Exploring the BSM parameter space with Neural Network aided Simulation-Based Inference

This paper demonstrates that likelihood-free Simulation-Based Inference, particularly the Neural Posterior Estimation method, offers a computationally efficient and accurate alternative to traditional MCMC for exploring high-dimensional Beyond the Standard Model parameter spaces, successfully inferring posterior distributions for both 5-parameter and 9-parameter pMSSM scenarios including dark matter constraints.

The Gist

Imagine you are a detective trying to solve a massive mystery: What is the universe made of beyond what we already know?

In the world of particle physics, scientists have a "Standard Model" (like a rulebook for how particles behave). But we know this rulebook is incomplete. It doesn't explain things like Dark Matter or why gravity is so weak. So, physicists have created "Beyond the Standard Model" (BSM) theories. Think of these theories as giant, complex video games with thousands of adjustable settings (parameters).

The problem? The "game" has so many settings that trying to find the specific combination that matches reality is like trying to find a single specific grain of sand on all the beaches on Earth, blindfolded.

The Old Way: The Slow, Exhaustive Search

Traditionally, scientists used a method called MCMC (Markov Chain Monte Carlo). Imagine you are in a dark room trying to find the exit. You take a step, check if you hit a wall, take another step, check again, and repeat. You might wander around for days, bumping into walls, just to find the door.

• The Issue: This method is incredibly slow and computationally expensive. It requires a supercomputer to run for days or weeks just to get a rough idea of where the "exit" (the correct physics) might be.

The New Way: The "AI Detective" (SBI)

This paper introduces a smarter approach using Artificial Intelligence (Neural Networks). Instead of wandering blindly, the AI learns the layout of the room by watching a simulator run millions of times.

The authors tested three different "AI detective" styles:

1. NPE (Neural Posterior Estimation): The AI learns to guess the settings directly based on the clues.
2. NLE (Neural Likelihood Estimation): The AI learns to guess how likely a specific setting is to produce the clues.
3. NRE (Neural Ratio Estimation): The AI learns to compare two settings to see which one fits better.

The "TARP" Test: Did the AI Cheat?

How do you know the AI isn't just guessing randomly? The authors invented a test called TARP (Test of Accuracy with Random Points).

• The Analogy: Imagine you hire a weather forecaster. To test them, you give them a random day's data and ask, "What's the chance it rains?" If they are good, their predictions should match reality perfectly over time. If they are bad, their predictions will be all over the place.
• The Result: The TARP test showed that NPE was the only detective that didn't cheat. It learned the rules perfectly. The other two (NLE and NRE) got confused and gave unreliable answers.

The Experiment: Two Levels of Difficulty

The team tested their AI on a specific theory called pMSSM (a popular version of the supersymmetry theory).

Level 1: The 5-Parameter Puzzle
They started with a simplified version of the theory with only 5 adjustable knobs (parameters).

• The Result: The NPE AI found the correct settings in 24 hours.
• The Comparison: The old MCMC method took 72 hours (3 times longer) and actually missed some important details about how the particles interact. The AI was faster and more accurate.

Level 2: The 9-Parameter Nightmare
Then, they made it harder. They added Dark Matter constraints, turning the puzzle into a 9-knob nightmare. This is like adding a new rule to the video game: "The character must also be invisible to the naked eye."

• The Challenge: Finding a solution that fits all the rules (Higgs boson data, flavor physics, AND Dark Matter) is incredibly hard. Most random guesses fail.
• The Result: Even with the added difficulty, the NPE AI succeeded. It found valid solutions, though it was less efficient than before (because the "good" area was so small).
• The Discovery: The AI revealed that the "Dark Matter particle" in this theory behaves differently depending on its weight:
  • Lighter particles (< 1.5 TeV): They are mostly "Bino" (a specific type of particle).
  • Heavier particles (1.5 - 2 TeV): They switch to being mostly "Wino."

Why This Matters

This paper proves that AI can replace the slow, brute-force methods of the past.

• Speed: It solves problems 3x faster.
• Accuracy: It finds the "true" answer more reliably than traditional math.
• Scalability: It can handle complex, multi-layered mysteries (like Dark Matter) that would take humans years to solve.

In a nutshell: The authors built a super-smart AI detective that learned to solve the universe's hardest physics puzzles. It didn't just guess; it learned the patterns, passed a strict honesty test (TARP), and found the answers faster and better than the old-school methods ever could. This opens the door to exploring even more complex theories of the universe in the future.

Technical Summary

1. Problem Statement

High Energy Physics (HEP) faces significant challenges in exploring the parameter spaces of Beyond the Standard Model (BSM) scenarios, such as the phenomenological Minimal Supersymmetric Standard Model (pMSSM).

• Computational Cost: Traditional methods like Markov Chain Monte Carlo (MCMC) require explicit calculation of the likelihood function at every step, which is computationally expensive for high-dimensional spaces (e.g., 19 parameters in pMSSM).
• Intractable Likelihoods: In many complex BSM scenarios, the likelihood function is not analytically tractable or is too costly to compute.
• Sampling Efficiency: As experimental constraints (from LHC, flavor physics, and Dark Matter) tighten, the "good" regions of the parameter space become sparse. Random scanning or standard MCMC often wastes computational resources exploring excluded regions.
• Validation: There is a lack of robust, necessary, and sufficient methods to validate the accuracy of posterior distributions generated by likelihood-free inference methods.

