Bayesian inference and uncertainty quantification for modeling of body-centered-cubic single crystals

This paper employs Bayesian model calibration and global sensitivity analysis to quantify uncertainties and identify key physical mechanisms in two distinct constitutive models for body-centered-cubic molybdenum single crystals, successfully validating their predictive capabilities across quasi-static to shock loading conditions.

The Gist

The Big Picture: Predicting the Unpredictable

Imagine you are an engineer trying to predict how a piece of metal (specifically, a single crystal of Molybdenum) will behave when you hit it with a hammer. But this isn't just a gentle tap; it's a shockwave, like a bullet hitting a target or a meteor striking the earth.

The problem is that metals are messy. Inside them, there are billions of tiny defects called dislocations (think of them as "kinks" or "wrinkles" in the fabric of the metal). When you squeeze or hit the metal, these kinks move, multiply, and tangle. This movement is what makes the metal bend or break.

Scientists have built computer models to simulate this. But these models have many "knobs and dials" (parameters) that control how the kinks behave. The trouble is, we don't know the exact settings for these knobs. If we guess wrong, our computer simulation might say the metal will hold up, when in reality, it shatters.

This paper is about finding the right settings for these knobs and figuring out which knobs actually matter when things get extreme.

---

The Two Competing Theories (The Models)

The researchers tested two different "rulebooks" (models) for how these metal kinks behave:

1. Model 1: The "Traffic Cop" Approach.

   • The Idea: This model treats the moving kinks like cars on a highway. It counts exactly how many cars (mobile dislocations) are on the road and how fast they are going. It assumes that if you need to move a lot of metal quickly (high speed), you need a lot of cars moving fast.
   • The Catch: It tries to track every single car. It's very detailed but complicated.

2. Model 2: The "Crowd Control" Approach.

   • The Idea: This model is simpler. It doesn't count individual cars. Instead, it assumes the crowd moves based on a general "mood" (temperature) and a fixed "speed limit" (a pre-set constant). It assumes the crowd gets denser as they get tired (strain hardening), but it doesn't worry about the specific traffic jams.
   • The Catch: It's easier to use, but it might miss the details of how the crowd actually moves during a sudden panic.

---

The Detective Work: Bayesian Calibration (The "Tuning" Phase)

To figure out which rulebook is better, the researchers used a statistical detective method called Bayesian Model Calibration.

• The Analogy: Imagine you have a radio with 100 knobs, and you want to tune it to get the clearest signal. You don't just guess; you listen to the static, turn a knob slightly, listen again, and repeat.
• What they did: They fed the computer models real-world data (how Molybdenum actually reacted to being squeezed at different speeds and temperatures). The computer then "tuned" the knobs in both Model 1 and Model 2 until the simulation matched the real-world data as closely as possible.

Result: Surprisingly, both models did a great job matching the "normal" tests (like squeezing the metal slowly). They both looked like winners so far.

---

The Stress Test: Global Sensitivity Analysis (The "Which Knob Matters?" Phase)

Here is where the paper gets interesting. Just because a model fits the data doesn't mean it understands why it fits. The researchers asked: "If we wiggle a specific knob, does the result change?" This is called Sensitivity Analysis.

• The Analogy: Imagine baking a cake. You can get a good cake with slightly different amounts of sugar or flour. But if you forget the eggs entirely, the cake collapses. Sensitivity analysis tells you which ingredients are the "eggs" (critical) and which are just "sprinkles" (optional).

The Findings:

• Model 1 (Traffic Cop): The results were highly sensitive to the number of "cars" (dislocations). If the initial number of kinks was low, the model predicted a very sharp, sudden reaction to the shock.
• Model 2 (Crowd Control): The results were insensitive to the number of kinks. It assumed the crowd would move the same way regardless of how many people were there.

---

The Real-World Test: The Plate Impact (The "Shock" Phase)

To see which model was truly superior, they tested them against a Plate Impact Experiment. This is like firing a copper plate at a Molybdenum crystal at supersonic speeds.

• The Experiment: They used three types of Molybdenum targets:

  1. Pristine: Very few kinks (clean metal).
  2. Pre-strained: Some kinks (bent metal).
  3. Heavily Pre-strained: Lots of kinks (very bent metal).

• The Reality: In the real world, the "Pristine" metal (few kinks) reacted almost the same as the "Heavily Pre-strained" metal. The shockwave behaved consistently regardless of how many kinks were inside.

