Young H. Cho

Methane emissions from plugged and unplugged abandoned wells dilute rapidly with air, causing conventional detection methods to fall short of detecting methane leaks from these sources. Commonly used technologies for methane emission detection often fall short of the detection limit required for low-rate emissions from abandoned wells, and are either dependent on environmental conditions or costly and highly energy consuming, barring them from becoming a scalable solution to this problem. We introduce here a new system for outdoor testing and show how the flux chamber detection limit is progressively reduced from 700 g/h to 2 g/h and ultimately to 1 g/h, meeting the US Department of the Interior (DOI) standard for monitoring equipment used on abandoned wells. Field deployment on an actual abandoned well also revealed intermittent emissions, which may serve as an indicator of deteriorating well integrity over time when monitored periodically. To forecast the emission event timing and intensity, a Liquid Time-Constant (LTC) and a gated recurrent neural network were trained on methane concentration time series collected during field deployment. As available well sites for physical testing are limited and atmospheric conditions are not controllable, a computational fluid dynamics (CFD) simulation framework integrated with machine learning (ML) was developed to optimize wellhead chamber geometry and size for both detectability and safety. This chamber is designed to be used for testing well sites in mass number that helps well abandonment ranking systems for both operators and regulatory bodies.

Methane emissions from plugged and unplugged abandoned wells dilute rapidly with air, causing conventional detection methods to underestimate methane leaks. We introduce here a new method for outdoor testing and show how the flux chamber detection limit is progressively reduced from 700 g/h to 2 g/h and ultimately to 1 g/h, meeting the US Department of the Interior (DOI) standard for monitoring equipment used on abandoned wells. Field deployment on an actual abandoned well also revealed intermittent emissions, which may serve as an indicator of deteriorating well integrity over time when monitored periodically. To forecast the emission event timing and intensity, a Liquid Time-Constant (LTC) and a gated recurrent neural network were trained on methane concentration time series collected during field deployment. As available well sites for physical testing are limited and atmospheric conditions are not controllable, a computational fluid dynamics (CFD) simulation framework integrated with machine learning (ML) was developed to optimize wellhead chamber geometry and size for both detectability and safety.

Summary: Data-driven models, such as neural networks, provide an alternative to physics-based simulations in predicting well behavior within unconventional reservoirs. However, these models struggle to interpret the factors behind production performance, the underlying physics of the process, and correlations unique to individual wells and their flow behaviors. These models are designed to approximate general trends present in data sets, which can hinder their ability to learn behaviors specific to certain data segments. In large unconventional fields, wells within a single formation can exhibit varied production behaviors, emphasizing the need to segment the data into more relevant subsets. Such data segmentation enables the development of local models (LMs) that capture regional or behavioral correlations. In contrast, a field-wide (global) AI model, trained on the entirety of the data, tends to produce averaged predictions, which can dilute specific correlations and reduce the overall predictive power. This work uses a transformer neural network to enable a behavior-based clustering process in predicting production from unconventional fields. By leveraging the attention mechanism inherent in transformers, we extract attention scores that link well properties to production behaviors. The information from these attention scores improves the clustering quality, which is crucial for training data-driven models on segmented data sets. The proposed approach highlights the advantages of developing behavior-centric (local) models that provide higher accuracy by capturing correlations distinct to a particular region or production behavior. Notably, the attention mechanism facilitates the isolation and curation of data subsets for training behavior-centric models, thereby improving prediction performance.

Summary: Neural network predictive models are popular for production forecasting in unconventional reservoirs due to their ability to learn complex relationships between well properties and production responses from extensive field data. The intricate flow behavior in hydraulically fractured unconventional reservoirs, which remains poorly understood, makes these statistical models particularly useful. Various neural network variants have been developed for production prediction in these reservoirs, each offering predictive capability of varying levels of granularity, accuracy, and robustness against noisy and incomplete data. Neural network predictive models that integrate physical principles are especially useful for subsurface systems, as they provide predictions that adhere to physical laws. This work introduces a new dynamic physics-guided deep learning (DPGDL) model that incorporates physical functions into neural networks and employs residual learning to compensate for the imperfect description of the physics, under variable data support. The new formulation allows for dynamic residual correction, avoids unintended bias due to less-than-ideal input data, and provides robust long-term predictions. The DPGDL model improves upon a static formulation by utilizing a masked loss function to enable learning from wells with varying production lengths and by improving the results when partially-observed timesteps are present. In addition, a sequence-to-sequence residual model has been developed to correct additional biases in the long-term predictions from the physics-constrained neural networks. Several synthetic data sets with increasing complexity as well as a field data set from the Bakken are used to demonstrate the performance of the new DPGDL model.

