BEGIN:VCALENDAR VERSION:2.0 PRODID:-//CERN//INDICO//EN BEGIN:VEVENT SUMMARY:Physics-Guided Evaluation and Regularization of Diffusion Models f or Calorimeter Simulation DTSTART;VALUE=DATE-TIME:20261029T064500Z DTEND;VALUE=DATE-TIME:20261029T070000Z DTSTAMP;VALUE=DATE-TIME:20261101T195953Z UID:indico-contribution-6197@cern.ch DESCRIPTION:Speakers: Alexandr Dorosh (Laboratory of methods for Big Data Analysis at HSE)\nFast calorimeter simulation requires accurate generative models and meaningful\nmeasures of physical fidelity. We integrate physic s-supervised distributional evaluation and residual regularization into ou r conditional diffusion pipeline for the first time. Shower generation is conditioned on particle entry coordinates and momentum\, allowing the mode l to represent variations in detector response across incident-particle co nfigurations.\n\nThe evaluation uses a Frechet Physics Distance - FPD - co mputed in a representation learned by a network trained to predict particl e kinematics and shower-shape observables. Comparing the means and covaria nce matrices of reference and generated embeddings connects mathematical d istributional assessment with physically relevant features\, complementing PRD-based scores and KL-divergence diagnostics. Physics supervision guide s feature learning towards quantities relevant to shower characterization\ , while the covariance comparison incorporates correlations between learne d features. The diffusion model predicts clean shower maps in a logarithmi c energy representation\, with a timestep-weighted penalty on negative pre dictions added to its reconstruction objective. This introduces a soft non negativity constraint during training while retaining DDIM sampling withou t a physics solver.\n\nOn the validation sample\, the revised pipeline ach ieves an FPD of $0.058$ after\nclipping negative deposits\, compared with $0.889$ under the same evaluation\nprotocol. Both Energy and Physical PRD AUC scores exceed $0.999$\, with the latter\nevaluated on events with fini te shower-shape features. PRD is computed on raw outputs\, preserving a se parate assessment before clipping. Explicit audits additionally examine ne gative deposits\, sparsity\, and total deposited energy\, identifying disc repancies not fully reflected in aggregate\nscores. These checks help dist inguish agreement in the selected feature spaces from physical admissibili ty of individual generated showers. The resulting framework supports syste matic model comparison and targeted improvements to calorimeter simulation \, with controlled ablations needed to isolate the contribution of residua l regularization.\n\nhttps://indico.particle.mephi.ru/event/626/contributi ons/6197/ LOCATION: Conference Hall 3 URL:https://indico.particle.mephi.ru/event/626/contributions/6197/ END:VEVENT END:VCALENDAR