Latent Space Dynamics Identification for Interface Tracking with Application to Shock-Induced Pore Collapse

The paper introduces LaSDI-IT, a data-driven framework that combines a revised auto-encoder with explicit interface-aware encoding and latent dynamics learning to efficiently and accurately model shock-induced pore collapse in high explosives, achieving high-fidelity accuracy with half the training data and a 106-fold speedup over conventional simulations.

The Gist

Imagine you are trying to predict how a drop of water will behave when it hits a hot pan. It splashes, boils, and changes shape instantly. Now, imagine doing this for a high-explosive material where a tiny bubble (a "pore") inside the material gets crushed by a massive shockwave. This process creates intense heat spots that could cause an explosion.

Simulating this on a supercomputer is like trying to film every single water molecule in slow motion. It's incredibly accurate, but it takes so much time and power that you can't run it a million times to test different scenarios. You need a shortcut—a "cheat code" that gives you the answer almost instantly without losing the important details.

This paper introduces that cheat code, called LaSDI-IT. Here is how it works, broken down into simple concepts:

1. The Problem: The "Blurry Photo" Effect

Standard AI models (neural networks) are great at learning smooth patterns, like the curve of a hill or the flow of a river. But they struggle with sharp edges.

Think of a neural network like a camera with a slightly out-of-focus lens. If you take a picture of a sharp, jagged rock, the camera tries to smooth out the edges to make the picture look "nice." In physics, this is bad. If you are modeling a shockwave hitting a pore, that "blurry" edge is exactly where the explosion happens. If the AI smooths it out, it misses the danger zone.

2. The Solution: The "Two-Track" Camera

The authors built a new type of AI camera called LaSDI-IT. Instead of just trying to guess the temperature of the material, this camera does two things at once:

1. It takes a photo of the temperature (the heat).
2. It takes a photo of the "mask" (a black-and-white map showing exactly where the material is and where the empty hole is).

The Analogy: Imagine trying to describe a picture of a red apple on a white table.

• Old AI: Tries to guess the color of every pixel. It might accidentally paint a little bit of red on the white table because it's confused about the edge.
• LaSDI-IT: First, it draws a perfect outline of the apple (the mask). Then, it fills in the red color only inside that outline. This way, the edge stays razor-sharp, and the white table stays perfectly white.

By teaching the AI to learn the "shape" of the hole separately from the "heat," it stops blurring the edges.

3. The "Smart Guessing" Game (Greedy Sampling)

Usually, to train an AI, you need thousands of examples. But running these physics simulations is expensive (like buying a lottery ticket every time you want to learn something). You can't afford 1,000 tickets.

The authors used a clever strategy called Greedy Sampling.

• The Analogy: Imagine you are a detective trying to solve a mystery in a large city, but you only have time to visit 4 locations.
• The Old Way: You pick 4 random spots. You might miss the crime scene entirely.
• The LaSDI-IT Way: You start with 4 spots. The AI looks at the map and says, "I'm very confused about what happens in this specific neighborhood." So, you go there next. Then it says, "Now I'm confused about this other street." You go there next.

The AI acts like a GPS that tells you exactly where to look next to learn the most. This allowed them to get 99% accuracy using only half the data they would normally need.

4. The Result: Speed vs. Accuracy

The final result is a model that is:

• Fast: It predicts the outcome 1 million times faster than the supercomputer simulation. It's the difference between waiting for a movie to download and watching it instantly on your phone.
• Accurate: It predicts the size of the "hot spot" (the danger zone) and the temperature with less than 9% error.
• Reliable: It correctly identifies how big the hole is and when the explosion might start.

Why Does This Matter?

This isn't just about explosives. This method is a universal tool for any problem where things change shape quickly and sharply.

• Medicine: Tracking how a tumor shrinks or grows.
• Engineering: Watching how a crack spreads through a bridge.
• Weather: Predicting the sharp edge of a hurricane.

In short, the authors built a "smart, fast, and sharp-eyed" AI that can watch a chaotic physical event, understand exactly where the edges are, and predict the future in a fraction of a second, saving scientists years of computing time.

Technical Summary

1. Problem Statement

The paper addresses a fundamental challenge in Reduced-Order Modeling (ROM): accurately capturing sharp, moving interfaces (discontinuities) in physical systems with limited data.

• Context: Many engineering problems, such as shock propagation, phase transitions, and fracture mechanics, involve steep gradients or discontinuities that standard ROM techniques struggle to resolve.
• Specific Application: The authors focus on shock-induced pore collapse in high explosives (HE). In this process, a shock wave compresses a porous HE material, causing pores to collapse and generating localized temperature spikes ("hot spots") that can trigger detonation.
• Challenges:
  • Spectral Bias: Standard neural networks (and autoencoders) preferentially learn smooth, low-frequency components, causing them to "smooth out" sharp interfaces (e.g., the boundary between the pore and the solid material).
  • Data Scarcity: High-fidelity simulations (e.g., using ALE3D) are computationally expensive, making it difficult to generate the large datasets required for training robust data-driven models.
  • Metric Sensitivity: Accurate prediction of the pore boundary is critical for calculating key metrics like pore area and hot spot size, which standard ROMs often fail to do due to reconstruction ambiguity.

