🔬 materials science

Posterior Inference of Hamiltonian Parameters from RIXS Spectroscopy

This paper introduces a novel simulation-based inference framework utilizing neural ratio estimation and conditional flow matching, enhanced by a physics-aware vision transformer, to efficiently infer full posterior distributions of Hamiltonian parameters from RIXS spectroscopy data, thereby enabling advanced uncertainty quantification and experimental design for transition metal compounds.

Original authors: Samuel Klein, Thomas M. Linker, Louis Conreux, Daniel Ratner, Apurva Mehta, Makoto Tachibana, Jiemin Li, Jonathan Pelliciari, Valentina Bisogni, Wei He, Xiangpeng Luo, Mark P. M. Dean, Marton K. Lajer

Published 2026-08-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Samuel Klein, Thomas M. Linker, Louis Conreux, Daniel Ratner, Apurva Mehta, Makoto Tachibana, Jiemin Li, Jonathan Pelliciari, Valentina Bisogni, Wei He, Xiangpeng Luo, Mark P. M. Dean, Marton K. Lajer, Michael Kagan, Joshua J. Turner, Yongqiang Cheng, Sean Gasiorowski

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 inside a tiny, invisible world. The crime scene is a piece of a special material, and the clues are hidden in a complex map of light. This field of science is called quantum materials research, where scientists study how electrons dance and interact to create things like magnetism or superconductivity. To understand these dances, researchers use a powerful tool called Resonant Inelastic X-ray Scattering (RIXS). Think of RIXS as a high-speed camera that fires X-ray photons at a material. When a photon hits an electron, it bumps it into a higher energy state, and then the electron falls back down, spitting out a new photon. By measuring the energy and direction of this "spit-back" photon, scientists get a 2D map of the material's internal energy levels.

However, reading this map is like trying to guess the recipe of a cake just by looking at a blurry photo of it. The photo (the spectrum) is the result of many hidden ingredients (the physical parameters) mixing together in a complicated way. For decades, scientists have had to manually tweak these ingredients, one by one, hoping the resulting photo looks like the real one. It's a slow, tedious process that often leaves you guessing about how the ingredients relate to each other. This paper introduces a new, smarter way to crack the code, turning a guessing game into a precise science that reveals not just the ingredients, but exactly how they depend on one another.

The New Detective: Teaching AI to Read the Light Map

This paper presents the first time scientists have used a technique called "simulation-based inference" to decode these complex light maps from RIXS experiments. Instead of manually tweaking the recipe, the authors trained a smart computer system to learn the relationship between the hidden ingredients (Hamiltonian parameters) and the resulting light map. They did this by creating a massive library of simulated maps, teaching the AI to recognize patterns, and then asking it to work backward from real experimental data to figure out the most likely ingredients.

The team tested this new method on two different nickel-based compounds: NiPS3, which is a bit more "covalent" (where atoms share electrons like best friends), and K2NiF4, which is more "atomic" (where atoms hold onto their electrons more tightly). These two materials represent opposite ends of the spectrum for how these quantum materials behave.

The Secret Weapon: A Specialized AI Eye

One of the biggest challenges was teaching the AI to look at the data correctly. Standard image-recognition AI (like the kind that identifies cats in photos) treats an image as a grid of square patches. But a RIXS map isn't just a picture; it has two very different axes. One axis represents the energy of the incoming X-ray, and the other represents the energy lost by the electron. The authors realized that squaring these up like a photo was the wrong approach.

Instead, they built a custom AI eye called a "column Vision Transformer." Imagine the light map as a stack of vertical slices. Each slice represents a specific incoming energy and contains a full spectrum of what happened at that moment. The AI looks at these vertical slices as whole units, rather than chopping them into squares. This "column" approach turned out to be a game-changer. It allowed the AI to understand the physical layout of the experiment much better, leading to sharper, more accurate guesses about the material's properties with fewer simulations than other methods would need.

Cracking the Code: What They Found

When the team applied this new method to real data from NiPS3 and K2NiF4, they didn't just get a single "best guess" for the ingredients. They got a full "probability map" (a posterior distribution) for every single parameter. This is a huge leap forward.

Seeing the Invisible Connections: Traditional methods give you a list of numbers, like "Ingredient A is 5, Ingredient B is 10." But they miss the secret handshake between them. The new method revealed that some parameters are deeply linked. For example, the "core spin-orbit coupling" (how an electron spins) and the "incident energy offset" (a tiny calibration shift) were found to be correlated. If one goes up, the other must go down to keep the light map looking the same. This kind of relationship is invisible to older methods but is clearly visible in the new AI's probability map.
Handling the "Fuzziness": Real-world measurements are always a bit blurry due to instrument limitations. The team treated this blurriness (called "energy-loss broadening") as a "nuisance parameter." Instead of guessing a single fixed value for the blur, the AI learned to average over all possible blur levels. This means the final answer for the material's properties is robust and doesn't depend on a lucky guess about how blurry the camera was.
The Results: For the covalent material (NiPS3), the AI's results matched up well with previous expert guesses and automated optimization attempts, but with the added benefit of showing the uncertainties and correlations. For the atomic material (K2NiF4), the method successfully identified how the material's internal forces differed from a pure, isolated atom, revealing the subtle effects of the crystal lattice.

Why This Matters

The most exciting part of this work is that once the AI is trained, it can instantly analyze new data without needing to run expensive simulations again. This "amortized" learning unlocks new possibilities for scientists. They can now combine data from multiple experiments to get an even clearer picture, or use the AI to design the next experiment to learn the most.

The authors are careful to note that their method relies on a specific model (the single-ion model) which simplifies some complex physics. For instance, in the NiPS3 material, there is a feature called a "Hund's exciton" that their model doesn't fully capture, which is why the AI's predictions aren't perfect in every single detail. However, the method successfully recovered the main features of the spectra and provided a rigorous, statistical way to understand the material's Hamiltonian parameters.

