🔬 materials science

Towards real-time surrogate-free Bayesian inversion for neutron reflectometry

This paper introduces a rapid, surrogate-free Bayesian inversion framework for neutron reflectometry that leverages exact gradients to enable highly efficient Hamiltonian Monte-Carlo and Variational Inference methods, offering state-of-the-art uncertainty quantification speeds and robustness for complex samples without sacrificing physical intuition.

Original authors: Max D. Champneys, Andrew J. Parnell, Philipp Gutfreund, Maximilian W. A. Skoda, Patrick A. Fairclough, Timothy J. Rogers, Stephanie L. Burg

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Max D. Champneys, Andrew J. Parnell, Philipp Gutfreund, Maximilian W. A. Skoda, Patrick A. Fairclough, Timothy J. Rogers, Stephanie L. Burg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you can only see the shadows cast by the culprit, never the culprit themselves. This is the daily reality for scientists who study the invisible world of surfaces and thin films. They use a technique called Neutron Reflectometry, which is like shining a super-powerful flashlight made of neutrons at a material. When these neutrons bounce off the surface, they create a pattern of ripples and waves. This pattern is the "shadow" that tells the scientists about the material's hidden layers, such as how thick they are or how dense they are. However, there's a catch: the shadow doesn't tell the whole story. The scientists have to work backward from the pattern to figure out what the layers actually look like. This is called an "inverse problem," and it's notoriously difficult, like trying to guess the ingredients of a cake just by tasting the crumbs.

For a long time, solving this puzzle has been slow and frustrating. The standard tools used by scientists are like trying to find a needle in a haystack by poking around randomly; they take a long time to converge on the right answer and often get stuck guessing the wrong things. Recently, some scientists tried to speed things up by training computer programs (artificial intelligence) to guess the answer instantly. But this is like hiring a magician who knows the answer but doesn't understand why it's the answer; you lose the physical intuition of how the layers actually work. The big question has been: Can we get the speed of the AI without losing the deep understanding of the physics?

This paper, titled "Towards Real-Time Surrogate-Free Bayesian Inversion for Neutron Reflectometry," proposes a clever new way to solve this puzzle. The authors, a team of researchers from the University of Sheffield and international labs, argue that we don't need to replace the physics with a "black box" AI guesser. Instead, they built a new mathematical toolkit that lets computers calculate the exact "slope" or "direction" of the solution instantly. Think of it like being lost in a foggy mountain range. The old way was to wander around randomly until you stumbled upon the peak. The new way uses a GPS that not only shows you where you are but also tells you exactly which way is "uphill" at every single step.

The paper demonstrates that by using these "gradients" (the mathematical equivalent of that GPS direction), scientists can solve the inverse problem much faster and more accurately than before. They tested their method on two very different challenges. First, they looked at a thick quartz film, a relatively simple case. Here, their new method found the answer in a fraction of the time it usually takes, and it was more precise than previous attempts. Second, they tackled a much harder problem: analyzing four different layers of organic light-emitting diode (OLED) devices, which are the screens in some modern electronics. These devices have complex, messy layers that change when heated. The team showed that their method could figure out the properties of all four devices simultaneously in less than 20 seconds, a task that would have taken hours or days with older methods.

Crucially, the authors didn't just find the "best" answer; they also calculated the "uncertainty" of that answer. This is like the detective not only saying "the culprit is in the kitchen" but also saying, "I'm 95% sure, but there's a small chance they're in the pantry." Their new approach, called Variational Inference, can do this uncertainty calculation in seconds, whereas the old methods might take hours. The paper suggests that this speed could eventually allow scientists to analyze data in real-time while experiments are happening, perhaps even adjusting the experiment on the fly. To help others use this new speed, the team has released their code as a free, open-source library called "refjax," allowing anyone to use these fast, gradient-based tools to study thin films without needing to be a math genius.

Technical Summary: Towards Real-Time Surrogate-Free Bayesian Inversion for Neutron Reflectometry

Problem Statement

Neutron reflectometry (NR) is a critical technique for characterizing surfaces and interfaces at the nanometer scale. However, NR is an indirect measurement technique; extracting physical properties (layer thickness, scattering length density, roughness) requires solving an inverse problem. Current approaches to this problem face significant limitations, particularly regarding uncertainty quantification (UQ) and computational efficiency:

Inefficiency in Complex Systems: Traditional sampling methods like Markov-Chain Monte-Carlo (MCMC) suffer from poor sample efficiency and slow convergence times, especially as the number of unknown parameters increases or when dealing with complex multilayer structures (e.g., organic LED devices).
Loss of Physical Intuition: Recently proposed surrogate machine learning models attempt to bypass the inverse problem by training neural networks to predict parameters directly from data. While fast, these approaches replace governing physical equations with black-box approximations, inevitably discarding physical intuition and requiring large, representative training datasets.
Lack of Gradient Utilization: Despite the revolution in automatic differentiation (AD) and gradient-based optimization in other fields of machine learning, NR software has largely remained reliant on gradient-free techniques (e.g., Levenberg-Marquardt, differential evolution) or heuristic methods, failing to leverage the exact gradients of the forward reflectivity model.

