Fast prediction of irradiation-induced cascade defects using denoising diffusion probabilistic model

• A machine learning method was proposed to rapidly and accurately predict the spatial coordinates of point defects at given primary knock atom (PKA) energies in W. • The computational scheme has the capability to generate the cascade collision defects across a broad spectrum of PKA energies, spanning from 0 to 100.0 keV. • The trained DDPM only needs 0.06 s to generate the cascade collision point defects at a specific PKA energy, which is at least five orders of magnitude faster than the MD simulation time. Irradiation-induced cascade collisions produce numerous point defects within materials, which can severely deteriorate their thermo-mechanical properties and overall performance. We propose a computational scheme that combines molecular dynamic (MD) simulations with a denoising diffusion probabilistic model (DDPM) to rapidly and accurately predict the spatial coordinates of point defects at any given primary knock atom (PKA) energy, ranging from 0 to 100.0 keV. Importantly, this capability extends to PKA energies that are exclusive from the training data set, demonstrating the robustness and generalizability of the model. The proposed scheme has been thoroughly validated by several designed indicators, including the Fréchet inception distance, the number of point defects, the distance from vacancies and self-interstitial atoms (SIAs) to their respective centroids, the inter-centroid distance between the vacancies and SIAs, the probability density of clustered defect sizes, and the sub-cascade number. Compared to MD simulations, the DDPM can generate point defects at a specific PKA energy at least ten thousand times faster. By offering a rapid and reliable means to model defect distributions across various energy levels, the proposed scheme benefits the comprehension of the cascade process and provides a valuable database for both experimental investigations and large-scale simulations.