🔬 materials science

Uncertainty-guided active learning for surrogate prediction of stream-finishing wear fields

This paper presents an uncertainty-guided active learning framework using deep ensembles to efficiently predict stream-finishing wear fields directly from workpiece geometry, achieving high accuracy with only 13% of the required discrete element method simulations by selectively targeting the most uncertain orientations.

Original authors: Anand Kumar, Puli Saikiran, Vineet Dawara, Koushik Viswanathan

Published 2026-08-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Anand Kumar, Puli Saikiran, Vineet Dawara, Koushik Viswanathan

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 smooth out a bumpy, complex sculpture using a swirling river of tiny, hard sand grains. This process, called "stream finishing," is like a high-tech version of sanding a wooden chair, but instead of a hand-held block, the part is submerged in a spinning vat of abrasive particles. The goal is to make the surface perfectly smooth and shiny. However, there's a catch: the sculpture doesn't just sit there; it has to be held at a specific angle. If you hold it one way, the sand hits the front hard but misses the back. Hold it another way, and the back gets polished while the front stays rough. To get a perfect finish on a complex shape, you might need to rotate the part through hundreds of different angles, polishing a little bit at a time.

The problem is that figuring out exactly which angles work best is incredibly difficult. You can't just guess, because the physics of how millions of tiny sand grains crash into a complex shape is chaotic and hard to predict. Scientists use powerful computer simulations to model these collisions, but running a simulation for just one angle can take hours. If you have hundreds of angles to check, it would take years of computer time to find the perfect recipe. This is where the story gets interesting: instead of brute-forcing every single possibility, can we teach a computer to learn the pattern quickly and tell us when it's guessing?

This paper presents a clever solution to that exact problem. The researchers built a "smart guesser" (a machine learning model) that learns from a few expensive computer simulations and then predicts what would happen for hundreds of other angles without running the full simulation. But here is the magic trick: this smart guesser doesn't just give an answer; it also raises a hand and says, "I'm not sure about this one." When the model is unsure, the system automatically runs the expensive, slow simulation just for that specific angle to learn the truth. By only running the slow simulations when absolutely necessary, they managed to predict the wear patterns for nearly 700 different angles using only about 13% of the computer time it would have taken to check them all.

The Problem: The "Sand River" Puzzle

Think of the stream-finishing process as a chaotic dance floor. You have a workpiece (like a turbine blade or a medical implant) standing in the middle, and a swirling crowd of abrasive particles (like silicon carbide sand) is spinning around it. Every time a particle bumps into the workpiece, it chips away a tiny bit of material. The amount of wear depends on how hard the particle hits, the angle of the hit, and how many particles hit that specific spot.

To get a uniform finish, engineers need to know exactly how the wear will look for every possible way they can hold the part. If they hold it straight up, the top gets hit hard. If they tilt it, the side gets hit. To find the perfect sequence of tilts, they would ideally need to map out the wear for every single angle. In the real world, there are 696 different "feasible" angles they might try.

Traditionally, scientists use a method called the Discrete Element Method (DEM) to simulate this. Imagine a video game where every single grain of sand is a character with its own physics. The computer tracks every collision, every bounce, and every scratch. It's incredibly accurate, but it's also slow. Running this simulation for one angle takes hours. Doing it for 696 angles? That's practically impossible for a new design.

The Solution: The "Uncertainty-Aware" Detective

The authors realized they didn't need to simulate every angle. Instead, they built a "surrogate" model—a fast, lightweight AI that learns the rules of the game from a few examples and then predicts the rest. But a normal AI has a flaw: it often acts confident even when it's wrong. If you ask it about a situation it hasn't seen before, it might just make up a plausible-sounding answer.

To fix this, the team used a technique called a "deep ensemble." Imagine asking 12 different experts to solve a puzzle. If they all agree, you can trust the answer. If they all give different answers, you know the puzzle is tricky and you need more information. In this paper, the "experts" are 12 separate neural networks trained on slightly different data. When they are asked to predict the wear for a new angle, the computer looks at how much they disagree. If they disagree a lot, the system knows, "Hey, this is an uncertain area; we need to run the slow, expensive simulation here to get the real answer."