In short, this paper doesn't just give a better answer; it changes the question. It moves the field from "What is the single best number?" to "What is the full story of how these numbers relate, and how sure are we?" By using a specialized AI that respects the physics of the experiment, the authors have opened the door to a new era of precise, uncertainty-aware quantum material discovery.

Technical Summary: Posterior Inference of Hamiltonian Parameters from RIXS Spectroscopy

Problem Statement
Resonant inelastic X-ray scattering (RIXS) is a critical tool for probing quantum materials, providing access to charge, spin, and orbital excitations. However, interpreting RIXS spectra to extract the underlying semi-empirical Hamiltonian parameters is a challenging inverse problem. The forward map from parameters to spectra is nonlinear, computationally expensive (requiring codes like EDRIXS), and often non-unique, where multiple parameter sets yield similar spectra. Historically, parameter extraction has relied on expert manual tuning or Bayesian optimization minimizing an $L_1$ spectral distance. These approaches typically yield point estimates with univariate confidence intervals that fail to capture correlations between parameters and lack formal uncertainty quantification. Furthermore, they often fix nuisance parameters (such as energy-loss broadening) to single point estimates, potentially biasing the results.

Methodology
The authors present the first application of Simulation-Based Inference (SBI) to RIXS spectroscopy. The framework replaces the intractable likelihood function with an explicit statistical model to infer the full joint posterior distribution $p(\theta | x_{obs})$ over Hamiltonian parameters $\theta$ given an observed spectrum $x_{obs}$ . The methodology proceeds in two stages:

Truncated Marginal Neural Ratio Estimation (TMNRE): To address the high computational cost of RIXS simulations, the authors employ TMNRE to efficiently restrict the prior distribution. This stage iteratively trains binary classifiers for each marginal parameter to identify and truncate regions of the parameter space with negligible posterior support, concentrating the simulation budget on high-probability regions.
Joint Density Estimation via Conditional Flow Matching: Once the prior is restricted, a joint density estimator is trained on simulations drawn from this restricted region. The authors utilize a conditional flow matching model, which learns a time-dependent velocity field to transport samples from a Gaussian base distribution to the posterior.
- Encoder Architecture: A key innovation is the use of a Column Vision Transformer (Column-ViT). Unlike standard image encoders that treat the RIXS map as a generic 2D image, the Column-ViT tokenizes the spectrum along the incident-energy axis (columns). This design respects the physical layout where each column represents a fixed incident energy probing a subset of resonant transitions, leading to better-covered and sharper posteriors.
- Nuisance Marginalization: The energy-loss broadening parameter $\sigma$ is treated as a nuisance parameter. The density estimator is conditioned on $\sigma$ , allowing the final posterior to be marginalized over the prior of $\sigma$ ( $p(\theta | x_{obs}) = \int q_\phi(\theta | x_{obs}, \sigma) p(\sigma) d\sigma$ ) without additional computational cost during inference.

Key Contributions and Results
The method is applied to two Ni $^{2+}$ compounds with distinct bonding characters: NiPS $_3$ (covalent) and K $_2$ NiF $_4$ (atomic).

Full Posterior Recovery: The approach successfully recovers full joint posteriors for 11 Hamiltonian parameters (including Slater integrals, spin-orbit couplings, and crystal field splitting) with a modest simulation budget (~108,000 evaluations), comparable to previous optimization-based studies.
Correlation Discovery: Unlike point estimators, the joint posterior reveals significant parameter correlations invisible to other methods. Notably, a strong diagonal correlation is observed between the core spin-orbit coupling ( $\zeta_c$ ) and the incident energy offset ( $\Delta\omega_{in}$ ), reflecting a degeneracy in the forward model where these parameters compensate for one another.
Posterior Predictive Consistency: The posterior predictive distribution closely matches the observed spectra for both materials. While the posterior mean parameters reproduce the main spectral features, the predictive bands demonstrate that the forward model possesses significant degeneracy; broad marginals in individual parameters can still yield a tight predictive band.
Validation: The authors validate the method using Tests of Accuracy with Random Points (TARP), confirming that the ensemble posterior has correct coverage. The results are consistent with reference solutions from Lajer et al. (Bayesian optimization) and He et al. (hand-fit) where parameters are well-constrained, but provide richer uncertainty quantification.
Robustness to Nuisance Parameters: Sensitivity analysis shows that the inferred Hamiltonian parameters are largely independent of the assumed energy-loss broadening $\sigma$ , confirming that the spectral features genuinely constrain the electronic structure rather than relying on implicit assumptions about broadening.

Significance and Claims
The paper claims that this work establishes a new paradigm for RIXS analysis by shifting from point estimates to full posterior distributions. The significance lies in three main areas:

Uncertainty Quantification: It provides a rigorous framework for quantifying uncertainty that naturally marginalizes over nuisance parameters, avoiding biases introduced by fixing instrumental parameters.
Amortized Inference: Once trained, the model can be applied almost instantly to new experimental datasets of materials sharing the same prior structure, enabling analyses previously unavailable, such as multi-measurement posterior fusion and active experimental design.
Physical Insight: By capturing parameter correlations and degeneracies, the method offers a more complete picture of the constraints imposed by the data, revealing that certain parameter combinations are constrained while others remain degenerate.

The authors acknowledge limitations, including the reliance on the single-ion EDRIXS model (which neglects charge-transfer physics and ligand degrees of freedom) and the assumption of negligible shot noise. However, they argue that the amortized SBI approach offers a scalable, auditable, and statistically robust alternative to traditional optimization techniques for extracting Hamiltonian parameters from complex spectroscopic data.

More like this