Geometric and Topological Deep Learning for Predicting Thermo-mechanical Performance in Cold Spray Deposition Process Modeling

This study develops and evaluates a geometric deep learning framework using graph neural networks to accurately predict thermo-mechanical responses in cold spray deposition, demonstrating that spatial graph-based aggregation significantly outperforms traditional and topological methods in modeling the process's highly non-linear behavior.

The Gist

The Big Picture: The "Cold Spray" Problem

Imagine you are trying to build a wall by throwing tiny metal marbles at a surface at supersonic speeds (faster than a bullet). This is called Cold Spray. Unlike welding, which melts metal, this process sticks the marbles together purely through the sheer force of the impact, like squishing a piece of clay so hard it fuses to the table.

The problem? It's incredibly hard to predict exactly what happens when a marble hits.

• Does it stick?
• Does it bounce off?
• Does it get hot enough to melt?
• Does it squish into a pancake or stay round?

To figure this out, scientists usually run complex computer simulations (like a video game physics engine). But these simulations are slow and expensive. Running them for every possible speed, temperature, and surface condition would take years.

The Solution: The "Smart Guessing" Machine

This paper introduces a new way to use Artificial Intelligence (AI) to act as a "shortcut." Instead of running the slow simulation every time, the AI learns from a library of past simulations and predicts the outcome instantly.

But here's the twist: The researchers didn't just use a standard AI. They used Geometric Deep Learning.

The Analogy: The "Party Guest" vs. The "Isolated Stranger"

Imagine you are trying to guess how much a person at a party will enjoy a specific song.

• Standard AI (The Isolated Stranger): It looks at one person alone. It sees they like jazz, so it guesses they will like the new jazz song. It ignores everyone else.
• Geometric Deep Learning (The Party Guest): It looks at the person and their friends. It sees that this person is standing right next to three other people who are dancing and smiling. It realizes, "Ah, this person is in a 'happy zone' with their friends, so they will probably love this song too."

In this paper, the "people" are the simulation results. The "friends" are other simulations with similar settings (e.g., similar speed and temperature). The AI learns that if a simulation at 500 m/s worked well, a simulation at 510 m/s (its neighbor) will likely work similarly.

The Experiment: The "Taste Test"

The researchers created a dataset of 100 different scenarios by changing three main ingredients:

1. Speed: How fast the particle flies (400 to 900 m/s).
2. Temperature: How hot the particle is before impact (300 to 600 Kelvin).
3. Friction: How "slippery" the surface is.

They then tested four different AI "chefs" to see who could predict the results (like how much the particle squished or how hot it got) best:

1. GraphSAGE: The "Social Networker." It looks at a sample and averages the results of its closest neighbors.
2. GAT (Geometric Attention Network): The "Smart Manager." It looks at neighbors but decides which ones matter most. (e.g., "This neighbor is very similar, so I'll listen to them closely. That other one is a bit different, so I'll ignore them.")
3. ChebSpectral: The "Mathematical Theorist." It tries to analyze the whole pattern using complex frequency math.
4. TDA-MLP: The "Shape Detective." It tries to find the overall "shape" or structure of the data.

The Results: Who Won?

The results were clear, like a race where two runners left the others in the dust.

• The Winners (GraphSAGE & GAT): These two were fantastic. They predicted the outcomes with 93% to 97% accuracy.

  • Why? Because they understood that in physics, things that are "close" in settings usually behave similarly. By looking at their neighbors, they learned the rules of the game perfectly.
  • The Champion: GAT was the best, especially at predicting how much the metal squished (plastic strain). It was like having a manager who knew exactly which clues to focus on.

• The Losers (ChebSpectral & TDA-MLP): These models struggled, with some predictions being worse than random guessing (negative accuracy scores).

  • Why? They tried to find complex global patterns or shapes that didn't exist in this specific data. They were overthinking the problem and missing the simple, local connections between similar scenarios.

The Key Discovery: Speed is King

The study also revealed something interesting about the physics of Cold Spray:

• Speed is the Boss: The speed of the particle is the single most important factor. If you change the speed, everything changes.
• Temperature and Friction are Sidekicks: They matter, but only when you already know the speed. If you ignore speed, temperature and friction look like random noise.

The Takeaway

This paper proves that for complex engineering problems like Cold Spray, AI works best when it respects the "neighborhood."

Instead of treating every data point as an isolated fact, the best AI models look at the context: "What happened to the similar cases nearby?" This approach allows engineers to skip the slow, expensive computer simulations and get instant, highly accurate predictions to design better coatings for airplanes, medical implants, and energy systems.

In short: The researchers built a super-smart "neighborhood watch" AI that learned the rules of metal impact by chatting with its neighbors, beating out the other AI models that tried to solve the puzzle all by themselves.

Technical Summary

1. Problem Statement

Cold spray is a solid-state additive manufacturing process where metallic particles are accelerated to supersonic velocities to bond with a substrate via extreme plastic deformation rather than melting. The mechanical integrity of the deposited layer depends on complex, highly coupled thermo-mechanical interactions governed by particle impact velocity, particle temperature, and interfacial friction.

While Finite Element (FE) modeling (e.g., using Abaqus) is the standard tool for investigating these single-particle impact mechanics, it is computationally expensive. Exhaustive parametric exploration of the input space (velocity, temperature, friction) via simulation is practically infeasible. Traditional machine learning surrogate models (e.g., standard MLPs, random forests) treat simulation samples as independent observations, failing to exploit the geometric relationships and spatial continuity that exist between samples in similar regions of the parameter space. This limitation is critical when data is sparse and the underlying physics are highly non-linear.

