🤖 AI

Explainable Computer Vision Framework for Automated Pore Detection and Criticality Assessment in Additive Manufacturing

This study introduces an explainable computer vision framework for additive manufacturing that detects pores in 3D tomographic volumes and uses machine learning with SHAP analysis to reveal that a pore's proximity to the specimen boundary, rather than its size or shape, is the dominant factor determining its criticality.

Original authors: Akshansh Mishra, Rakesh Morisetty

Published 2026-02-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Akshansh Mishra, Rakesh Morisetty

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 baking a giant, complex cake using a 3D printer. Instead of flour and eggs, you are melting metal layer by layer. Sometimes, tiny air bubbles get trapped inside the cake as it cools. In the world of engineering, these bubbles are called pores, and they can make the metal weak or cause it to break later.

For a long time, engineers have used special "X-ray cameras" (called computed tomography) to take 3D pictures of these metal cakes to find the bubbles. But here was the problem: the computers that found the bubbles were like black boxes. They would point at a bubble and say, "This one is dangerous!" but they couldn't explain why. It was like a teacher giving you an "F" on a test without telling you which questions you got wrong.

This paper introduces a new, transparent framework that acts like a helpful tutor. It doesn't just find the bubbles; it explains exactly why some are dangerous and others are harmless.

How the System Works

The Detective Work: The system takes thousands of 2D slices of the metal cake and stacks them to build a 3D model. It then uses a "flashlight" (a mathematical threshold) to highlight the bright spots where the bubbles are hiding. It filters out the giant outer shell of the cake and focuses only on the tiny internal bubbles.
The Map: Once it finds 500 bubbles, it draws a map connecting them. Imagine a spiderweb where every bubble is a node, and the strings connecting them represent how close they are to each other. This helps the computer see the "neighborhood" of each bubble.
The Teacher (AI): The system uses a smart computer program (Machine Learning) to guess how "critical" or dangerous each bubble is. But unlike the old black boxes, this one uses a special tool called SHAP (think of it as a "reasoning engine") to break down its decision. It asks, "How much did the bubble's size matter? How much did its shape matter? How much did its location matter?"

The Big Surprise: Location is Everything

The most exciting discovery in this paper is what the "reasoning engine" found.

The Old Belief: Engineers used to think the size of the bubble was the most important thing. They assumed a giant bubble was always worse than a tiny one.
The New Reality: The AI revealed that size barely matters at all. Whether a bubble is tiny or medium-sized, it doesn't change the danger level much.
The Real Hero (or Villain): The distance to the surface is the only thing that truly matters.

The Analogy:
Think of the metal part like a fortress wall.

A bubble deep inside the wall is like a small crack in the middle of a thick brick. It's annoying, but the wall is still strong.
A bubble right next to the surface is like a crack right at the edge of the wall. Even if it's tiny, it's dangerous because it's right where the wall is weakest and where stress hits first.

The study found that bubbles near the surface are orders of magnitude more dangerous than bubbles in the middle. The computer's "reasoning" showed that the distance to the edge was the only factor that really drove the prediction. The shape of the bubble (was it round? was it long?) and its size were essentially noise.

What This Means for the Future

Because this new system is explainable, engineers can finally trust the computer. They don't have to guess why a part was rejected. They can see clearly: "This part failed because the bubbles are too close to the surface, not because they are too big."

This insight tells manufacturers that they shouldn't just worry about making fewer bubbles in the middle of the metal. Instead, they need to focus their efforts on smoothing out the surface layers and ensuring the outer edges are perfect. It shifts the strategy from "make everything perfect" to "protect the edges, because that's where the trouble starts."

In short, this paper gave us a pair of glasses that lets us see why a 3D printed part might fail, revealing that where a defect is located is far more important than how big it is.

