About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Molecular Approaches to Ceramics; Synthesis, Processing, Modeling, and Derived Ceramics
|
| Presentation Title |
Deep Learning-Based Framework for ReaxFF Development in Reactive Material Systems |
| Author(s) |
Zihan Wang, Asma Ul Hosna, Mozhdeh Mirakhory, Adri van Duin, Wei Chen |
| On-Site Speaker (Planned) |
Zihan Wang |
| Abstract Scope |
ReaxFF enables reactive molecular dynamics (MD) simulations by describing bond formation, bond dissociation, and charge redistribution through bond-order interactions and dynamic charge equilibration. However, molecular configurations evolve dynamically during MD, while conventional ReaxFF development relies on a fixed training database that may not cover the chemical space generated under reactive conditions. To address this gap, we propose a deep learning-based feedback framework for ReaxFF development. Molecules from both the MD simulation set and the ReaxFF training set are reconstructed as molecular graphs and embedded into a shared low-dimensional latent space using Graph2Vec, where relative distances represent molecular similarity. Density-based analysis identifies frequently sampled MD configurations that are poorly covered by the training set. Representative configurations are then selected and added back to the training database, enabling iterative ReaxFF refinement toward MD-sampled chemistry. This framework is demonstrated on CCC-relevant carbon-, oxygen-, and nitrogen-containing systems, with potential extension to other compositions. |