Summary Compared to clastic reservoirs, volcanic reservoirs exhibit higher heterogeneity. Lithological facies type is one of the most important indicators of favorable volcanic reservoirs. Traditionally, facies are identified by core observation or log classification. However, spatial-distribution characteristics and geological conceptual models, which are important in the early stages of exploration, are seldom incorporated quantitatively in facies prediction. Based on previous work, a new method has been developed to incorporate volcanic spatial information with limited well data (three wells) to improve facies prediction. This method was applied to a volcanic clastic reservoir of the Cretaceous Yingchen member of the Xinshan fault depression, northeastern China. For better well control, an artificial neural network (ANN), a beta-Bayesian method (BBM), and a discriminant analysis (DA) algorithm, were used to predict log-based facies. Confidence analysis was applied to evaluate the log facies prediction. Analysis of variance (ANOVA) verifies that the overall prediction accuracy is above 82%. Indicator kriging was used to estimate the conditional probabilities of facies occurrence given residual thickness. This is based on the assumption that the residual thickness of the volcanic formation is controlled by distance from the eruption center, a major factor defining the geological facies. The geological conceptual models (areal sedimentary facies maps and diagenetic facies maps) were converted into the conditional probability of facies occurrence in given geological settings using multinomial logistic regression. These conditional probabilities were combined with well-log facies data within a Bayesian framework. Three favorable reservoirs were predicted based on the method above, and the predictions were proved by the subsequent drilling. Introduction Volcanic reservoir quality is controlled by both lithofacies and diagenetic effects. Traditionally, these effects are qualified by core observation and well-log (especially image-log) interpretation. The spatial distribution of reservoirs is characterized by high-resolution seismic interpretation. Even though the importance of these features is well known among the geological community, it is difficult to quantify and integrate these data into reservoir modeling and flow simulation. The area of study is in the early Cretaceous (Yingchen) volcanic formation, Xinshan fault depression, Songliao basin, northeastern China. Several factors have made the traditional approach less practical. First, the available well data are limited (three wells were drilled in this area). Second, the volcanic formation is deeply buried (3000 to 6000 m) in the Xinshan area. It has a high seismic amplitude contrast with bounding sedimentary formations but low intralayer reflection. Seismic properties appear homogeneous for most volcanic formations (Zhao 1999); furthermore, compared to a clastic reservoir, the spatial distribution of volcanic reservoirs is less continuous. Reservoir quality is highly heterogeneous because of complex lithology, facies and diagenetic overprints. Detailed characterization is, thus, more difficult to conduct. There are two major challenges:to quantify conceptual geological models and integrate them with other types of data (e.g., seismic and log data) andto accurately identify volcanic lithofacies with limited well-log data. To incorporate diverse data into lithofacies prediction, reservoir-quality assessment, and uncertainty reduction, two types of methods are frequently used: statistical (geostatistics) methods and ANN methods. Geostatistics methods such as cokriging or indicator cokriging (Goovaerts 1998; Deutsch and Journel 1998; Yarus and Chambers 1994) are versatile and widely used. However, these methods require cross-variogram models for all indicators, which are tedious and time-consuming to construct. In addition, this approach does not guarantee better results (Deutsch and Journel 1998). Alternatively, object-based simulation, such as Boolean simulation (Yarus and Chambers 1994), can be used to integrate geological models into reservoir modeling and reproduce the exact geometry of the facies model. It requires good geometric parameters, which can be estimated from outcrop analogs or detailed seismic interpretation. However, those parameters are seldom available and are less representative in this fault-complicated area. ANN methods are powerful for integrating high-dimensional data and expressing complex, nonlinear relationships between input and output. Several successful applications using neural networks for data incorporation and facies prediction have been reported (Wong et al. 1995; Siripitayananon et al. 2001; Bhatt and Helle 2002). However, the computation cost of ANNs is high, and they do not provide any estimate of uncertainties.