Optimized LSTM neural networks for material damage predictions

We present a predictive model using Long Short-Term Memory (LSTM) neural networks to forecast component material conditions at timepoints (t + n) based on monitored historical sample data up to time t. LSTM architectures are renowned for their ability to capture long-term dependencies and are particularly well-suited for sequential data. In this work we introduce an improved and optimized LSTM model designed to enhance the performance and efficiency of sequential data predictions. Our approach includes a statistical evaluation that utilizes experimental process information extracted as time series data to develop multi-layer models for predicting the condition of component materials. We demonstrate the effectiveness of the new approach on a case study using NASA Ames milling data set by Agogino and Goebel (2007). Results suggest that the new method can precisely predict the component condition while significantly improving the efficiency. The proposed modeling approach can also be directly applied for predictions of other sequential data, as for example materials corrosion data, required for the design and safety evaluation of nuclear power plants. • Proposed a predictive model based on Long Short-Term Memory (LSTM) deep neural networks to forecast component material condition. • Included a statistical evaluation that utilizes experimental process information to improve optimization. • Designed a multi-layer deep neural network model specifically for sequential data prediction. • Demonstrated effectiveness of the approach using the NASA Ames milling data set, showing precise prediction of component condition with significantly improved efficiency of the standard LSTM model. • The approach can be applied to other sequential data, such as materials corrosion data for nuclear power plant design and safety evaluation.