Please note that the SPE VBS presentation on May 3rd 2016, can unfortunately not be followed remotely on Lync. OMV informed us that this is due to technical reasons with several links in the presentation. We are very sorry that we have to cancel this event!

Artificial neural networks are a family of models inspired by biological systems and can be considered as a type of artificial intelligence that attempts to imitate the way a human brain works. In contrast to deterministic models, in which all necessary computations are available by well-known modus operandi like mathematical, physical or other models, neural networks do not need such information since they have the capability to learn. They can learn relations in-between data, they can learn how to make optimal decisions and they can extract human knowledge from data. That learing is – although strongly simplified – similar to human learning. It works by creating and optimising artificial synaptic connections between the processing elements – the computer equivalent of biological neurons. Synapses are modified throughout a training phase by some sophisticated rules until the actual outpur of the network satisfies the demands. Once a network is trained it can be applied to data it never has seen before. The advantage of suchlike models is that – independent of the challenge – no formal description of the background mechanisms is necessary and solutions can be obtained anyway. The disadvantage on the other hand is that no information about such mechanisms is provided, neural networks are heuristic model builders and do not provide information on how a model is designed.

The purpose of this presentation is to provide an overview about the different types of neural networks and how they are successfully applied. Included are the training paradigms supervised and unsupervised learning, when and for which types of networks they are applied. Reinforcement learning will be shortly addressed. The talk will also compare feed forward to recurrent networks and which challenges – especially in the oil business – can be addressed with such tools. Since the training of neural networks is always targeted on good generalisation properties, some general rules are discussed for improving the chance that a network performs well on data it never has seen before. Several real world examples are presented.

Dr. Rudolf K. Fruhwirth is graduated Petroleum Engineer of the University of Leoben. At present he study wellbore heat transfer models at the Chair of Petroleum and Geothermal Energy Recovery. After finishing his PhD in 1986 he started as a geophysicist and run through several organisations during his professional career including the Austrian Academy of Science. He has gained more than 30 years of research experience in the fields of oil-, gas & groundwater exploration as well as in data analysis & computational intelligence. Rudolf is author of numerous papers and associate lecturer at the University of Leoben.