How They Did It: Learning the Rules, Not Just the Answers

The researchers didn't just ask the AI to guess the final wear pattern. That would be like asking a student to guess the final score of a football game without knowing the rules. Instead, they taught the AI to predict the three underlying ingredients that cause the wear:

Normal Impact Velocity: How fast the particles hit the surface head-on.
Tangential Impact Velocity: How fast they slide across the surface.
Particle Flux: How many particles hit a specific spot per second.

Once the AI predicts these three smooth, easy-to-learn numbers, the researchers plug them into a known physics formula (the Finnie model) to calculate the final wear. This is much smarter because the raw physics of the collisions is messy and jagged, but the speed and number of particles change smoothly as you rotate the part. It's easier for a computer to learn a smooth curve than a jagged mountain range.

The Results: A Smart, Efficient Loop

The team started by running the slow, expensive simulations for 52 carefully chosen angles (the "seeds"). They used these results to train their ensemble of 12 AI experts. Then, the real magic began:

The AI looked at the remaining 644 angles it hadn't seen yet.
It identified the 10 angles where the 12 experts disagreed the most (the most uncertain ones).
The computer ran the slow, expensive simulation for just those 10 angles.
The AI learned from these new results and updated its knowledge.
They repeated this loop five times.

By the end, they had only run the expensive simulation for 92 angles out of the 696 possible ones (about 13% of the total). Yet, their predictions were incredibly accurate.

For the angles the model was confident about (low uncertainty), the predicted wear pattern matched the real simulation almost perfectly, with a correlation score of 0.97. This means the AI could tell you exactly which parts of the sculpture would get worn down and which would stay safe. Even for the angles where the model was less sure, it was honest about its uncertainty. The "uncertainty score" the AI gave itself was a truthful predictor of how wrong it might be. If the score was high, the prediction was shaky; if it was low, the prediction was solid.

Why This Matters

This approach changes the game for manufacturing complex parts. Instead of spending months running simulations for every possible angle, engineers can now use this "uncertainty-guided" method to find the best polishing strategy in a fraction of the time. The system knows when it's guessing and when it knows, directing the computer's power only where it's needed.

The paper shows that with just 92 simulations, they could reconstruct the wear fields for the entire set of 696 orientations with high reliability. The "low uncertainty" group grew with every step, meaning that with just a few more rounds of this smart learning, they could likely cover the entire set of angles with high confidence. This isn't just a faster way to do the same thing; it's a way to solve problems that were previously too expensive to tackle, ensuring that everything from airplane blades to hip implants gets the perfect, smooth finish they need.

Technical Summary: Uncertainty-Guided Active Learning for Surrogate Prediction of Stream-Finishing Wear Fields

Problem Statement

Stream finishing is an abrasive process where a workpiece is held stationary within a rotating container of granular media. The quality of the surface finish depends critically on the workpiece's orientation relative to the media flow, as different orientations expose distinct surface regions to varying impact velocities and particle fluxes. To achieve uniform or targeted finishing, engineers must identify optimal sequences of orientations.

Accurately predicting the spatial wear-rate distribution for a given orientation requires high-fidelity simulations using the Discrete Element Method (DEM) coupled with microscopic erosion models (e.g., the Finnie model). While DEM provides accurate predictions of particle–surface interactions, it is computationally prohibitive for design optimization. Simulating hundreds of feasible orientations (e.g., 696 configurations generated by discretizing rotation angles) for a single new geometry can take thousands of hours. Exhaustive simulation is therefore impractical, necessitating a data-driven surrogate model that can predict wear behavior from a limited number of DEM runs. However, standard neural networks require large datasets and lack inherent mechanisms to indicate when their predictions are unreliable, making it difficult to know where to allocate further expensive DEM simulations.

Methodology

The authors propose an uncertainty-guided active learning framework that employs a deep ensemble surrogate to predict the primary kinematic fields governing erosion, rather than the wear rate directly.

1. Physics-Informed Input Features

Instead of predicting the non-linear wear rate, the surrogate predicts three smooth primary fields derived from the Finnie wear model:

Mean normal impact velocity ( $v_n$ )
Mean tangential impact velocity ( $v_t$ )
Particle impact flux ( $\phi$ )

The wear rate is subsequently reconstructed analytically using the Finnie relation. This approach separates the smooth physical dependencies (velocity and flux) from the sharp non-linearities of the erosion law, making the learning task easier and more physically transparent.