2. Methodology

A. Data Generation (Finite Element Simulation)

• Geometry: A 3D model of a spherical aluminum particle ($r=40\,\mu m$) impacting a cylindrical aluminum substrate ($r=250\,\mu m$, depth $250\,\mu m$).
• Material Model: Thermo-mechanically coupled constitutive framework using the Johnson-Cook plasticity model (accounting for strain hardening, strain-rate sensitivity, and thermal softening) and a linear shock velocity–particle velocity equation of state (EOS).
• Meshing: Adaptive mesh refinement (ALE) with fine meshing ($2.5\,\mu m$) at the impact zone to capture extreme deformation gradients.
• Parametric Study: An automated Python framework in Abaqus generated 100 simulation cases varying:
  • Particle Velocity ($V$): 400–900 m/s
  • Particle Temperature ($T_p$): 300–600 K
  • Friction Coefficient ($\mu$): 0.10–0.50
• Outputs: Five target variables: Max Equivalent Plastic Strain (PEEQ), Average Contact PEEQ, Max Temperature, Max Von Mises Stress, and Deformation Ratio.

B. Geometric Deep Learning Framework

Instead of treating data as a flat table, the authors constructed a k-nearest-neighbor (k-NN) graph in the normalized feature space. Each simulation case is a node, and edges connect similar parameter sets, weighted by a Gaussian kernel based on Euclidean distance.

Four distinct algorithms were implemented and compared:

1. GraphSAGE: An inductive graph neural network using weighted mean message passing to aggregate neighbor information.
2. Chebyshev Spectral GNN: Uses Chebyshev polynomial approximations of the graph Laplacian for spectral convolution.
3. TDA-augmented MLP: A standard Multi-Layer Perceptron enriched with Topological Data Analysis (TDA) descriptors (persistent homology-based connectivity/density proxies).
4. Geometric Attention Network (GAT): Uses an attention mechanism to weigh the importance of neighboring nodes dynamically, modulated by geometric edge weights.

3. Key Contributions

• Novel Application: First application of geometric and topological deep learning specifically for cold spray particle impact prediction.
• Inductive Bias: Demonstrated that incorporating spatial graph-based neighborhood aggregation serves as a physically motivated inductive bias, capturing the smooth, continuous nature of physical responses better than point-wise models.
• Comparative Analysis: Provided a rigorous benchmarking of spectral, topological, and attention-based graph learning strategies against a complex, non-linear thermo-mechanical dataset.
• Feature Space Visualization: Detailed 2D contour and 3D surface analyses revealing that particle velocity is the dominant driver of deformation and thermal response, while temperature and friction act as secondary, coupled modulators.

4. Results

A. Feature Space Topology

• Velocity Dominance: Contour plots confirmed that plastic strain and deformation ratio are overwhelmingly governed by velocity. When velocity is excluded from 2D projections (e.g., Temperature vs. Friction), the response surfaces become irregular, sparse, and physically unintuitive.
• Non-Linearity: The relationship between inputs and outputs is highly non-linear. For instance, Von Mises stress exhibits competing mechanisms (hardening vs. softening) leading to localized peaks rather than monotonic trends.

B. Model Performance

The models were evaluated using $R^2$, Mean Squared Error (MSE), and Mean Absolute Error (MAE).

• Top Performers: GraphSAGE and GAT consistently outperformed all other methods.
  • GAT achieved the highest accuracy for Max PEEQ ($R^2 = 0.97$) and Max Temp ($R^2 = 0.95$).
  • GraphSAGE excelled at Avg Contact PEEQ ($R^2 = 0.98$) and Deformation Ratio ($R^2 = 0.94$).
  • Both models achieved $R^2 > 0.93$ across most targets, indicating robust generalization.
• Poor Performers:
  • ChebSpectral and TDA-MLP performed significantly worse.
  • TDA-MLP yielded negative $R^2$ values for several targets (e.g., $R^2 = -0.21$ for Max Temp), indicating it failed to learn the underlying relationship and performed worse than a simple horizontal mean baseline.
  • ChebSpectral showed moderate performance ($R^2 \approx 0.3-0.5$) but failed to capture the fine-grained non-linearities.

C. Visualization

Predicted vs. Actual plots for GraphSAGE and GAT showed tight clustering along the diagonal identity line, whereas TDA-MLP and ChebSpectral showed scattered, random distributions.

5. Significance and Conclusion

This study establishes that spatial graph-based neighborhood aggregation is a superior strategy for surrogate modeling in cold spray processes compared to traditional MLPs or purely topological/spectral approaches.

• Physical Interpretability: The success of GraphSAGE and GAT confirms that the thermo-mechanical response of cold spray is locally continuous; knowing the response of "nearby" parameter sets significantly aids in predicting the response of a specific set.
• Efficiency: These models offer a computationally efficient alternative to running thousands of expensive FE simulations for process optimization.
• Future Work: The framework provides a foundation for extending these models to multi-particle deposition scenarios and predicting microstructure-level properties, moving from single-particle mechanics to full-scale coating performance.

In summary, the paper proves that treating simulation data as a geometric graph rather than independent points unlocks the ability to accurately model the complex, non-linear physics of cold spray deposition, with Graph Attention Networks (GAT) and GraphSAGE emerging as the most effective tools for this domain.