This seminar presents two machine learning contributions to ATLAS physics.

The first part introduces Log Gaussian Cox Processes (LGCP) for non-parametric background modeling in low-statistics searches. Inspired by challenges in low-mass diphoton resonance searches where complex shapes and limited Monte Carlo statistics challenge traditional methods, LGCP uses Bayesian inference to learn background distributions without assuming functional forms. Validation demonstrates automated, flexible fitting with built-in uncertainty quantification applicable to a broad range of analyses.

The second part covers ditau trigger development for ATLAS Phase 2. With HL-LHC luminosity increasing tenfold, new trigger strategies are needed to maintain efficiency while controlling rates. I present preliminary performance results, the trigger architecture, and future implementation steps.