Foreseeing the Future: Redefining Hanford Planning with Machine Learning and Operations Research - 22139

Washington River Protection Solutions (WRPS) is the tank operations contractor that oversees the management, retrieval, and treatment of radioactive, hazardous tank waste for the Hanford mission. Ever since Hanford ceased to be a production facility several decades ago, its primary function has been safe waste storage; consequently, it has experienced a relatively low operational tempo. However, as Hanford shifts to delivering waste feed to the Hanford Tank Waste Treatment and Immobilization Plant (WTP), that operational tempo will increase significantly. With this tempo increase, there will also be an increase in operational complexity. The cleanup effort will require the integration of dozens of unique facilities and processes, each governed by its own set of operating logic, physical configuration, and constraints. Additionally, each facility has multiple interface points, making the operations of any one process potentially significant to the operations of others. Traditionally, the primary methods for developing operational plans have been engineering judgement, operator experience, and lessons learned. Although these can be good starting points for assessing current risk, they are also vulnerable to human biases, misconceptions, and 'group think.' They also provide limited foresight into the complex interactions of existing and proposed equipment and facilities. To combat this blind spot, WRPS has been developing and utilizing operations research (OR) models that serve as 'digital twins' for each of the major current and future facilities within the Hanford complex. These models provide the capability to mathematically and logically link these multifaceted processes to one another and to then vary their inputs. This, in turn, allows for significantly increased insight into operational dependencies that would otherwise be far too complicated to fully grasp and analyze. These insights have been instrumental in preparing the Hanford Site in key areas for the upcoming increase in operations. However, as more models are developed and current models mature, the amount of data to analyze has increased exponentially. Extending model functionality to analyze further operational constraints has also become more time consuming. Future insights and recommendations from the OR models will depend on analytical capability and model execution speed. Machine learning (ML) tools have been developed to address these issues. These tools are being used to streamline data analysis and expand OR model functionality. To date, three different ML tools have been developed that support modeling of the Effluent Treatment Facility (ETF) and Waste Feed Delivery (WFD) system. The ETF processes a diverse array of site-generated waste streams. Each of these waste streams carries a unique set of processing constraints that can impact how the facility operates. To avoid compounding these constraints, the streams are typically stored and treated separately. However, if certain off-normal conditions are met, mixing may become necessary. Enhancing the ETF model functionality to account for these mixtures has led to the development of the Surrogate Flowsheet Model (SFM) ML tool. The SFM is an ML model that replicates the functionality of an ETF flowsheet, which is necessary to generate treatment profiles for mixed waste streams. The SFM replicates this functionality with far less impact on the OR model runtime. The WFD system has many key parameters that affect overall performance. Each parameter is accounted for in the Double Shell Tank (DST) model and can be varied to study parameter impact and importance. The Parameter Analysis ML tool has been developed to identify the most important parameter and the specific contribution of each parameter to overall model performance. One of the identified key parameters in the DST model is the order in which waste is transferred through the WFD system. There are any number of ways to move waste around the system but predicting the most effective ones is close to impossible without the use of simulation and modeling analysis. The Transfer Order ML tool addresses this issue by identifying the most effective simulated transfer orders and the key characteristics of those orders that contribute to WFD system success. As the Hanford site continues its preparations for waste treatment operations, OR Modeling will play an increasingly pivotal role in operational decision making and planning. To realize the full benefit of these models and the insights they uncover, pursuit of innovative methods and technologies is essential, especially as data and complexity increase. This paper describes part of that journey and demonstrates how ML tools have proven highly valuable in augmenting OR Model functionality and analysis. (authors)