About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
PINN Approach to Accelerated Spin Dynamics Simulation |
| Author(s) |
Doyl Dickel, Hala Ben Messaoud, Christopher Barrett |
| On-Site Speaker (Planned) |
Doyl Dickel |
| Abstract Scope |
While machine learned interatomic potentials have greatly increased the accuracy of atomistic simulations and are considerably faster than ab initio methods, they often lag significantly behind older empirical formalisms in terms of computational efficiency. This is particularly the case for magnetically sensitive potentials where the inclusion of atomic descriptors which include spin degrees of freedom can result in spin dynamic performance two or more orders of magnitude slower than simple Heisenberg exchange potentials. We present here a framework, based on physics informed neural networks, applicable to multiple styles of MLIP, which trivializes the extra computational expense of spin dynamics compared to lattice dynamics, while achieving similar accuracies to existing machine learned potentials. We demonstrate this framework for a pair of magnetic potentials and show its performance in dynamic simulation. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Magnetic Materials, Computational Materials Science & Engineering, Machine Learning |