatic changes are likely to influence general flammability characteristics of fuel, affecting load distribution in vegetation strata, floristic composition, and live and dead fuel ratio. In addition, variations in precipitation and mean temperature could directly affect the fuel water status, and consequently flammability, and length of critical periods of high ignition danger for Mediterranean ecosystems. Mediterranean shrub species are an important component of both Mediterranean vegetation communities and understorey vegetation. They also constitute the surface fuels primarily responsible for the ignition and the spread of wildland fires in Mediterranean forests. Although fire spread and behaviour are dependent on several factors, the water content of live fuel plays a crucial role in determining wildland fire danger, especially in the Mediterranean shrubland, where live fuel is often the main component of the available fuel which catches fire. Several studies have proved the relation between live fuel moisture content (FMC) and flammability and have verified the relations between live FMC and actual fire occurrence. Considering the observed climatic variations and foreseen future scenarios, an evaluation of the impact of these variations on fire danger seems essential. Therefore, the main aims of this work are i) to identify useful tools to determine possible impacts of expected climate change on fuel status in Mediterranean shrublands, and ii) to predict the potential impacts of climate change on the length of fire danger season due to live fuel status. The study was carried out in Sardinia (Italy). Moisture content of live fuel was determined periodically during five years on four shrub species (Cistus monspeliensis L., Juniperus phoenicea L, Pistacia lentiscus L., and Rosmarinus officinalis L.). Meteorological variables were also recorded. Data sets were used to determine the accuracy of meteorological drought codes in estimating live fuel moisture patterns in Mediterranean areas. Threshold values of codes for estimating the starting and ending dates of high ignition danger periods were estimated by percentile and logistic analysis. Finally, a methodology to identify the length of critical periods of high ignitability risk in maquis ecosystems was proposed. Potential climate change impact on length of ignition danger season in Mediterranean Basin shrublands was determined using the meteorological code values and future climate scenarios at a local scale derived from an advanced high resolution Regional Climate Model. Starting and ending dates of danger season were calculated for the period May-October using data from the Regional Climate Model (1961 - 1990: baseline; 2071 - 2100 A1B scenario). Changes in length of fire danger season due to live fuel water content were finally analysed. Results suggest that future climate projections could determine changes in fuel moisture dynamics and affect the duration of the danger fire season causing a delay in the ending dates of high ignition danger periods.

Wildland fire models and simulators developed in the last twenty years are increasingly applied in different ecosystems and countries of the world in order to predict fire behaviour and effects, and to support fire prevention and management programs. In addition to the simplifications and preliminary assumptions of fire spread models, which are mainly based on a semi-physical approach, criticisms of fire simulators frequently concern the need of high resolution environmental data (fuel model maps, canopy cover, etc.), and in particular data on environmental variables that quickly change in space and time (e.g. wind speed and direction). The rapid increase in computational resources and numerical methods of the last years helps to provide high temporal and spatial resolution data on wind fields in order to improve the accuracy of simulations at both local and regional scale. Several studies report on the application of computational fluid dynamics (CFD) models to produce high resolution data of wind flow. CFD models are based on the numerical solution of a number of partial differential equations expressing conservation of mass, momentum and energy (when buoyancy effects are not negligible) and make use of a number of simplifications (turbulence models) to reduce the range of length scales which are directly resolved. Most of these works validate the proposed approaches in simplified domains, where separated flows and circulation regions are limited. In addition, application of CFD models requires a precise specification of both simulation domain characteristics and boundary conditions (in terms of flow direction and intensity); unfortunately, especially in areas characterized by complex terrains, measured data of wind intensity and direction collected close to the limits of the computational domain (fundamental to prescribe appropriate boundary conditions) are often not available. The aims of this work were (1) to develop a methodology for simulating wind flow in domains with complex orography, by matching data within the domain, (2) to compare wind fields predicted by different commercial or open-source software and (3) to validate the simulation results on a set of experimental data collected in different areas of the computational domain.

Level-set approaches are efficient and versatile methods for solving interface tracking problems and have been used in recent years to describe wildland fire propagation. Being based on an Eulerian description of the spread problem, their numerical implementation offers improved computational agility and better portability to parallel computing environments with respect to vector-based simulators. The use of a continuous representation of the fire perimeter in place of the binary formulation used in Cellular Automata avoids the commonly observed distortion of the fire shape. This work presents an algorithm for fire-spread simulation based on a level-set formulation. The results are compared to the ones obtained by two well-known Cellular Automata simulators under homogeneous conditions, and to the ones given by a well-know vector-based fire-spread simulator under realistic slope and wind conditions. According to this work, the level-set approach provides better results, in terms of accuracy, at a much reduced computational cost.