• Model 1's Failure: It predicted that the "Pristine" metal would react very differently (a huge spike in pressure) because it had so few kinks to move. It failed to match reality.

• Model 2's "Success" (Sort of): It predicted the same reaction for all three, which matched reality. BUT, it got there for the wrong reason. It ignored the kinks entirely, which is dangerous because in other scenarios, the kinks do matter.

The Verdict: Model 1 had the right physics (it knew kinks mattered) but was missing a piece of the puzzle. Model 2 got the right answer by accident (by ignoring the problem), which makes it unreliable for future predictions.

---

The Fix: Adding a "Nucleation" Mechanism

The researchers realized Model 1 was missing a crucial step. When the metal is hit with a massive shock, new kinks don't just move; they pop into existence out of nowhere (nucleation).

• The Analogy: Imagine a calm lake (pristine metal). If you throw a stone, ripples move. But if you hit it with a sledgehammer, the water doesn't just ripple; it explodes and creates new waves instantly. Model 1 only accounted for the ripples, not the explosion.

They added a "nucleation" term to Model 1. This allowed the model to create new kinks when the stress got too high.

• The Result: Suddenly, Model 1 worked perfectly! It predicted that even the "Pristine" metal would behave like the others because the shock created new kinks instantly, leveling the playing field.

---

The Takeaway

1. Don't trust a model just because it fits the data. You need to understand why it fits. (Model 2 looked good but was "lucky").
2. Uncertainty is key. We need to know which parts of our models are shaky.
3. Physics matters. To predict extreme events (like explosions or meteor impacts), you can't just use simple rules. You need to account for complex behaviors like the sudden creation of new defects (nucleation).

In short: The paper used advanced statistics to tune two computer models of metal. They found that to predict how metal behaves under extreme shock, you must account for the fact that the metal can instantly create new "defects" to handle the stress, not just move the ones it already has.

Technical Summary

1. Problem Statement

Predictive modeling of body-centered-cubic (bcc) refractory metals, specifically molybdenum, under extreme loading conditions (from quasi-static to shock loading) faces significant challenges due to:

• Complex Deformation Mechanisms: The interplay of slip activity, dislocation interactions, and the breakdown of the standard Schmid law in bcc crystals.
• High-Dimensional Parameter Uncertainty: Incorporating multiple deformation mechanisms (e.g., thermal activation, phonon drag, dislocation density evolution) increases the number of model parameters, leading to strong correlations and substantial uncertainties.
• Limitations of Calibration: Traditional Bayesian Model Calibration (BMC) can identify parameters that fit data but often fails to diagnose why a model fails to reproduce experimental data or to identify the specific physical mechanisms causing discrepancies, especially when extrapolating to unexplored regimes (e.g., shock loading).

2. Methodology

The authors employed a rigorous statistical framework combining Bayesian Model Calibration (BMC) and Variance-Based Global Sensitivity Analysis (GSA) to evaluate two distinct physics-based bcc single crystal plasticity models.

A. The Models

Two models were compared, both based on continuum crystal plasticity but prioritizing different physical assumptions:

1. Model 1 (Nguyen et al. 2021a): Incorporates mobile dislocation kinetics. It explicitly couples the plastic slip rate to the mobile dislocation density via the Orowan relation. It accounts for the transition from thermally-activated glide to a phonon drag-dominated regime at high strain rates, including mobile-immobile dislocation transitions.
2. Model 2 (Lee et al. 2023b): Based on thermally-activated glide with a fixed pre-exponential factor ($\dot{\gamma}_0$). It utilizes dislocation-density-based hardening laws informed by dislocation dynamics simulations but does not explicitly couple mobile dislocation density to slip rate in the same manner as Model 1.

B. Bayesian Model Calibration (BMC)

• Data: Experimental stress-strain data covering a wide range of temperatures (195–500 K), strain rates ($10^{-5}$ to $300 \text{ s}^{-1}$), and crystallographic orientations ($\langle 100 \rangle, \langle 110 \rangle, \langle 111 \rangle$).
• Procedure: Used the SEPIA package (Simulation-Enabled Prediction, Inference, and Analysis). A Gaussian Process (GP) emulator was trained on 100 parameter samples to mimic the computationally expensive crystal plasticity simulations.
• Inference: Markov Chain Monte Carlo (MCMC) sampling was used to infer posterior distributions for key parameters (e.g., dislocation multiplication coefficients, activation energies, reference strain rates).

