Accelerating Casimir Force Computations Between Complex Particles Using a Transformer-Based Scattering Framework

Casimir forces play a fundamental role in determining the collective behavior of particulate systems, from colloidal assemblies to cosmic dust aggregates. Accurate evaluation of these fluctuation-induced forces requires integrating over all electromagnetic modes while accounting for the detailed geometry and composition of the interacting particles. However, for irregularly shaped or compositionally heterogeneous objects, such calculations based on scattering-matrix formulations become computationally intensive, often prohibiting many-body simulations. In this work, we develop a Transformer-based deep learning framework to serve as a surrogate model for rapid prediction of electromagnetic scattering properties of arbitrary compact 3D objects. Trained on high-fidelity datasets generated from rigorous scattering computations, the model enables fast and accurate reconstruction of the full scattering matrix across frequency spectra. This framework accelerates Casimir force computations by two to three orders of magnitude while maintaining physical fidelity, offering a scalable route to investigate many-body interactions and collective evolution in granular and astrophysical systems.