This paper is concerned with the description of the constitutive behaviour of geomaterials of wide ranging character using artificial neural networks (ANNs). The basic idea is to increase the possible options for selection of constitutive models for geotechnical analysis and design through adoption of a neural network representation of the geomaterial constitutive behaviour. First the context of ANNs in this role is presented, followed by a brief description of their operation. Then a general conceptual framework for the use of ANNs for geomaterial constitutive modelling is given and a case study demonstrating the use of the technique for a stiff overconsolidated clay (Vallericca Clay) is presented. Predictions of mechanical behaviour using the ANN are extremely accurate within the stress space explored in the experimental part of the research. The advantages and limits of applicability of the technique are finally discussed.
Ce papier discute la description du comportement constitutif de geomateriaux à caracterè tres variable, en utilisant des reseaux neroux artificiels (RNA = ANN). L'idee de base est d'augmenter les options disponibles pour la selection de modeles constitutif pour analyse geotechnique et dessein à travers I'adoption d'une representation par reseau neural du comportement constitutif du geomateriel. Premièrernent, Ie contexte des RNAs dans ce rôle est presente, suivi d'une brève description de leur operation. Une assature generale, et conceptuelle pour l'utilisation des RNAs pour modeller constitutivement des geomateriaux est donnee, et ainsi qu'une etude d'un cas demontrant l'untilisation de la technique pour une craie rigids trop consolidee (Vallericca Clay). Les predictions du comportement constitutif utilisant Ie RNA sont extremement precis dans I'espace explore pour Ie stress dans la part experiment ale de la recherche. Les avantages et limites sur I'application de la technique sont discutes.
Diese Arbeit handelt ueber die Beschreibung von dem konstitutiven Verhalten von Geomaterialien mit breit gefacherten Charakter unter der Benutzung von kuenstlichen neuralen Netzwerken (artificial neural networks = ANNs). Die Basisidee ist es die möglichen Option fuer die Auswahl von konstitutiven Modellen fur geotechnische Analyse und Design durch die Adoption von einer neuralen Netzwerkreprasentation von dem geomaterialen, mechanischen Verhalten zu verstarken. A1s erstes wird der Zusammenhang von ANNs in dieser Rolle prasentiert, gefolgt von einer kurzen Beschreibung ueber ihre Arbeitsweise Es wird ein generelles konzeptionelles Geruest fuer den Gebrauch von ANNs fur geomateriale konstitutive Modelle gegeben und anhand einer Fallstudie die Benutzung dieser Technik fuer einen steifen ueberkonsolidierten Lehm (Vallericca Clay) demonstriert. Vorhersagen ueber konstitutive Verhaltensweisen unter der Benutzung von ANNs sind ext rem akkurat innerhalb des Spannungbereiches in dem experimentellen Teil dieser Studie. Es werden weiter die Vorteile und Grenzen der Anwendbarkeit dieser Technik diskutiert.
Figure I schematically shows the relation between primary activities in rock and soil engineering. At the largest scale there is the engineering problem to be solved, that is, the design of the measures to be taken to attain the objective whether it be stabilisation of slopes, securing underground excavations, etc. This activity inevitably leads to some kind of engineering analysis, which involves the use of classification schemes, empirically based analyses, numerical methods, or stability assessment of the overall problem in limit states. Depending on the scope and complexity of the engineering analysis, it is often necessary to undertake a detailed investigation of the mechanical behaviour of the materials at hand, whether they be the natural geomaterials or manmade materials introduced, for example, as structural elements for the purpose of support or reinforcement. Artificial neural networks (ANNs), which are essentially optimisation techniques developed in the computer science discipline, are becoming a part of activities at all of these levels. One example of use at the upper level of the engineering problem, treated as a whole, is the prediction of debris flows occurring as a result of rainfall, where a time series of rain gauge measurements were correlated with debris flow events to produce a debris flow risk indicator (Hirano et al, 1994). Also, ANNs have been used to enhance interpretative aspects of rock mass classification approaches (Millar and Hudson, 1993), and have been employed as indicators of underground opening stability (Feng et at. 1993). This type of application of ANNs in geomechanics lies on the boundary between the level of engineering analysis and the core level which concerns with the mechanical behaviour of geomaterials.
