High-Dimensional Bayesian Calibration of Expensive Nuclear Models with Differentiable Emulation

This paper introduces DREAM, a differentiable emulation strategy that enables efficient, gradient-based Bayesian calibration of expensive nuclear models by compressing legacy parameter-dependent operators offline, thereby allowing Hamiltonian Monte Carlo methods to rapidly converge on high-dimensional posteriors with exact likelihood gradients at minimal computational cost.

The Gist

The Big Problem: The "Black Box" and the "Blind Search"

Imagine you are trying to tune a very complex, expensive machine (like a nuclear physics model) to match real-world data. This machine has 18 different knobs you can turn (parameters) to change how it behaves.

The problem is twofold:

1. It's slow: Turning the knobs and seeing what happens takes a long time (minutes per try).
2. It's a "Black Box": The machine was built years ago using old code. It tells you the result, but it refuses to tell you which way to turn the knobs to get a better result. It gives no "gradients" (directional hints).

Because the machine gives no hints, scientists have to use a "blind search" method. They try random combinations of knobs, check the result, and hope they get closer. To find the perfect setting in a space with 18 knobs, they might need to try the machine 100,000 times. At minutes per try, this would take days or weeks of computer time, and even then, they might get stuck in a local "good enough" spot rather than finding the best spot.

The Solution: DREAM (The "Smart Map" Strategy)

The author introduces a new method called DREAM. Think of it as building a high-speed, GPS-enabled map of the machine's behavior before you start your journey.

Here is how DREAM works in two steps:

Step 1: The Offline "Snapshot" Phase (Making the Map)
Before doing any real calculations, the author runs the old, slow machine at hundreds of different settings on a grid.

• The Analogy: Imagine taking a photo of the machine at every possible combination of knobs.
• The Trick: Instead of saving every single photo (which is too much data), the author uses a mathematical compression technique (called SVD) to realize that all these photos are actually just slight variations of a few "master images."
• The Result: They create a tiny, compressed "dictionary" of how the machine behaves. This is done once and takes about 37 minutes.

Step 2: The Online "Real-Time" Phase (Driving the Car)
Now, when the computer needs to test a new setting during the search:

• The Analogy: Instead of driving the slow machine, the computer looks at its "dictionary" and instantly reconstructs what the machine would have done at that setting.
• The Superpower: Because this reconstruction is built using modern, differentiable math (like a smart video game engine), the computer doesn't just get the result; it instantly knows exactly which way to turn the knobs to improve the result.
• The Speed: This happens in less than one millisecond (0.001 seconds).

The Result: From Days to Minutes

By using this "Smart Map," the author replaced the blind search with a guided search (called Hamiltonian Monte Carlo).

• Old Way: A blind search trying 100,000 times would take days and might still get lost.
• DREAM Way: The guided search found the perfect answer in 27 minutes on a single graphics card.
• Accuracy: The "map" was so accurate that the tiny errors in the map were 20 times smaller than the natural uncertainty in the physics model itself. This means the result is trustworthy and not just an artifact of the shortcut.

What Did They Actually Find?

The author tested this on a specific nuclear reaction: a deuteron (a heavy hydrogen nucleus) hitting a Nickel-58 atom.

1. The Physics: They successfully mapped out how the deuteron is "absorbed" by the surface of the nickel atom.
2. The Discovery: They found that the "surface absorption" (how the atom eats the deuteron) is about 40% stronger than previous standard models predicted.
3. The Asymmetry: They found a significant difference between how protons and neutrons interact with the surface. However, the author is careful to say this is a "representative payoff" of the method, not a final, settled law of physics. They suggest that to be sure, this method needs to be applied to more data sets (different energies) in the future.

The Bottom Line

The paper doesn't claim to have solved all of nuclear physics. Instead, it claims to have built a universal tool that allows scientists to use powerful, fast, gradient-based search methods on old, slow, "black box" nuclear models.

• The Metaphor: It's like taking a slow, old car that has no GPS and building a real-time navigation system for it. You don't change the car's engine; you just give it a brain that knows exactly where to go, turning a multi-day journey into a 27-minute drive.

The author concludes that this method works for any nuclear model where the parameters change smoothly, opening the door for much more precise and faster analysis of complex nuclear reactions in the future.

Technical Summary

Technical Summary: High-Dimensional Bayesian Calibration of Expensive Nuclear Models with Differentiable Emulation

Problem Statement
Full Bayesian calibration of expensive nuclear models has historically been impractical due to two compounding obstacles. First, likelihood evaluations require costly quantum-mechanical solves (often taking minutes per evaluation). Second, the parameter-dependent operators within these solvers are typically implemented in legacy codes (e.g., Fortran) that provide no gradient information. Without exact likelihood gradients, samplers must rely on gradient-free methods (e.g., affine-invariant ensemble samplers), which degrade sharply in high-dimensional, correlated posteriors. This forces gradient-free approaches to require on the order of $10^5$ evaluations to explore the parameter space, rendering full Bayesian inference computationally prohibitive for complex models like coupled-channels or many-body systems.