Each surface triangle of the workpiece is described by an 11-dimensional feature vector computable solely from the geometry and orientation, including:

Rotated centroid position and unit normal.
Baseline (unrotated) coordinates to maintain facet identity across orientations.
Radial distance from the container axis.
Alignment between the facet normal and the local flow direction (windward vs. leeward).

2. Deep Ensemble and Epistemic Uncertainty

To address the need for reliability estimates, the authors utilize a deep ensemble of $M=12$ independently initialized feedforward neural networks. Each network is trained on a different random 80% subset of the available training data.

Prediction: The ensemble mean ( $\mu^*$ ) serves as the final prediction.
Uncertainty: The variance between the ensemble members ( $\sigma_e$ ) serves as an estimate of epistemic uncertainty. High disagreement among members indicates regions of the input space where the model is under-constrained by data.

3. Active Learning Loop

The framework operates via an iterative loop designed to maximize information gain:

Seeding: An initial set of 52 "seed" orientations is selected using farthest-point sampling based on a geometric descriptor of the orientation space. This ensures the initial training set covers the feasible domain without requiring extrapolation.
Training & Prediction: The deep ensemble is trained on the current pool of DEM data and predicts the primary fields and uncertainty ( $\sigma_e$ ) for all remaining feasible orientations.
Selection: The 10 orientations with the highest epistemic uncertainty are selected for new DEM simulations.
Update: These new simulations are added to the training pool, and the loop repeats.

Key Results

The framework was tested on a stainless steel workpiece with 696 feasible orientations. The active learning loop was executed for five iterations, training the surrogate on a total of 92 simulations (approximately 13% of the total design space).

Prediction Accuracy: The surrogate achieved high Spearman rank correlations ( $\rho_s$ $ρ_{s}$ ) on held-out validation sets:
- Normal impact velocity: $\rho_s = 0.93$
- Tangential impact velocity: $\rho_s = 0.89$
- Particle impact flux: $\rho_s = 0.93$
Truthfulness of Uncertainty: A critical finding is that the predicted epistemic uncertainty is a truthful predictor of realized error. There is a strong negative correlation between the predicted uncertainty ( $\sigma_e$ $σ_{e}$ ) and the achieved accuracy ( $\rho_s$ $ρ_{s}$ ) on held-out orientations ( $\rho_s \approx -0.8$ $ρ_{s} \approx - 0.8$ ).
- Low Uncertainty: Orientations with low predicted uncertainty showed near-exact agreement with DEM wear maps ( $\rho_s = 0.97$ ).
- High Uncertainty: As uncertainty increased, the reconstruction of the wear field degraded in a controlled manner. While the exact wear magnitude ( $R^2$ ) dropped significantly, the model retained partial ability to rank wear regions ( $\rho_s$ remained non-zero), preserving the identification of high-wear zones.
Efficiency: By iteration 5, the "low-uncertainty" band (defined as $\rho_s \geq 0.90$ ) covered approximately 64% of the normal velocity field and 55% of the tangential velocity field across the entire feasible set, with the trend indicating that full coverage is achievable with only a modest number of additional iterations.

Significance and Claims

The paper claims that this framework successfully circumvents the computational cost of exhaustive DEM simulation for stream finishing by:

Decoupling Physics from Learning: By predicting smooth kinematic fields and applying the non-linear wear law analytically, the model achieves higher accuracy and physical interpretability than direct wear regression.
Calibrated Uncertainty: The deep ensemble provides reliable epistemic uncertainty estimates that accurately anticipate prediction errors before DEM simulations are run. This allows for an efficient active learning strategy where computational resources are directed only to the most uncertain orientations.
Generalizability: The input features are derived purely from triangulated mesh geometry, making the approach applicable to arbitrary complex shapes without re-engineering the feature set.

The authors conclude that the framework enables the construction of part orientation strategies based on wear-rate distributions at a fraction of the cost of traditional methods, with the uncertainty quantification providing a reliable metric for the fidelity of the reconstructed wear fields.