Amplitude Surrogates in Particle Physics

• Amplitude surrogates are machine-learned models that approximate complex scattering amplitudes, enabling fast event simulations for the LHC.
• They utilize symmetry-aware architectures and advanced interpolation methods to reduce phase-space errors to O(10⁻⁵) with microsecond evaluations.
• Robust uncertainty quantification via heteroscedastic loss and Bayesian techniques ensures reliable integration into Monte Carlo workflows.

Amplitude surrogates for particle physics are machine-learned models that approximate the costly scattering amplitudes fundamental to LHC theory predictions. These surrogates accelerate event simulation by replacing direct numerical evaluation—often dominated by multi-loop or high-multiplicity matrix elements—with efficient neural architectures and interpolation schemes, while providing rigorous control over accuracy and uncertainty. Their deployment underpins next-generation Monte Carlo integration, shower simulation, and large-scale event generation at collider experiments.

1. Mathematical Formulation and Model Architecture

The central object of interest is the true amplitude $\Atrue(x)$, where $x$ parameterizes the kinematics (external four-momenta and quantum numbers). An amplitude surrogate is a parametric mapping

$\hat\A_\theta : x\;\longmapsto\;\hat\A(x)\equiv\hat\A_\theta(x)\;\approx\;\Atrue(x)$

typically implemented as a neural network. When uncertainty quantification is required, the output may comprise both a predicted mean $\mu_\theta(x)$ and variance $\sigma^2_\theta(x)$.

Activation function choice is critical for approximation fidelity. Systematic tests reveal that pointwise activations such as ReLU and leaky-ReLU yield $\mathcal{O}(10^{-2})$ relative errors, while GELU and GroupKAN layers achieve $\mathcal{O}(10^{-3}\!-\!10^{-4})$ errors. Full Kolmogorov–Arnold Networks do not outperform optimized fixed activations, but GroupKANs approach GELU-level precision without hyperparameter scanning (Bahl et al., 2024).

Architectures incorporating physical symmetries—DeepSets leveraging Mandelstam invariants and Lorentz-equivariant models such as GATr—reduce phase-space errors to $\lesssim10^{-5}$ per point and match integration tolerances required for LHC predictions (Bahl et al., 2024, Bresó et al., 2024). Model size, depth, and input feature engineering (standardized invariants, log-amplitudes) are out-of-the-box determinants for scaling and performance.

2. Uncertainty Quantification and Calibration

Amplitude surrogates address systematic and statistical limitations by pairing each prediction with a learned uncertainty estimate. Training typically involves maximizing a heteroscedastic Gaussian likelihood: $L(\theta)=\sum_i \left[ \frac{(\A_i-\mu_i)^2}{\sigma_i^2}+\ln \sigma_i^2 \right]$ where $\mu_i=\mu_\theta(x_i)$ and $\sigma_i=\sigma_\theta(x_i)$. This drives the network toward agreement with ground truth (fit term) while regularizing the predicted uncertainty.

Bayesian Neural Networks (BNNs) supply a principled framework for full uncertainty estimation, sampling weight posteriors $q(\theta)$ and aggregating predicted means/variances: $\mu(x)=\E_q[\mu_\theta(x)],\qquad \sigma^2_{\mathrm{tot}}(x)=\E_q[\sigma_\theta^2(x)]+\Var_q[\mu_\theta(x)]$ Repulsive ensembles offer a less costly alternative, penalizing function-space proximity among network members to cover diverse minima and inflate the aggregate variance (Bahl et al., 2024).

Calibration is verified using pull distributions: $p_i=\frac{\Atrue(x_i)-\mu(x_i)}{\sigma(x_i)}$ A correctly calibrated surrogate yields pull histograms with mean $\approx0$, width $\approx1$. BNN and heteroscedastic models achieve this across noise, size, and symmetry stress tests. Post-hoc global rescaling adjusts underconfident ensemble predictions (Bahl et al., 2024, Bahl et al., 2 Jan 2026).

