Abstract
During Run 2 of the Large Hadron Collider at CERN, the LHCb experiment has spent more than 80% of the pledged CPU time to produce simulated data samples. The upgraded LHCb detector, being commissioned now, will be able to collect much larger data samples, requiring many more simulated events to analyze the collected data. Simulation is a key necessity of analysis to interpret signal, reject background and measure efficiencies. The needed simulation will exceed the pledged resources, requiring an evolution in technologies and techniques to produce these simulated samples. In this contribution, we discuss Lamarr, a Gaudi-based framework to deliver simulated samples parametrizing both the detector response and the reconstruction algorithms. Generative Models powered by several algorithms and strategies are employed to effectively parametrize the high-level response of the multiple components of the LHCb detector, encoding within neural networks the experimental errors and uncertainties introduced in the detection and reconstruction process. Where possible, models are trained directly on real data, leading to a simulation process completely independent of the detailed simulation.
The Computational Challenge of LHCb Simulation
In high-energy physics, simulation is the backbone of data analysis. To understand if a signal is "real," researchers must compare it against simulated data where the "ground truth" is known. However, during Run 2 of the Large Hadron Collider (LHC), the LHCb experiment spent over 80% of its total computing power just generating these simulations.
The traditional "detailed simulation" approach, utilizing tools like Geant4, tracks every single particle as it interacts with detector material. With the LHCb upgrade for Run 3, the volume of data will increase significantly, making traditional simulation methods economically and technically unsustainable. The field requires a way to produce high-quality simulated samples without the massive CPU overhead.
Methodology: Parametrizing Detector Response
Lamarr introduces a framework that shifts the paradigm from simulating interactions to parametrizing outcomes. Instead of calculating how a particle interacts with every layer of silicon or lead, Lamarr uses a pipeline of machine learning models to predict detector response based on the particle's initial properties. This approach skips the most computationally expensive parts of the chain, such as digitization and reconstruction.
Key technical components include:
- Simplified Physics: Charged particles are propagated through the magnetic field using a "single $p_T$ kick" approximation rather than full step-by-step integration.
- Gradient Boosting Decision Trees (GBDTs): These are trained to predict tracking acceptance and reconstruction efficiency—essentially deciding if a particle would have been "caught" by the detector.
- Generative Adversarial Networks (GANs): GANs are used to predict momentum smearing, generate correlation matrices, and reproduce the high-level response of Particle Identification (PID) systems like the RICH and MUON detectors. By using GANs, the framework captures non-Gaussian tails and complex correlations often missed by traditional parametric formulas.
- Direct Training on Real Data: Uniquely, some models in Lamarr can be trained directly on real calibration data, potentially correcting for systematic differences between simulation and reality.
Validation and Performance
The framework was validated using the semileptonic decay $\Lambda_b^0 \to \Lambda_c^+ \mu^- \bar{\nu}_\mu$ as a benchmark. This decay contains four different types of charged particles, providing a rigorous test for the parametrization models. Validation results for the Impact Parameter (IP) $\chi^2$ and PID selection efficiencies show that Lamarr almost perfectly overlaps with detailed Geant4-based simulations, capturing the complex kinematic dependencies required for high-precision physics.
Lamarr achieves massive speedups compared to traditional methods:
- Detailed Simulation: ~2,500 HS06·s per event.
- Lamarr (Full Event): ~500 HS06·s per event.
- Lamarr (Particle Gun Mode): ~1 HS06·s per event, representing a 2,500x speedup.
Conclusion and Future Work
While Lamarr is highly effective for charged particles, work is ongoing to improve parametrizations for neutral particles in calorimeters. The framework currently focuses on high-level reconstructed objects and may not be suitable for analyses requiring low-level raw detector data. Future developments will focus on extending the feature set and optimizing the code for even more complex event topologies, ensuring LHCb can meet the simulation demands of the Run 3 era.
Summary updated