Affiliation:
1. Caliche Private Limited
Abstract
Abstract
Chemical enhanced oil recovery involves enormous combinations of chemicals, surfactants, etc. The reservoir properties such as temperature, capillary pressure, permeability keeps changing, making the process of identification of suitable chemicals even more challenging. Data driven modelling holds solutions for the complexity involved in identification of most suitable parameters for chemical enhanced oil recovery. Over the last decade, Artificial Intelligence has found its numerous applications in different branches of chemistry. From the selection of chemicals to the operating conditions during synthesis all can be estimated by the use of deep learning models. This paper presents yield prediction which is of high economic significance for chemical enhanced oil recovery, because they enable calculation of investment versus return.
These models give us the conversion of reaction into products before performing the lab experiment. This will help chemists in selecting high performance chemicals for specific reservoirs without spending time on costly iterative chemical processes. These models require application of deep learning architectures like transformers and natural language processing techniques like tokenization for the prediction task. Encoder models like BERT are used for receiving the information on chemical reactions in text-based form for a reaction which is then combined with a regression extension layer to give us the desired reaction yield.
We demonstrate our model on a HTE dataset with an excellent prediction score. Efforts are also made on the USPTO patent dataset which covers a wide variety of chemical reaction space. The USPTO patent dataset consists of almost every chemical reaction published since late 1970s till 2006. Diverse techniques starting with Multi Layer Perceptrons, Sequence to sequence modelling, Long short term memory models and finally transformers are employed for the improvement of accuracy of patent reactions. The paper presents detailed comparative results of predicting chemical reaction yield, and the enhancements that it will bring to Chemical Enhanced Oil Recovery.
Reaction yield prediction models receive very little attention in spite of their enormous potential of determining the reaction conversion rates and its contribution to chemical enhanced oil recovery processes . The paper introduces a novel approach of modelling chemical reaction yield with deep learning models to the petroleum community. Unprecedented result of accuracy beyond 90% in predicting chemical reactions yield and its significance in chemical enhanced oil recovery has been proposed in the paper.