Causation-guided mechanism identification and interpretable reduced-order modeling of damage-driving grain-boundary stress in creep

This paper presents a causation-guided machine-learning framework that integrates crystal-plasticity simulations to identify key microstructural mechanisms governing grain-boundary stress in creep and distills them into an interpretable, robust reduced-order model for predicting damage-driving stress under complex loading conditions.

The Gist

Imagine a metal alloy, like the super-strong steel used in jet engines, as a giant mosaic made of millions of tiny, individual tiles called grains. When these engines run hot for a long time, the metal slowly stretches and deforms, a process called creep. Eventually, this leads to cracks forming along the lines where the tiles meet (the grain boundaries).

The big problem for engineers is that predicting exactly where and why these cracks start is incredibly hard. It's like trying to predict which specific tile in a mosaic will crack first, knowing that the pressure on that tile depends on the shape of the tile, the angle of the line next to it, the texture of the tile itself, and how its neighbors are pushing back. There are too many variables, and they all interact in complicated, non-linear ways.

This paper acts like a detective trying to solve that mystery. Here is how they did it, explained simply:

1. The Detective's Toolkit: "Causation Entropy"

Usually, scientists look at data and say, "These two things happen at the same time, so they must be related." But that's like seeing that ice cream sales and shark attacks both go up in July and concluding that ice cream causes shark attacks. They are just correlated, not causal.

The authors used a special mathematical tool called Causation Entropy. Think of this as a "truth filter." It asks: "If I already know everything else about this situation, does knowing this specific detail actually tell me anything new about where the stress is?"

They tested 18 different clues (like the angle of the grain boundary, how easily the metal slips, and how stiff the grains are). The filter sorted them to find the four "super-clues" that truly drive the stress:

1. The Angle: How tilted the grain boundary is relative to the force.
2. The Slip Pass: How easily the metal's internal "slip" can jump from one grain to the next.
3. The Creep Climb: A specific way the metal relaxes stress at high temperatures (like a slow-motion dance of atoms).
4. The Stiffness Mismatch: How different the "hardness" is between the two grains meeting at the boundary.

2. Building a Simple Map (Reduced-Order Modeling)

Once they found the four super-clues, they didn't just leave it there. They built a simple, easy-to-read map (a mathematical formula) that predicts the stress using only those four clues.

Imagine you have a massive, confusing encyclopedia of weather data. Instead of reading the whole book to predict rain, this team found that you only need to look at the barometer, the wind speed, the humidity, and the cloud shape to get it right 80% of the time. Their map is that simple, but it's built on the physics of the metal, not just a guess.

3. The "Stress Test" (Does it work in new situations?)

To make sure their map wasn't just a lucky guess for one specific scenario, they tested it in two new situations:

• Multiaxial Loading: Instead of pulling the metal in just one direction, they pulled it from multiple angles (like squeezing a stress ball from all sides).
  • Result: The map still worked! The four super-clues remained the most important, even though the forces were more complex.
• Tricrystal Systems: They added a third grain to the mix, creating a "junction" where three tiles meet.
  • Result: The original map started to struggle because it only looked at the immediate neighbors (local). It was like trying to predict traffic at a three-way intersection by only looking at two cars.
  • The Fix: They added a "neighborhood watch" feature to the map. By including information about the other grain boundaries nearby (non-local information), the map became accurate again. This showed that their method is flexible enough to grow when the situation gets more complex.

4. The "Black Box" vs. The "Glass Box"

The authors also tested their method against standard "Black Box" AI models (like complex neural networks). These AI models are great at guessing the answer but terrible at explaining why.

• When they fed the AI the original 18 clues, it was okay at guessing.
• When they fed the AI only the 4 super-clues (plus their simple mathematical shapes), the AI got much better at guessing.

This proves that their "truth filter" didn't just find random numbers; it found the actual physical ingredients that matter. It's like showing that a chef doesn't need 50 spices to make a great soup; they just need salt, pepper, garlic, and onions. If you give a robot chef just those four, it makes better soup than if you give it a bucket of random spices.

The Bottom Line

The paper doesn't claim to have built a new engine or cured a disease. Instead, it built a better way to understand and predict how metal fails under heat.

They took a messy, high-dimensional problem (too many variables) and distilled it down to a simple, interpretable story: The stress on a metal grain boundary is mostly about the angle, the slip, the climb, and the stiffness mismatch. By focusing on these four, they created a model that is accurate, easy to understand, and works even when the conditions change.

Technical Summary

Technical Summary: Causation-guided Mechanism Identification and Interpretable Reduced-Order Modeling of Damage-Driving Grain-Boundary Stress in Creep

Problem Statement
In polycrystalline superalloys, grain-boundary (GB) local stress is the primary driving force for the initiation and evolution of long-term creep damage, specifically intergranular cavitation. While classical continuum models treat GB stress as an effective mean field, real polycrystals exhibit spatially heterogeneous stress states arising from crystallographic anisotropy, orientation mismatch, and neighboring grain constraints. Although Crystal-Plasticity Finite Element (CPFE) methods can resolve these local fields, the high-dimensional, nonlinear, and strongly coupled relationships between GB stress and numerous microstructural characteristics (crystallographic, geometric, and micromechanical) make it difficult to identify the truly governing mechanisms. Traditional statistical correlation methods (e.g., Pearson correlation) fail to capture nonlinear dependencies, while standard machine learning (ML) approaches often provide model-based interpretability (e.g., SHAP values) that does not equate to physical causality. Consequently, there is a need for a framework that can distinguish direct causal mechanisms from incidental correlations to construct interpretable, reduced-order models for GB stress.

