Orateur
Description
Sampling parameter spaces of beyond-the-Standard-Model (BSM) scenarios is computationally challenging due to high dimensionality, complex likelihoods, and stringent experimental constraints. We explore likelihood-free, neural network-based Simulation-Based Inference (SBI) methods to address this problem, focusing on three amortized approaches: Neural Posterior Estimation (NPE), Neural Likelihood Estimation (NLE), and Neural Ratio Estimation (NRE). Their performance is evaluated using the Test of Accuracy with Random Points (TARP), posterior sample efficiency, and computational cost. Applied to the scalar sector of the phenomenological MSSM (pMSSM), NPE yields accurate posterior distributions with minimal samples and outperforms MCMC. Extending to a 9-parameter pMSSM with dark matter observables, the efficiency decreases but the amortized SBI framework continues to produce reliable posteriors, identifying viable points that are predominantly bino-dominated up to $\sim 1.5$ TeV and wino-dominated between 1.5-2 TeV.