C. Global Sensitivity Analysis (GSA)

• Method: Used pyBASS (Bayesian Adaptive Spline Surfaces) to perform functional GSA.
• Goal: To decompose the output variance (stress/velocity) into contributions from individual parameters and their interactions. This was applied to both uniaxial stress responses and plate impact simulations to identify which physical mechanisms dominate under different loading conditions.

D. Validation: Plate Impact Tests

The calibrated models were validated against plate impact experiments (Kanel et al. 2022) on [100] single crystal molybdenum with varying initial dislocation densities (pristine, 0.6% prestrained, 5.4% prestrained). The simulations predicted free surface velocity profiles to assess the elastic-plastic transition and Hugoniot Elastic Limit (HEL).

3. Key Results

A. Calibration Performance

• Both models successfully reproduced the experimental uniaxial stress-strain data across the calibration regime with similar accuracy (RMSD $\approx$ 55–62 MPa).
• However, the posterior distributions revealed that parameters governing high-strain-rate behavior in Model 1 (e.g., $C_M$, $\rho_{0,sat}$) had wider uncertainties due to a lack of high-strain-rate calibration data.

B. Sensitivity Analysis Insights

• Model 1: The initial yield and early hardening are highly sensitive to the dislocation multiplication coefficient ($C_M$) and the evolution of mobile dislocation density. At high strain rates, the model relies heavily on the Orowan relation and drag-limited velocity.
• Model 2: The initial response is controlled primarily by the fixed pre-exponential factor ($\dot{\gamma}_0$). It is largely insensitive to the evolution of dislocation density at the onset of yield.
• Mechanistic Difference: Model 1 treats mobile dislocations as active mediators of plastic strain accommodation, whereas Model 2 uses dislocation density primarily for strain hardening.

C. Plate Impact Validation & Discrepancies

• Model 1 Failure: While Model 1 captured the general trend, it failed to reproduce the invariance of the Hugoniot Elastic Limit (HEL) amplitude with respect to sample thickness. It predicted a significant attenuation of the elastic precursor in pristine (low dislocation density) samples, which contradicts experimental observations.
• Model 2 Failure: Model 2 showed almost no sensitivity to initial dislocation density, failing to capture the distinct elastic-plastic transition behaviors observed in pristine vs. prestrained samples.
• Root Cause: The GSA revealed that the discrepancy in Model 1 was driven by the rate of mobile dislocation evolution. The model lacked a mechanism to rapidly generate dislocations at the shock front to prevent excessive stress relaxation.

D. Model Improvement

• To address the HEL attenuation issue in Model 1, the authors added a stress-dependent dislocation nucleation term to the mobile dislocation evolution equation.
• Result: This addition successfully captured the invariance of the HEL amplitude across different sample thicknesses by allowing rapid stress relaxation near the impact surface without excessive attenuation over distance.

4. Key Contributions

1. Systematic UQ Framework: Demonstrated a robust workflow combining BMC and functional GSA to not only calibrate parameters but also diagnose the physical origins of model failure in high-dimensional spaces.
2. Mechanistic Differentiation: Clearly distinguished the predictive capabilities of two distinct bcc plasticity models, showing that mobile dislocation kinetics (Orowan relation + drag limit) are critical for capturing shock loading responses, which simple thermally-activated models (Model 2) miss.
3. Identification of Missing Physics: Used sensitivity analysis to pinpoint that dislocation nucleation (or depinning) is a necessary mechanism to accurately model the transient elastic-plastic transition in pristine bcc crystals under shock loading.
4. Quantification of Uncertainty Sources: Showed that uncertainties in high-strain-rate predictions stem largely from data sparsity in calibration regimes and the specific sensitivity of the model to dislocation multiplication rates.

5. Significance

• Advancing Predictive Modeling: The study highlights that accurate modeling of extreme loading conditions in bcc metals requires more than just fitting stress-strain curves; it necessitates capturing the correct kinetics of dislocation generation and motion.
• Guidance for Future Models: The work suggests that treating the pre-exponential factor ($\dot{\gamma}_0$) as a constant is an oversimplification for shock loading. Future models must incorporate mechanisms for stress-induced dislocation nucleation or depinning to handle the elastic-plastic transition in pristine materials.
• Methodological Impact: The integrated UQ approach provides a blueprint for developing continuum models that are not only calibrated but also physically robust across a wide range of loading conditions, from quasi-static to shock, and offers a pathway to incorporate machine learning surrogates and non-Schmid effects in future research.