Physics-based reservoir simulation is the backbone of many decision-making processes in the oil and gas industry. However, due to being computationally demanding, simulating a model multiple times in iterative studies, such as history matching and production optimisation, is extremely time intensive. This downside results in it being practically impossible to update the model every time a set of new data are available. One of the popular solutions for this problem is creating a proxy model of the desired reservoir. However, the consequence of this approach is that such a proxy model can only represent one corresponding reservoir, and, for every new reservoir, a new proxy model must be rebuilt. Additionally, when the overall runtime is considered, it can be observed that, in some cases, iteratively running a numerical reservoir simulation may be quicker than the process of building, validating and using a proxy model. To overcome this obstacle, in this study, we used deep learning to create a data-driven simulator, deep net simulator (DNS), that enables us to simulate a wide range of reservoirs. Unlike the conventional proxy approach, we collected the training data from multiple reservoirs with completely different configurations and settings. We compared the precision and reliability of DNS with a commercial simulator for 600 generated case studies, consisting of 500,000,000 data points. DNS successfully predicts 45%, 70% and 90% of the cases with a mean absolute percentage error of less than 5%, 10% and 15%, respectively. Due to the indirect dependency of DNS on the initial and boundary conditions, DNS acts incredibly fast when compared with physics-based simulators. Our results showed that DNS is, on average, 9.25E+7 times faster than a commercial simulator.

基于物理的油藏模拟是石油和天然气行业许多决策过程的基础。但是，由于计算量大，因此在迭代研究（例如历史匹配和生产优化）中多次模拟模型非常耗时。不利的一面是，实际上每当有一组新数据可用时就无法更新模型。针对该问题的流行解决方案之一是创建所需储层的代理模型。但是，这种方法的结果是，这样的代理模型只能代表一个相应的储层，并且对于每个新的储层，都必须重建一个新的代理模型。此外，在考虑整体运行时时，可以观察到，在某些情况下，迭代运行数值油藏模拟可能比建立，验证和使用代理模型的过程要快。为了克服这一障碍，在本研究中，我们使用深度学习创建了一个数据驱动的模拟器，即深网模拟器（DNS），该模拟器使我们能够模拟各种储层。与传统的代理方法不同，我们从配置和设置完全不同的多个油藏中收集了训练数据。我们将DNS的精度和可靠性与商用模拟器进行了比较，生成了600个案例研究，其中包括500,000,000个数据点。DNS成功地预测了45％，70％和90％的情况，平均绝对百分比误差分别小于5％，10％和15％。由于DNS对初始条件和边界条件的间接依赖性，与基于物理的模拟器相比，DNS的运行速度非常快。我们的结果表明，DNS平均比商用模拟器快9.25E + 7倍。