Summary: The complexity of physics-based modeling of fluid flow in hydraulically fractured unconventional reservoirs, together with the abundant data from repeated factory-style drilling and completion of these resources, has prompted the development and application of data-driven statistical models for predicting hydrocarbon production performance. More recently, machine learning algorithms have been widely studied in developing data-driven prediction models for unconventional reservoirs. These models often require a large amount of high-quality training data with sufficient range to avoid excessive extrapolation and produce reliable predictions. Unlike statistical models, physics-based models represent causal relations between input and output variables to provide predictions beyond available data. While a detailed physics-based description of fluid flow in unconventional reservoirs is not yet available, approximate physical flow functions have been proposed to capture the general production behavior of unconventional wells. These physical functions can be augmented with the available data to improve data-driven methods by constraining the models to adhere to the general production trends. In this paper, we develop a physics-constrained data-driven model by embedding physical flow functions into neural network models. Since the performance of the physics-constrained model depends on the relevance of the embedded physics, a sizeable residual prediction error is expected if the collected data do not sufficiently match the embedded model. The residual model typically represents errors in the description of inputs or any missing physical phenomenon. We compensate for such errors through residual learning, where an auxiliary neural network is designed to learn the complex relationship between the input parameters (such as formation and completion properties) and the expected prediction residuals. The new physics-guided deep learning (PGDL) model augments any physics-constrained prediction model with residual learning to increase its prediction accuracy. Several synthetic and actual field data sets, from the Bakken play, are used to demonstrate the performance of the PGDL model.

Summary: Constructing reliable data-driven models to predict well production performance (e.g., estimated ultimate recovery, cumulative production, production curves, etc.) for unconventional reservoirs requires large amounts of data. However, when considering unconventional reservoirs in their early stages of development, where data and the wells drilled are limited, one may benefit from leveraging available data and/or pretrained models from other more developed fields. Transfer learning, the process of storing knowledge gained while solving one problem (source data) and applying it to solve a different but related problem (target data), provides a workflow for alleviating data needs in training a data-driven model in fields with limited data. However, a pitfall in the application of transfer learning is the possibility of negative transfer, that is, transferring incorrect or irrelevant knowledge to the target data. In particular, the black-box nature of most data-driven models, e.g., neural networks, support vector machines, and random forest, makes it difficult to completely interpret the contribution of different source models used for knowledge transfer. Hence, ranking the viability of source models for transfer learning can reduce the risk of negative transfer and improve the prediction performance. In this paper, we illustrate the impact of negative transfer and how it can be identified, and present a new approach for ranking multiple source models based on their positive transfer contribution. Finally, we propose a framework to build a reliable model to predict well production performance by combining multiple sources of information into one network to be transferred and retrained with limited data in fields at their early stages of development.

Today's CMOS technologies allow larger circuit designs to fit on a single chip. However, this advantage comes at a high price of increased process-voltage-temperature (PVT) variations. FPGAs and their designs are no exceptions to such variations. In fact, the same bit file loaded into two different FPGAs of the same model can produce a significant difference in power and thermal characteristics due to variations that exist within the chip. Since it is increasingly difficult to control physical variations through manufacturing tasks, there is a need for practical ways to sense chip variations to provide a way for circuit designers to compensate or avoid its negative effects. One of the most critical aspects of such variation is power. Therefore, we developed and demonstrated a high accuracy on-chip on-line Energy-per-Component (EPC) measurement technology on Xilinx FPGAs since 2011. However, we found that the hardware overhead associated with such method limited the use of the technology. Therefore, our follow-up work in Energy-per-Operation (EPO) on Spartan FPGA with OpenRISC SoC produced an equally accurate power monitoring technology with drastically lower hardware overhead. While this method made our technology more practical for SoC designs on FPGAs, it did not produce component level power dissipation data that previous EPC method provided. Therefore, we extend this prior work with a new algorithm to extract EPC values from EPO result. Despite the lower hardware overhead, this change ended up improving the accuracy of the power result by unraveling the instruction-level abstraction into component-level energy consumption.

Despite scalable computing resource facilities and industry's ability to accurately model individual components with simulations, fine-grained physical level power simulation for today's typical digital design is found to be too expensive, especially, when parametric variations need to be accounted for modern processes. This paper presents a new linear solver based on-line modeling method that exploit existing slow but accurate simulators to obtain highly accurate sub-circuit power estimates while using significantly less processing and memory resources. Comparative simulation runs for a number of benchmark circuits using industry standard SPICE simulators and our simulator suggest that our technique can accelerate existing simulation speed up to 100x with less memory without sacrificing power estimate accuracy.

Accurate energy estimates for applications running on FPGA based processors at runtime can enable advanced energy management algorithms to dynamically tune the system to optimize its energy consumption and in turn increase the battery life in the mobile systems. This paper presents an online methodology to monitor energy per instruction on FPGA platform that can be used to accurately estimate energy consumed by the application at runtime that is chip specific and suppresses effects of noise. We then extend this methodology to account for energy consumed by the peripherals. Results suggest our methodology is immune to noise and can estimate energy online with less than 3% deviation from total measured energy at less than 2% measurement overhead.

Analog-to-digital (ADC) converters are the most direct and accurate method of power measurement. However it is impractical to instrument ADCs for large number of circuits in a digital system to measure power. In this paper, we present a high resolution digital power measurement technique that is scalable, uses lower power, and is less invasive than ADCs. The core concepts of this new technique has been demonstrated using an evaluation design on FPGA platform. Our latest design can simultaneously monitor 350 different instrumentation points on FPGA circuits and on-board components at 128 kS/sec for each point. The calculated total power differs from those acquired using low noise ADC by less than 3%. Furthermore, an offline data analysis indicates that our method is resilient to instrumentation errors and noise.