Methodology

The authors propose a surrogate-free Bayesian inversion framework that utilizes automatic differentiation (AD) to compute exact gradients of the forward specular reflectivity model with respect to unknown parameters. This enables the application of high-performance gradient-based inference schemes previously unexploited in NR.

Core Components

Forward Model & Automatic Differentiation:
- The forward reflectivity is calculated using the Abeles matrix formalism, a standard physical model for thin films.
- Instead of approximating this model with a neural network, the authors implement the calculation in JAX, a Python library that supports just-in-time (JIT) compilation, parallelism, and automatic differentiation.
- This allows for the exact computation of $\nabla_\theta \hat{R}(Q, \theta)$ , the gradient of the predicted reflectivity with respect to the parameters $\theta$ (slab and instrument parameters).
Gradient-Based Optimization:
- The framework employs stochastic gradient descent (specifically the ADAM optimizer) to minimize error metrics (e.g., $\chi^2$ ).
- This approach is shown to be highly effective for finding optimal parameters in high-dimensional spaces, avoiding local minima more effectively than traditional gradient-free methods.
Hamiltonian Monte-Carlo (HMC) for Inference:
- To perform full Bayesian inference, the authors utilize HMC (specifically the No-U-Turn Sampler, NUTS).
- HMC constructs a pseudo-dynamical system where the gradient of the log-posterior guides the sampling trajectory.
- By leveraging the exact gradients from the forward model, HMC achieves significantly higher sample efficiency (Effective Sample Size) compared to gradient-free MCMC, requiring far fewer iterations to converge.
Variational Inference (VI) for Approximate UQ:
- For scenarios requiring rapid inference on large datasets or complex models, the authors employ Variational Inference.
- VI transforms the inference problem into an optimization problem by approximating the true posterior $p(\theta|R)$ with a parametric distribution $q(\theta; \phi)$ .
- The parameters $\phi$ are optimized to maximize the Evidence Lower Bound (ELBO), utilizing the gradients of the forward model. This enables approximate UQ on the order of seconds rather than hours.

Software Implementation

The authors released refjax, an open-source Python library built on the JAX ecosystem. It provides:

Standardized interfaces for forward reflectivity calculations (including smearing kernels).
Native support for automatic differentiation, allowing users to apply any JAX-compatible optimizer or sampler (e.g., ADAM, HMC, VI) directly to NR data.

Key Results

The methodology was validated through two case studies:

Crystalline Quartz Film (Benchmark):
- Optimization: Using ADAM, the method achieved a $\chi^2$ error of 1.302, slightly outperforming the previously reported value of 1.32 for the same dataset, despite all parameters being free to fit.
- Inference (HMC): The NUTS sampler converged efficiently with only 2,000 samples per chain. The posterior mean achieved a $\chi^2$ of 1.22.
- Efficiency Comparison: When compared to a gradient-free Sample-Adaptive MCMC (SA-MCMC), HMC demonstrated superior convergence. The Effective Sample Size (ESS) for HMC increased monotonically, whereas SA-MCMC showed signs of slow mixing and decreased ESS after initial exploration. HMC reached the "well-converged" threshold (ESS $\approx$ 200) much faster than SA-MCMC.
Organic LED (OLED) Multilayer Devices (High Complexity):
- Scenario: A co-fit of four different OLED devices (varying annealing temperatures) involving 59 unknown parameters (layer thicknesses, SLDs, roughnesses, and shared instrument parameters).
- Variational Inference: The VI approach successfully inferred the joint posterior distribution.
- Performance: Each inference run completed in under 20 seconds on standard hardware.
- Physical Insights: The model successfully identified physical trends, such as increased SLD in the TFB layer for annealed samples (indicating solvent removal and density equilibration) and changes in interfacial widths. The inferred Silicon backing SLD (2.09 $\pm$ 0.0178) closely matched the literature value (2.07), validating the physical plausibility of the solution.

Significance and Claims

The paper claims a "step-change" in the speed and efficiency of NR inference without sacrificing physical fidelity.

Surrogate-Free Accuracy: Unlike ML surrogates, this approach uses the exact governing equations, preserving physical intuition and eliminating the need for large, labeled training datasets.
Computational Efficiency: The integration of AD enables:
- Order-of-magnitude gains in sample efficiency via HMC compared to traditional MCMC.
- Rapid approximate UQ via Variational Inference, reducing inference time from hours to seconds.
Scalability: The framework is designed to handle high-dimensional parameter spaces and complex, multi-device co-fitting scenarios that are computationally prohibitive for traditional methods.
Future-Proofing: By providing an open-source library (refjax), the authors aim to bridge the gap between NR and modern ML/optimization techniques, facilitating real-time analysis for high-throughput and kinetic experiments.

The authors position this work as a foundational step toward enabling real-time, surrogate-free Bayesian inversion, ensuring that NR data analysis can keep pace with the rapid advances in experimental throughput and machine learning.