2. Methodology

The authors propose a Simulation-Based Inference (SBI) framework, specifically utilizing amortized methods. These methods train neural networks to learn the mapping between parameters ($\theta$) and observables ($x$) without explicitly calculating the likelihood function $P(x|\theta)$.

Core SBI Methods Evaluated

The paper compares three amortized SBI approaches:

1. Neural Posterior Estimation (NPE): Directly learns the posterior distribution $P(\theta|x)$ using a neural density estimator (Masked Autoregressive Flow - MAF).
2. Neural Likelihood Estimation (NLE): Learns the likelihood function $P(x|\theta)$, which is then combined with the prior to derive the posterior.
3. Neural Ratio Estimation (NRE): Learns the likelihood ratio $r(x, \theta) = P(x|\theta)/P(x)$ using a binary classifier (ResNet) to distinguish joint samples from marginal samples.

Validation: TARP Test

To ensure the reliability of the inferred posteriors, the authors employ the Test of Accuracy with Random Points (TARP).

• Mechanism: It tests the coverage probability of the estimated posterior. It checks if the true parameters fall within the predicted credible intervals at the expected frequency.
• Significance: Unlike simple coverage tests, TARP is argued to be both necessary and sufficient to detect inaccurate inferences (e.g., over- or under-confidence).

Case Studies

The framework is tested on two scenarios using the LtU-ILI (Learning the Universe Implicit Likelihood Inference) code:

1. pMSSM5 (5 Parameters): Focuses on the scalar sector ($M_2, \mu, \tan\beta, M_A, A_t$) constrained by Higgs mass, coupling strengths, and flavor observables ($B \to X_s\gamma$, $B_s \to \mu^+\mu^-$, etc.).
2. pMSSM9 (9 Parameters): Adds Dark Matter (DM) constraints (relic density $\Omega h^2$, spin-independent/spin-dependent scattering cross-sections) and varies additional mass parameters ($M_1, M_3, m_{\tilde{l}}, m_{\tilde{q}}$).

3. Key Results

Performance Comparison (NPE vs. NLE vs. NRE)

• TARP Validation: The NPE method consistently passed the TARP test, producing unbiased posterior distributions. In contrast, NLE and NRE failed the TARP test, showing significant deviations (S-shaped curves) indicating biased inference.
• Sample Efficiency: NPE achieved high posterior sample efficiency (surviving points within $3\sigma$ of constraints) with fewer training samples. For pMSSM5, NPE saturated in efficiency at $1 \times 10^5$ samples, whereas NLE and NRE required the full dataset and still performed poorly.
• Computational Speed: NPE was significantly faster. For the pMSSM5 case, NPE completed in 24 hours, while the equivalent MCMC run took 72 hours. NPE was 2–4 times faster than NLE/NRE for training.

pMSSM5 Analysis (Higgs & Flavor)

• Posterior Distributions: NPE correctly identified the correlation between the trilinear coupling $A_t$ and the Higgs mass, disfavoring $A_t \approx 0$ (which leads to $m_h < 123$ GeV).
• Comparison with MCMC: While MCMC allowed a broader, less constrained parameter space (failing to capture the dip at $A_t=0$), NPE produced a more faithful posterior distribution that aligned better with theoretical expectations and previous literature.

pMSSM9 Analysis (Including Dark Matter)

• Efficiency Drop: Adding DM constraints drastically reduced the viable parameter space. Random sampling efficiency dropped to $\sim 0.2\%$, while NPE maintained an efficiency of $\sim 5.0\%$ (25x improvement over random).
• DM Phenomenology: The SBI-predicted viable points revealed distinct mass-dependent dominance:
  • $\lesssim 1.5$ TeV: Dominated by Bino-like LSPs (satisfying relic density via co-annihilation with charginos/stops/staus or the heavy Higgs funnel).
  • $1.5 - 2.0$ TeV: Dominated by Wino-like LSPs (satisfying relic density via chargino co-annihilation).
  • Higgsino: No viable Higgsino-dominated points were found in the explored range (up to 2 TeV) due to constraints from XENON1T direct detection limits.

4. Key Contributions

1. Systematic Benchmarking: Provided a rigorous comparative analysis of amortized SBI methods (NPE, NLE, NRE) in the context of high-energy physics, demonstrating that NPE is superior for this specific class of problems.
2. Validation Framework: Successfully applied the TARP test to validate SBI methods in HEP, establishing a robust protocol for verifying posterior accuracy beyond simple likelihood checks.
3. Efficiency Gains: Demonstrated that amortized SBI (specifically NPE) can reduce computational time by a factor of 3 compared to MCMC while providing more accurate posterior distributions.
4. Phenomenological Insights: Mapped the viable pMSSM parameter space under combined Higgs, flavor, and DM constraints, identifying specific mass ranges for Bino and Wino dominance and ruling out Higgsino dominance up to 2 TeV under current constraints.

5. Significance

This work addresses a critical bottleneck in modern particle physics: the inability to efficiently explore high-dimensional, constrained parameter spaces. By proving that likelihood-free inference via NPE is both faster and more accurate than traditional MCMC for BSM scenarios, the paper advocates for a paradigm shift in how phenomenological studies are conducted. The successful application of TARP validation ensures that these "black box" neural network methods can be trusted for scientific discovery, paving the way for their use in future LHC analyses and Dark Matter searches.