Methodology: The DREAM Framework
The paper introduces DREAM (Differentiable Reduced-basis Emulator with Automatic gradients for Markov-chain inference), a strategy designed to enable gradient-based Bayesian inference for non-affine, expensive legacy operators. The framework operates in two distinct stages:

1. Offline Stage (Pre-computation):

   • The parameter-dependent operator is treated as a black box. It is sampled on a dense grid of parameters using the legacy code.
   • The resulting ensemble of "snapshots" (matrix representations of the operator) is compressed using Singular Value Decomposition (SVD).
   • Crucially, the compression is applied to the operator itself, not just a surrogate observable. For the specific application (Continuum-Discretized Coupled-Channels, CDCC), the reduced optical potential is decomposed into a parameter-independent part and a sum of form factors multiplied by linear depth parameters. The nonlinear geometry dependence is captured in the form factor matrices, which are compressed into a low-rank basis (POD modes) and scalar expansion coefficients.

2. Online Stage (Inference):

   • The legacy solver is never called during inference. Instead, a differentiable engine (implemented in JAX) reconstructs the operator for any proposed parameter vector $\Omega$.
   • Reconstruction: The scalar coefficients from the SVD are interpolated (using bilinear interpolation) based on the proposed geometry parameters to reconstruct the reduced operator $V_{red}(\Omega)$.
   • Solving: The reduced Galerkin system is solved using direct linear algebra.
   • Differentiation: The entire pipeline—from interpolation to matrix assembly, linear solve, and likelihood evaluation—is composed of automatically differentiable operations. Reverse-mode automatic differentiation (AD) is applied through the linear solve using implicit differentiation. This allows the calculation of the exact likelihood gradient $\nabla_\Omega \mathcal{L}$ with respect to all parameters at the cost of only one additional forward solve, regardless of the parameter dimension $N$.
   • Sampling: The exact gradients enable the use of Hamiltonian Monte Carlo (HMC) and the No-U-Turn Sampler (NUTS), which scale efficiently with dimensionality, unlike gradient-free mixers.

Key Contributions

• Operator-Level Differentiability: The construction is operator-level, requiring only that the parameter-dependent operator varies smoothly enough to admit a low-rank SVD decomposition. It does not require modifying the underlying physics solver or deriving analytic derivatives of the legacy code.
• Exact Gradients via Implicit Differentiation: The paper demonstrates how to obtain exact likelihood gradients for a linear solve within a differentiable framework, avoiding the memory overhead of naive back-propagation through iterative solvers.
• Integrated Workflow: DREAM combines reduced-basis emulation (for speed) with automatic differentiation (for gradients), creating a workflow where the emulator's accuracy is directly quantified against the legacy code, and the gradients are exact with respect to the emulator.

Results: d + $^{58}$Ni Elastic Scattering
The framework was demonstrated on a CDCC analysis of deuteron ($d$) + $^{58}$Ni elastic scattering at 21.6 MeV, involving 18 optical-potential parameters.

• Performance: NUTS sampling converged on a single GPU in approximately 27 minutes with 24,000 post-warmup samples and zero divergent transitions. In contrast, a gradient-free sampler starting from the same point failed to converge correctly, producing biased marginal means (up to 3.4$\sigma$) while failing to trigger convergence alarms.
• Accuracy: The mean emulator error was found to be more than an order of magnitude smaller than the inferred model discrepancy ($\beta \approx 0.076$). Consequently, the posterior distribution is governed by the reaction model physics rather than the surrogate approximation.
• Physics Findings: The calibrated posterior constrained the average deuteron surface absorption to $\bar{W}_d = 11.7 \pm 0.8$ MeV, roughly 40% higher than the Koning-Delaroche systematics. It also supported a substantial proton/neutron asymmetry in the surface absorption, though the authors note this is a model-dependent extraction requiring multi-energy data for confirmation. The model discrepancy was inferred to be $\sim 8\%$, well below the experimental uncertainty.

Significance and Claims
The paper claims that the primary bottleneck for full Bayesian calibration is not the cost of a single solve, but the absence of exact likelihood gradients in legacy operators. DREAM resolves this by making expensive, non-affine operators differentiable without rewriting the underlying physics code.

The authors position this work as a methodological proof of concept. They explicitly state that the CDCC application demonstrates the framework is "workable end-to-end" for calibrations previously inaccessible to gradient-based inference. The paper does not claim the specific physics results (e.g., the asymmetry) are final interpretations; rather, they are presented as a "representative payoff" showing that the framework can sharpen multi-energy datasets into full physics interpretations. The authors emphasize that the construction is portable to any model where the parameter-dependent operator admits low-rank SVD compression, including multi-energy and multi-observable optical-potential calibrations, though they acknowledge that applying this to high-dimensional nuclear structure matrix elements (e.g., with non-affine momentum cutoffs) remains an open question for future work.