Methodology
The authors propose a causation-guided machine learning framework that integrates physics-based modeling with information-theoretic attribution. The methodology proceeds through the following steps:

1. Physics-Based Data Generation: Dislocation-climb-affected CPFE simulations are performed on minimal grain clusters (bicrystals and tricrystals) of Inconel 617 at 950°C. The model incorporates both dislocation glide and climb to capture steady-state creep behavior. Data is generated under uniaxial and multiaxial (high triaxiality) loading conditions.
2. Feature Space Construction: A candidate space of 18 physically motivated characteristics is defined. These include:
   • Crystallographic features: Mean orientation, misorientation angle, slip transmission factor, and Schmid-type indicators for both slip and climb.
   • Microstructural geometry: GB inclination angle.
   • Micromechanical properties: Elastic modulus mismatches (global and local, series and parallel configurations).
3. Causation Entropy Analysis: Instead of relying on simple correlation, the authors employ causation entropy (CE), an information-theoretic metric based on conditional entropy. CE quantifies the unique information a candidate variable contributes to the target (GB normal stress) after conditioning on all other variables. This allows for the identification of direct causal relevance and the filtering of redundant or indirect associations.
4. Reduced-Order Modeling: Based on the CE ranking, a library of candidate functions (including linear, polynomial, trigonometric, and saturation terms) is constructed. A linear regression model is trained using only the top-ranked candidate functions to create an interpretable reduced-order representation of the GB normal stress.
5. Validation and Transferability: The framework is validated by testing the transferability of the identified characteristics and functional forms from uniaxial bicrystal data to multiaxial bicrystal and tricrystal systems. Additionally, the extracted features are used as inputs for representative ML models (Random Forest, GBRT, GPR, MLP) to assess improvements in predictive performance across different model classes.

Key Results

• Dominant Mechanisms: The causation entropy analysis identified four dominant characteristics governing GB normal stress:
  1. GB Inclination Angle ($\phi_{GB}$): The most significant factor, consistent with classical stress transformation mechanics.
  2. Slip Transmission Factor ($m'$): Governs crystallographic compatibility and slip transfer across the boundary.
  3. Climb-Related Schmid-Type Indicator ($m_n$): Reflects the propensity for diffusion-assisted climb and stress relaxation, proving more influential than the conventional slip-based Schmid factor in creep.
  4. Elastic Modulus Mismatch: Specifically, the maximum in-plane Young's modulus mismatch, highlighting the role of local micromechanical contrast.
• Reduced-Order Model Performance:
  • For bicrystals under uniaxial loading, the CE-derived regression model achieved a test-set $R^2$ of 0.803, reducing the variance of the prediction residual by 93.7% compared to the raw data.
  • Transferability: The same functional form and characteristic hierarchy remained effective under multiaxial loading (high triaxiality, $\eta=3$) with a test-set $R^2$ of 0.804, demonstrating that the governing physical mechanisms are robust to changes in macroscopic stress state.
  • Nonlocal Extension: For tricrystal systems, a purely local description (using only bicrystal-derived features) resulted in underfitting ($R^2 = 0.590$). However, by introducing physically interpretable nonlocal features (averaged characteristics of neighboring GBs), the model was systematically extended, recovering predictive performance ($R^2 = 0.776$).
• Surrogate Model Enhancement: Using the top 18 CE-selected candidate functions as inputs significantly improved the predictive accuracy (lower RMSE, higher $R^2$) of all tested ML models (RF, GBRT, GPR, MLP) compared to using the original 18 raw characteristics. This confirms that the extracted functions encode governing information more efficiently.

Significance and Claims
The paper claims to establish a principled, information-theoretic route for causal-mechanistic identification in micromechanics. Its primary contributions are:

1. Interpretability: It moves beyond "black-box" ML or simple correlation to provide a physically transparent, reduced-order representation of GB stress driven by specific mechanisms (geometry, compatibility, climb, and mismatch).
2. Mechanism Distillation: It demonstrates that minimal grain clusters can serve as controlled vehicles for distilling governing mechanisms that remain valid across complex loading conditions and grain configurations.
3. Scalability and Extensibility: The framework successfully handles the transition from local (bicrystal) to nonlocal (tricrystal) interactions by systematically augmenting the feature space with physically meaningful terms, rather than requiring a complete retraining of the model structure.
4. Efficiency: The extracted candidate functions serve as a high-information-density representation that enhances the performance of diverse machine learning models, suggesting a pathway for developing efficient, interpretable surrogate models for microscale damage prediction.

The authors conclude that this framework provides a solid basis for future studies on nonlocal effects, GB damage initiation, and the development of reduced-order constitutive descriptions for creep in complex polycrystalline microstructures.