La complexité des applications numériques de calcul scientifique ne cesse de croître. Cette difficulté revêt alors deux formes. La première est une complexité logicielle qui nécessite l'intégration de divers codes de calcul toujours plus sophistiqués et spécialisés à la simulation de phénomènes physiques complexes. La seconde forme de complexité est calculatoire où les composants de calcul nécessitent toujours plus de ressources et de capacité de stockage afin de modéliser, entre autres, les phénomènes au plus proche de la physique « réelle ».<br /><br />Dans le domaine de la simulation des procédés assistée par ordinateur, la complexité logicielle est masquée par le standard CAPE-OPEN qui répond aux besoins d'intégration de codes tiers. Il propose des spécifications d'interfaces, basées sur une approche par composants logiciels tels que DCOM ou CORBA. Cette thèse apporte une solution à la complexité calculatoire. Pour cela, nous étudions le problème de la distribution de la charge de calcul des simulations des procédés sur des architectures de type grappe de calcul dont les composants sont au standard CAPE-OPEN. Une exécution distribuée performante requiert la distribution des activités concurrentes de l'application tout en minimisant le volume de données à échanger via le support de communication.<br /><br />Dans ce contexte, nous présentons une analyse fine du schéma d'exécution des simulations de procédés qui conduit à la conception de deux environnements distribués d'exécution. Le premier nous a permis de quantifier le gain atteignable sur une grappe de calcul à travers la simulation de plusieurs cas tests métier. Toutefois, les contraintes technologiques industrielles se sont avérées peu propices à l'implémentation d'un environnement distribué visant à s'approcher de l'optimal. Par conséquent, la définition d'un second prototype basé sur le moteur exécutif KAAPI a été mené à bien. Afin de répondre à nos besoins, nous l'avons étendu aux techniques « statiques » d'exécution. Fort de cet environnement, nous avons entrepris d'étudier différentes politiques d'ordonnancement. L'environnement KAAPI couplé à notre extension ouvre de larges perspectives d'études dans le cadre plus large des applications numériques de calcul scientifique.

This work presents the first known attempt to model the dairy business from a multiscale modelling perspective. The multiscale nature of the dairy industry is examined with emphasis on those key decision making and process scales involved in production. Decision making scales identified range from the investor level to the plant operator level, and encompass business, production, plant, and operational levels. The model considers scales from the production manager to the unit operation scale. The cheese making process is used to demonstrate scale identification in the context of the important phenomena and other natural levels of scrutiny of interest to decision makers. This work was a first step in the establishment of a multiscale system model capable of delivering information for process troubleshooting, scheduling, process and business optimization, and process control decision-making for the dairy industry. Here, only material transfer throughout a process, use of raw materials, and production of manufactured product is modelled. However, an implementation pathway for adding other models (such as the precipitation of milk protein which forms curd) to the system model is proposed. The software implementation of the dairy industry multiscale model presented here tests the validity of the proposed: • object model (object and collection classes) used to model unit operations and integrate them into a process, • mechanisms for modelling material and energy streams, • method to create simulations over variable time horizons. The model was implemented using object oriented programming (OOP) methods in conjunction with technologies such as Visual Basic .NET and CAPE-OPEN. An OOP object model is presented which successfully enabled the construction of a multiscale model of the cheese making process. Material content, unit operation, and raw milk supply models were integrated into the multiscale model. The model is capable of performing simulations over variable time horizons, from 1 second, to multiple years. Mechanisms for modelling material streams, connecting unit operations, and controlling unit operation behaviour were implemented. Simple unit operations such as pumps and storage silos along with more complex unit operations, such as a cheese vat batch, were modelled. Despite some simplifications to the model of the cheese making process, the simulations successfully reproduced the major features expected from the process and its constituent unit operations. Decision making information for process operators, plant managers, production managers, and the dairy business manager can be produced from the data generated. The multiscale model can be made more sophisticated by extending the functionality of existing objects, and incorporating other scale partial models. However, increasing the number of reported variables by even a small number can quickly increase the data processing and storage demands of the model. A unit operation’s operational state of existence at any point of time was proposed as a mechanism for integrating and recalculating lower scale partial models. This mechanism was successfully tested using a unit operation’s material content model and is presented here as a new concept in multiscale modelling. The proposed modelling structure can be extended to include any number of partial models and any number of scales.