pyMAISE: A Python platform for automatic machine learning and accelerated development for nuclear power applications

Abstract

Despite significant advancements in artificial intelligence and machine learning (AI/ML) algorithms and their potential in nuclear engineering applications , the field still lacks a framework that automates ML model development and deployment for nuclear engineering problems. To address this, pyMAISE ( M ichigan A rtificial I ntelligence S tandard E nvironment) is introduced, which is a Python package that features automated hyperparameter tuning, model explainability, model training, validation, postprocessing, and deployment for various ML models relevant to nuclear engineering. pyMAISE provides a platform for researchers to demonstrate new models on benchmarked datasets and currently supports nine benchmark problems across reactor physics and design, reactor control, thermal hydraulics , fuel performance, safety analysis, and anomaly detection . In this work, pyMAISE is demonstrated in three applications: critical heat flux prediction, microreactor power prediction, and fault detection in electronic signals. pyMAISE provided efficient model search and performance results that meet or exceed other studies. For critical heat flux prediction, feedforward neural networks (FNN) and random forests were the top models achieving R 2 = 0 . 999 when six input features were used. FNN was the best performer for predicting microreactor quadrant power R 2 = 0 . 97 , 0.26 greater than the closest classical ML model. In fault detection, pyMAISE models achieved 81% test accuracy in detecting faulty signals using long short-term memory, which may prevent various accident scenarios that could cause facility downtime. As pyMAISE continues to develop through a multi-phase approach, the goal is to integrate uncertainty quantification and deployment tools to expedite the creation of explainable and licensable AI technologies for nuclear power plants .

Topics & Concepts