2. Methodology: LaSDI-IT

The authors propose LaSDI-IT (Latent Space Dynamics Identification for Interface Tracking), a data-driven framework that extends the existing LaSDI/GPLaSDI approach. The methodology consists of three core components:

A. Interface-Aware Autoencoder (The Core Innovation)

To overcome spectral bias and interface ambiguity, the authors modified the standard autoencoder architecture:

• Dual-Output Decoder: Instead of just reconstructing the physical field (temperature $T$), the decoder jointly reconstructs:
  1. The scaled physical field ($\tilde{T}'$).
  2. A binary indicator function ($\phi_{pred}$) representing the material domain (1 for active material, 0 for the pore/void).
• Reconstruction Logic: The final predicted state is computed as $\tilde{T}_{pred} = \tilde{T}' \times \mathbb{1}_{\phi_{pred} > 0.5}$. This ensures that regions identified as "pore" are strictly zero, eliminating spurious temperature values in the void.
• Modified Loss Function: The training loss is split into two parts:
  • Value Loss ($J_{val}$): Minimizes error in the physical field within the active domain.
  • Mask Loss ($J_{mask}$): Minimizes error in the indicator function (treating it as a classification task).
• Adaptive Scaling: Standard normalization (mean/std) is recalculated excluding the zero values in the pore region. This prevents the pore's zero values from skewing the statistics of the active material, allowing the network to focus on the variance of the physical field.

B. Latent Dynamics Identification

• The high-dimensional state is encoded into a low-dimensional latent space $z$.
• The evolution of $z$ is modeled as a system of Ordinary Differential Equations (ODEs): $\dot{z} = \Xi(p) \Phi(z)$.
• For this specific application, the authors restricted the dynamics to a linear form (reducing overfitting risk and computational cost) while relying on the autoencoder to capture the non-linear manifold of the solution.
• Coefficients $\Xi$ are identified via regression (SINDy-style) on the encoded training data.

C. Parametric Interpolation & Greedy Sampling

• Gaussian Process (GP): To generalize to unseen parameters (pore orientation and size), GP regression is used to interpolate the ODE coefficients $\Xi(p)$ across the parameter space.
• Uncertainty-Driven Greedy Sampling: Instead of uniform sampling, the framework uses the predictive uncertainty of the GP to select new training points.
  • The algorithm samples the latent dynamics with Monte Carlo draws of the coefficients.
  • The parameter point with the highest solution uncertainty is selected as the next training point.
  • This allows the model to achieve high accuracy with significantly fewer training simulations.

3. Key Contributions

1. LaSDI-IT Framework: A novel ROM framework specifically designed for systems with moving sharp interfaces, combining latent dynamics learning with explicit interface tracking.
2. Interface-Aware Architecture: The introduction of a dual-output autoencoder that separates the reconstruction of physical values from the geometric boundary, effectively solving the "smoothing" problem of standard neural networks.
3. Data Efficiency: Demonstrated that Gaussian Process-based greedy sampling can reduce the required training dataset size by 50% while maintaining accuracy comparable to models trained on uniformly sampled data.
4. Physical Fidelity: Successfully recovered critical physical metrics (pore area, hot spot size, max temperature) that are usually lost in standard ROMs.

4. Results

The method was tested on shock-induced pore collapse simulations (ALE3D code) with parameters: pore orientation ($0^\circ$–$20^\circ$) and major axis length ($1.0$–$1.2$ $\mu$m).

• Reconstruction Accuracy:
  • Standard AE: Failed to resolve the pore boundary, resulting in non-zero temperatures inside the pore and high point-wise errors near the interface.
  • LaSDI-IT: Achieved < 9% relative prediction error across the parameter space. The pore boundary was sharply defined with zero spurious values.
• Metric Recovery:
  • Accurately predicted hot spot area (defined by $T > 800$ K) and pore area.
  • Correctly captured the timing and evolution of the maximum temperature, though peak values were slightly underestimated (a common issue with latent space smoothing).
• Computational Speed:
  • The surrogate model prediction is $10^6$ times faster than the high-fidelity simulation (0.033s vs. ~10 mins per parameter point).
• Training Efficiency:
  • Starting with only 4 corner points, the greedy sampling added 11 points to reach a total of 15.
  • This 15-point model achieved accuracy comparable to a 36-point uniformly sampled model, proving the efficiency of the uncertainty-guided sampling strategy.

5. Significance and Future Outlook

• General Applicability: While demonstrated on high explosives, the framework is applicable to any multiphase flow, fracture mechanics, or phase change problem where interface fidelity is crucial.
• Bridging the Gap: It bridges the gap between data-driven efficiency and physical fidelity, enabling real-time prediction and design optimization for complex discontinuous systems.
• Future Work: The authors plan to extend the method to multi-pore interactions, more complex geometries, and coupling with reactive chemistry models.

In summary, LaSDI-IT represents a significant advancement in reduced-order modeling by explicitly addressing the "spectral bias" limitation of neural networks in the context of moving interfaces, offering a highly efficient and accurate tool for simulating shock-induced phenomena in energetic materials.