Technical Summary: Explainable Computer Vision Framework for Automated Pore Detection and Criticality Assessment in Additive Manufacturing

Problem Statement
Internal porosity remains a primary defect mode in additively manufactured components, significantly compromising structural integrity and limiting industrial adoption. While computed tomography (CT) enables the visualization of internal defects, current quality control methods rely heavily on manual or semi-automated analysis that lacks physical insight into which specific pores pose the greatest risk. Furthermore, traditional machine learning models used for failure prediction often function as "black boxes," offering predictions without mechanistic justification. This opacity prevents engineers from understanding the physical basis of criticality assessments, thereby hindering confidence in automated systems and obstructing effective process optimization.

Methodology
The study proposes an explainable computer vision framework designed to detect pores in three-dimensional tomographic volumes and assess their criticality with mechanistic transparency. The methodology proceeds through the following stages:

Data Preprocessing and Segmentation: Input data consists of sequential grayscale TIFF slices reconstructed into a 3D volumetric dataset. High-intensity voxels are isolated using intensity-based thresholding ( $I_{thr} = 250$ ) to suppress background noise. Connected component labeling identifies individual bright regions. The largest component, corresponding to the specimen boundary or imaging artifacts, is excluded. Remaining components are classified as pores if their volume ( $A_k$ ) satisfies the condition $2 < A_k < 0.01 A_{max}$ , where $A_{max}$ is the size of the largest connected component.
Feature Extraction: Each of the 500 identified pores is characterized by geometric and spatial descriptors, including:
- Size: Total voxel count.
- Shape: Aspect ratio and extent.
- Spatial Position: Centroid coordinates and the minimum normalized distance from the pore centroid to the specimen surface.
Network Construction: A pore interaction network is constructed where pores act as nodes. Connections (edges) are established based on a percentile-based Euclidean distance criterion (specifically the lowest 20% of pairwise distances) to ensure consistent network density without arbitrary global thresholds. This resulted in a network of 500 nodes and 24,950 edges.
Machine Learning and Explainability: A supervised machine learning model predicts a scalar criticality score for each pore using the extracted descriptors. To interpret the model, SHapley Additive exPlanations (SHAP) are employed. Feature importance is quantified using mean absolute SHAP values, and the directionality of contributions is analyzed via beeswarm and dependency plots.

Key Results
The analysis of the segmented tomographic volume and the subsequent machine learning model yielded the following findings:

Dominance of Surface Proximity: SHAP analysis reveals that normalized surface distance is the overwhelmingly dominant feature, contributing more than an order of magnitude greater importance than all other descriptors combined. The model output is effectively approximated by the surface distance contribution, with other features acting as negligible modifiers.
Inverse Relationship with Criticality: There is a strong, nearly linear inverse relationship between surface distance and predicted criticality. Pores located closer to the specimen surface exhibit significantly higher criticality scores, while those deeper within the material show progressively lower scores. This indicates a boundary-driven failure mechanism.
Minimal Influence of Size and Geometry: Contrary to conventional assumptions, pore size provides minimal influence on criticality predictions. While larger pores show a weak positive contribution, they do not act as primary drivers. Geometric descriptors, including aspect ratio, extent, and axial position, demonstrate negligible impact, with SHAP values tightly clustered around zero.
Spatial Distribution: The reconstructed 3D network shows that pores are distributed along the full specimen height but are strongly localized near the outer radial boundary, forming an annular arrangement with a sparsely populated core.

Significance and Claims
The paper claims that this interpretable framework enables transparent defect assessment by shifting the focus from opaque algorithmic predictions to physical understanding. The primary significance lies in the revelation that spatial location (specifically proximity to the boundary) governs pore criticality more than intrinsic pore morphology or size.

This mechanistic insight suggests that quality assurance strategies and process controls in additive manufacturing should prioritize the mitigation of near-surface defects over the reduction of bulk porosity. By providing actionable insights into the physical basis of criticality, the framework supports data-driven decisions for process optimization. The authors position this work as a step toward validating surface-driven criticality mechanisms, with future work proposed to extend the methodology across different alloy systems and manufacturing processes to validate generalizability.

How the System Works

The Big Surprise: Location is Everything

What This Means for the Future

More like this