Predicting Complex Erosion Profiles in Steam Distribution Headers with Convolutional and Recurrent Neural Networks

Abstract

The effects of erosion due to particle impingement continue to be of immense concern in various energy and technology industries. Brute force computational fluid dynamics (CFD) approaches allow accurate predictions of complex erosion processes; however, these large-scale calculations can be very computationally expensive. Specifically, when different initial conditions are required to analyze the system, the CFD simulations must be restarted de novo without recourse to previously converged cases. To address these issues, we harness convolutional neural network (CNN) and long-and short-term memory (LSTM) machine learning approaches to predict complex surface erosion profiles in steam distribution headers for the first time. We show that this hybrid machine learning approach can accurately predict entire particle trajectories and surface erosion profiles when only initial positions and velocities are inputted into the algorithm. Most importantly, our approach is 600 times faster than conventional CFD calculations and gives impressive R 2 scores of 0.91 and 0.71 for particle trajectories and surface erosion profiles, respectively. Taken together, our hybrid machine learning approach is a promising technique for accurately predicting surface erosion rates but with a significantly reduced computational cost.

Understanding and Designing a High-Performance Ultrafiltration Membrane Using Machine Learning

“…To understand the built models and thus provide insightful guidance for future membrane fabrication, identifying potential controlling fabrication parameters for membrane properties and performance is essential. In previous work, feature importance analysis has been performed to explain different models (such as random forest and neural network), providing insight into the roles of input features. , Here, the importance and impact of each feature on the targets were analyzed using the SHAP method . The Shapley value for input feature x (of n total input features) given the prediction p by the built ML model was calculated by Ø x false( p false) = ∑ S ⊆ N / x false| S false| ! false( n − | S | − 1 false) ! n ! false[ p ( S ∪ x ) − p ( S ) false] where S is the subset for each feature without feature x , p ( S ∪ x ) represents the predictions by the built ML model considering feature x , and p ( S ) represents the predictions without considering feature x .…”

Yield and Properties Prediction Based on the Multicondition LSTM Model for the Solvent Deasphalting Process

“…Due to the memory function of LSTM, LSTM is gradually applied to industrial time series analysis and prediction. A machine learning method combining CNN and LSTM can accurately predict the entire particle trajectory and surface erosion profile . LSTM can effectively predict the CH 4 leakage source of chemical processes, and it is found that the model with larger time steps has better performance and stronger generalization ability .…”

“…A machine learning method combining CNN and LSTM can accurately predict the entire particle trajectory and surface erosion profile. 33 LSTM can effectively predict the CH 4 leakage source of chemical processes, and it is found that the model with larger time steps has better performance and stronger generalization ability. 34 A maximum correntropy criterion-based LSTM (MCC-LSTM) neural network is proposed to develop a reliable soft sensor model for quality prediction, which can identify the outliers and reduce their negative effects on the prediction in some extent.…”

Modeling of extrusion through a Garvey die: Cost and performance

“…In the same spirit of using GA to test new approaches to improve extrusion technologies, there are studies pointing to machine learning as a possible means to reduce computational costs. Yang et al [ 28 ] studied the prediction of erosion profiles in steam distribution headers with convolutional and recurrent neural networks. They used a medium size server with 32 processors 300 GB of RAM and 3,125 simulations to train their algorithms in a system with 451,997 finite elements.…”