Speaker
Description
Ion therapy employs protons and heavier ions (e.g., helium, carbon, oxygen) for cancer treatment due to their advantageous physical and biological properties, particularly effective against radio-resistant tumors. However, precise modeling of nuclear fragmentation processes, which critically influence dose distributions, biological effectiveness, and overall treatment accuracy—especially with heavier ions—remains challenging. Monte Carlo (MC) simulations are considered the gold standard for dose calculation in ion therapy, with Geant4 being one of the most widely used tools. Nonetheless, Geant4 currently lacks dedicated nuclear interaction models for energies below 100 MeV/u, leading to discrepancies compared to experimental data. Advanced nuclear interaction models, such as the Boltzmann-Langevin One Body (BLOB) model, when interfaced with Geant4, provide improved accuracy but are prohibitively computationally intensive (~4 minutes per interaction), thus impractical for routine clinical use.
To address these limitations, we propose a hybrid modeling approach combining deep learning (DL) techniques with classical nuclear interaction models. Specifically, we developed a physics-informed neural network explicitly designed to incorporate interaction symmetries within its architecture and process nucleons as batch inputs, enabling general applicability across various nuclear reactions. As a proof-of-concept, our model was trained on two computationally demanding tasks: calculating the mean-field potential and computing Hamiltonian derivatives with respect to generalized coordinates in two nuclear interaction models implemented in Geant4—Quantum Molecular Dynamics (QMD) and Light Ion QMD (LiQMD).
We trained our DL models on extensive datasets generated by Geant4 simulations, encompassing multiple ion species and clinically relevant energies (90–130 MeV/u), achieving highly accurate emulation of both potential and Hamiltonian derivatives (Median Relative Error ≤ 0.84%). Integrating the DL models into Geant4, we replaced their corresponding classical methods and conducted several fragmentation simulations. Comparisons of double-differential cross sections for fragment production between classical and hybrid simulations showed that while potential emulation primarily yielded accurate cross sections for lighter fragments, Hamiltonian derivative emulation provided high accuracy across all fragment species, demonstrating strong generalization, interpolation, and extrapolation capabilities.
This study demonstrates the feasibility and potential of our approach, paving the way for future developments. Upcoming work includes increasing computational efficiency through an optimized implementation that leverages GPU acceleration and extending this hybrid methodology to more complex models, such as BLOB, ultimately aiming at precise and rapid nuclear fragmentation modeling for clinical ion therapy planning and validation.