3. Interpolation Frameworks and Sparse Grids

Amplitude surrogates in both low and high dimensions exploit state-of-the-art mathematical interpolation schemes:

• Polynomial (Chebyshev) interpolation: Tensor-product grids and Chebyshev nodes deliver exponential convergence for analytic amplitudes but require $10^4$–$10^5$ points in five dimensions for percent-level accuracy (Bresó et al., 2024).
• B-splines: Local tensor-product bases (cubic preferred) afford smooth approximations with competitive sample efficiency.
• Spatially adaptive sparse grids: Hierarchical basis functions, greedily or balanced adapted, reach 1% error with $O(5\times10^3-10^4)$ points, offering rigorous error control and pointwise adaptivity.

Neural methods (MLP, L-GATr) and sparse grids match in sample efficiency, with surrogates evaluating at microsecond timescales per phase-space point.

4. Scaling Laws, Resource Estimates, and Performance

Systematic investigation reveals that amplitude surrogate accuracy scales as a power law in model size, dataset, and compute: $$\varepsilon(N)\simeq(N_c/N)^{\alpha_N},\quad \varepsilon(D)\simeq(D_c/D)^{\alpha_D},\quad \varepsilon(C)\simeq(C_c/C)^{\alpha_C}$$ Exponents are bounded from below by the intrinsic phase-space dimension $d=3n_f-4$ (for $n_f$ final-state particles), typically $\alpha\gtrsim4/d$. Empirical studies confirm that the number of training points, not network size, dominates resource demand for target accuracy ($O(10^4)$ for $d=5$) (Bahl et al., 19 Jan 2026).

Speed-up in practice is dramatic: surrogate evaluation is thousands of times faster than direct loop codes (from milliseconds or seconds to microseconds, or $10^{3}$–$10^{4}\times$ acceleration) (Bahl et al., 2024, Bresó et al., 2024). Adaptive training strategies (e.g., HDBSCAN clustering, KDE-based targeted sampling) further reduce residual error in difficult phase-space regions with modest data increase (Bahl et al., 2 Jan 2026).

5. Deployment Strategies and Integration in Simulation Workflows

Best practice recommendations for surrogate deployment at the LHC include:

1. Choosing symmetry-aware architectures (DeepSets+invariants, Lorentz-equivariant models) for $\mathcal{O}(10^{-5})$ accuracy.
2. Using heteroscedastic loss to capture systematic limitations and facilitate uncertainty calibration.
3. Wrapping the last layers in a BNN for joint systematics/statistics estimation.
4. Validating on held-out pull distributions ($\mathrm{mean}\approx0$, $\mathrm{width}\approx1$).
5. Monitoring $\sigma(x)$ during production and reverting to exact evaluations if uncertainty exceeds user-defined budgets.

Modern event generators benefit from on-the-fly uncertainty thresholds and robust event weighting. Surrogates may supplement exact codes, with uncertainty proxies governing selective fallback to high-precision matrix-element calls. This enables end-to-end uncertainty propagation and robust cross-section predictions (Bahl et al., 2024, Beccatini et al., 11 Dec 2025).

6. Impact on Particle Physics Applications

Amplitude surrogates are now indispensable for collider simulations requiring multi-loop corrections, high-dimensional integrals, and efficient calorimeter shower modeling. In highly granular calorimetry, point-cloud generative surrogates (diffusion+flow architectures) achieve superior fidelity-speed trade-offs compared to regular grid approaches, supporting full-physics benchmarks with $>100\times$ speed-up and percent-level accuracy (Buss et al., 21 Nov 2025).

Multi-jet processes leverage Catani–Seymour factorization within neural frameworks: the network learns only a smooth correction factor, relying on analytic dipole limits. Uncertainty estimates permit strict selection for surrogate-dominated event chains, achieving up to $20\times$ acceleration with all key observables controlled to 1% accuracy (Beccatini et al., 11 Dec 2025).

The geometric view of amplitudes (amplituhedra, matroid polytopes, positive geometry) underlies the analytic landscape and offers combinatorial insights into new QFT models and factorization theorems (Lam, 29 Sep 2025).

---

Amplitude surrogate construction, uncertainty quantification, scaling analysis, and deployment form a unified methodology for controlled, fast precision predictions in particle physics. Their integration establishes a new standard for simulation workflows at the LHC and beyond.