🔬 materials science

Bayesian model comparison of type-I and type-II ultrafast demagnetization dynamics

This study employs Bayesian model comparison to demonstrate that finite experimental temporal resolution and noise significantly obscure the intrinsic differences between type-I and type-II ultrafast demagnetization dynamics, often rendering their statistical discrimination inconclusive and highlighting the sensitivity of such classifications to experimental conditions.

Original authors: Hiroki Wadati, Tetsuro Ueno

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hiroki Wadati, Tetsuro Ueno

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 trying to identify two different types of raindrops falling on a window. One type (let's call it Type-I) makes a single, smooth splash. The other type (Type-II) makes a quick splash followed immediately by a slower, secondary ripple.

In the world of ultrafast physics, scientists study how magnets lose their magnetism in a fraction of a second after being hit by a laser. They have long tried to sort these events into "Type-I" (one-step) or "Type-II" (two-step) categories. However, this new paper argues that sorting them is much harder than it looks, and often, the "weather" (the experiment) makes it impossible to tell the difference.

Here is a simple breakdown of what the authors, Hiroki Wadati and Tetsuro Ueno, discovered:

1. The Problem: The "Fuzzy Camera" Effect

Think of your experimental equipment as a camera trying to take a picture of a very fast event.

The Reality: The magnet might actually be doing a complex two-step dance (Type-II).
The Blur: Your camera has a "shutter speed" that isn't fast enough (finite temporal resolution) and the lens is a bit grainy (noise).
The Result: When you look at the photo, the complex two-step dance looks suspiciously like a simple one-step splash. The blur smears out the details, making the two different types of magnets look almost identical.

The authors found that this "blur" is so strong that it creates a large "gray zone" where you simply cannot tell if you are looking at a Type-I or Type-II event, even if you know the truth beforehand.

2. The Old Way: Guessing by Eye

Previously, scientists would look at their data graphs and try to fit a curve to them.

The Flaw: It's like looking at a blurry photo and saying, "That looks like a cat," or "That looks like a dog." You might be right, but you might also be wrong.
The Trap: Because the data is blurry, a simple "one-step" model often fits the data just as well as a complex "two-step" model. If you just look at the lines, you can't be sure which one is the real story.

3. The New Way: The "Math Detective" (Bayesian Model Comparison)

Instead of just guessing, the authors used a statistical tool called Bayesian Model Comparison. Think of this as a math detective that asks two questions:

How well does the story fit the evidence? (Does the curve match the dots?)
Is the story too complicated? (Do we really need a complex explanation, or is a simple one enough?)

The detective uses a score (called the BIC) to decide.

If the data is clear and sharp, the detective can confidently say, "This is definitely a two-step event!"
If the data is blurry and noisy, the detective says, "I can't tell. The evidence isn't strong enough to pick a winner."

4. What They Found

The authors created fake data (simulations) to test this detective. They found:

The "Inconclusive Zone": There is a huge range of conditions where the data is so blurry that the detective refuses to make a call. In these cases, claiming a magnet is Type-I or Type-II is statistically meaningless.
The Real World Test: They applied this to real data from a material called NiCo2O4. The detective confirmed that this material does show a complex two-step behavior, but only because the data was clear enough to overcome the blur. If the experiment had been noisier, the result might have been inconclusive.

The Big Takeaway

The paper concludes that labeling ultrafast demagnetization as "Type-I" or "Type-II" isn't just about the material itself; it's also about how good your experiment is.

If your "camera" (experiment) is too blurry or too noisy, you might be looking at a complex two-step event but mistakenly calling it a simple one-step event. The authors suggest that scientists should stop just "eyeballing" their graphs and start using these statistical tools to admit when the data is too fuzzy to make a definitive claim.

In short: Don't trust your eyes when the picture is blurry. Use the math detective to tell you when you simply don't have enough information to choose a side.

Technical Summary: Bayesian Model Comparison of Type-I and Type-II Ultrafast Demagnetization Dynamics

Problem Statement
Ultrafast demagnetization dynamics, following femtosecond laser excitation, are phenomenologically classified into "type-I" (single-step) and "type-II" (two-step) responses. While type-I dynamics are common in elemental ferromagnets (e.g., Ni, Co) and type-II in rare-earth or ferrimagnetic systems, distinguishing between them is often ambiguous in practice. Finite experimental temporal resolution and noise can obscure intrinsic dynamics, causing distinct physical models to yield visually similar fits. Conventional analysis relies on least-squares fitting and visual inspection of residuals, which lacks a quantitative basis for comparing models with different structural complexities. This raises a fundamental question: under what experimental conditions can time-resolved data reliably discriminate between type-I and type-II dynamics, and when do experimental limitations render such classifications statistically inconclusive?

