Location
Havener Center, St. Pat's Ballroom C
Presentation Date
April 21, 2023, 11:30am-12:30pm
Session
Session 1
Description
In this study, we developed deep learning potentials for ultra-high temperature high entropy driven diborides of MB2 where M is the transition metal made of Ti, Zr, Hf, Nb, Ta. The materials are being considered as the heat shield for hypersonic vehicles due to their high thermal conductivity and thermal stability. We used quantum mechanical data as our input for training and validation. Specifically, we use the energy, force and stress data from molecular dynamics simulations at elevated temperatures using quantum mechanics calculations. We used the DeePMD neural network code and varied the hyper-parameters such as cut-off radius, batch size, the number of hidden layers, maximum number of neighbors per atom site, etc. We showed the effect of these parameters on the accuracy of the potential models. Similarly, we showed the effect on the simulated stability of the crystal structure at 2000K. Overall, we have been able to develop a new potential capable of modeling the thermal stability of the diboride compounds.
Meeting Name
32nd Annual Spring Meeting of the NASA-Mo Space Grant Consortium
Document Type
Presentation
Document Version
Final Version
File Type
text
Language(s)
English
Rights
© 2023 The Authors, all rights reserved.
Development of Deep Learning Potentials for Ultra-high Temperature High-Entropy Driven Diborides
Havener Center, St. Pat's Ballroom C
In this study, we developed deep learning potentials for ultra-high temperature high entropy driven diborides of MB2 where M is the transition metal made of Ti, Zr, Hf, Nb, Ta. The materials are being considered as the heat shield for hypersonic vehicles due to their high thermal conductivity and thermal stability. We used quantum mechanical data as our input for training and validation. Specifically, we use the energy, force and stress data from molecular dynamics simulations at elevated temperatures using quantum mechanics calculations. We used the DeePMD neural network code and varied the hyper-parameters such as cut-off radius, batch size, the number of hidden layers, maximum number of neighbors per atom site, etc. We showed the effect of these parameters on the accuracy of the potential models. Similarly, we showed the effect on the simulated stability of the crystal structure at 2000K. Overall, we have been able to develop a new potential capable of modeling the thermal stability of the diboride compounds.
