Performance of Simulation-Based Inference (SBI) and its GPU-parallelized version (mean ± standard deviation across various simulation setups).

Training times are reported in minutes and simulation times, defined as the time required to establish N pairs ( θn , xn ) through model simulation, are reported in seconds. The last column (KL) measures the Kullback-Leibler (KL) divergence to estimate the accuracy of posterior approximations. As no ground truth data is available for our model, the GPU-parallelized version of SBI is employed to estimate the posterior. We utilize a substantially larger set of summary statistics as detailed in Table 1, and the number of samples drawn from the prior distribution has been increased to 2 × 105. The differences in performance metrics highlight the efficiency gains with GPU parallelization, particularly in reduced simulation times across all simulation counts and dimensions. The mean and standard deviation are obtained across all 10 runs.