Methodology
The authors reformulate the classification problem as a Bayesian model-comparison task. The methodology involves three primary components:

Phenomenological Modeling: Two models are defined to describe intrinsic magnetization $M(t)$ . The Type-I model assumes a single demagnetization channel followed by recovery, while the Type-II model incorporates both fast and slow demagnetization components.
Instrumental Convolution: To account for finite temporal resolution, the intrinsic models are convolved with a Gaussian instrumental response function (IRF). The authors derive exact analytic expressions for these convolved signals, $y(t)$ , utilizing the complementary error function. This step is crucial as it mathematically demonstrates how the IRF smooths early-time curvature, reducing observable differences between the two models.
Synthetic Data and Bayesian Inference: Synthetic pump-probe datasets are generated from the Type-II model (ground truth) with added Gaussian noise and varying IRF widths. The authors employ the Bayesian Information Criterion (BIC) to compare the Type-I and Type-II models against the data. The BIC difference is used to calculate an approximate Bayes factor ( $K_{II/I}$ ), which quantifies the statistical preference for one model over the other while penalizing model complexity (Type-II has 5 parameters vs. 3 for Type-I).

Key Contributions

Analytic Formulation: The paper provides closed-form analytic expressions for Gaussian-convolved Type-I and Type-II demagnetization models, explicitly accounting for finite temporal resolution.
Quantitative Distinguishability Map: By sweeping IRF full width at half maximum (FWHM) and noise amplitude ( $\sigma_n$ ), the authors construct a "model preference map." This map identifies specific regimes where the data are statistically inconclusive (Bayes factor near zero), demonstrating that even intrinsically two-step dynamics can appear statistically consistent with single-step models under poor experimental conditions.
Application to Real Data: The framework is applied to experimental ultrafast demagnetization data from NiCo $_2$ O $_4$ thin films. The analysis quantitatively confirms a strong statistical preference for the Type-II model, consistent with previous literature, and validates the extracted time constants.

Results

Impact of Resolution and Noise: The study reveals that convolution with the IRF significantly reduces the observable differences between intrinsic responses. As temporal broadening (IRF width) and noise increase, the distinguishability between Type-I and Type-II models degrades, eventually leading to regimes where no model can be statistically favored.
Limitations of Least-Squares: In representative synthetic datasets, both Type-I and Type-II models produced visually similar fits to the noisy, convolved data. However, residual analysis showed systematic deviations for the Type-I model, which the BIC analysis correctly identified as statistically inferior despite the visual similarity.
Experimental Validation: Analysis of NiCo $_2$ O $_4$ data yielded a $\log_{10} K_{II/I}^{\text{BIC}}$ of 82.4, indicating overwhelming evidence for the Type-II model. This preference persisted even when the analysis window was restricted to early-time dynamics, suggesting the two-step nature is intrinsic to the demagnetization phase rather than just the recovery.

Significance and Claims
The paper argues that the distinction between Type-I and Type-II dynamics is inherently conditional on experimental resolution and noise levels, rather than being an absolute property of the material system. The authors claim that:

Probabilistic Assessment: Apparent discrepancies in the literature regarding demagnetization types may stem from differences in experimental resolution and analysis strategies rather than genuine physical inconsistencies.
Statistical Inconclusiveness: In regimes where the Bayes factor approaches unity, data do not support a definitive classification. The authors advocate for treating such cases as "statistically inconclusive" rather than forcing a binary classification.
Methodological Shift: The Bayesian framework replaces heuristic or qualitative classifications with a probabilistic assessment that naturally balances goodness-of-fit against model complexity (Occam's razor). This provides a transparent and reproducible basis for evaluating competing dynamical hypotheses in ultrafast magnetism.

The authors conclude that while their phenomenological models do not provide a microscopic description of demagnetization, the proposed framework is essential for determining the warranted level of model complexity for a given dataset and for interpreting ultrafast pump-probe experiments under realistic experimental constraints.